835618

Journal of Probability and Statistics
Actuarial and Financial Risks: Models,

Statistical Inference, and Case Studies
Guest Editors: Ricardas Zitikis, Edward Furman, Abdelhakim Necir,

Johanna Nelehov, and Madan L. Puri

Guest Editors: Ricardas Zitikis, Edward Furman, Abdelhakim Necir,
Johanna Neslehova, and Madan L. Puri
Copyright q 2010 Hindawi Publishing Corporation. All rights reserved.
This is an issue published in volume 2010 of Journal of Probability and Statistics. All articles are open access articles
distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and repro-
duction in any medium, provided the original work is properly cited.
Editorial Board
Mohammad F. Al-Saleh, Jordan Debasis Kundu, India Man Lai Tang, Hong Kong
Vo V. Anh, Australia Michael Lavine, USA Robert J. Tempelman, USA
Zhidong Bai, China Nikolaos E. Limnios, France A. Thavaneswaran, Canada
Ishwar Basawa, USA Chunsheng Ma, USA Peter van der Heijden, The Netherlands
Shein-chung Chow, USA Eric Moulines, France Rongling Wu, USA
Murray Clayton, USA Hung T. Nguyen, USA Kelvin K. W. Yau, Hong Kong
Dennis Dean Cox, USA Madan L. Puri, USA Philip L. H. Yu, Hong Kong
Junbin B. Gao, Australia Jose Mara Sarabia, Spain Daniel Zelterman, USA
Arjun K. Gupta, USA Lawrence A. Shepp, USA Ricardas Zitikis, Canada
Tomasz J. Kozubowski, USA H. P. Singh, India
Contents
Actuarial and Financial Risks: Models, Statistical Inference, and Case Studies, Ricardas Zitikis,
Edward Furman, Abdelhakim Necir, Johanna Neslehova, and Madan L. Puri
Volume 2010, Article ID 392498, 3 pages
Forest Fire Risk Assessment: An Illustrative Example from Ontario, Canada, W. John Braun,
Bruce L. Jones, Jonathan S. W. Lee, Douglas G. Woolford, and B. Mike Wotton
The Foundations of Probability with Black Swans, Graciela Chichilnisky

Asteroids: Assessing Catastrophic Risks, Graciela Chichilnisky and Peter Eisenberger

Continuous Time Portfolio Selection under Conditional Capital at Risk,

Gordana Dmitrasinovic-Vidovic, Ali Lari-Lavassani, Xun Li, and Antony Ware
Individual Property Risk Management, Michael S. Finke, Eric Belasco, and Sandra J. Huston
On Some Layer-Based Risk Measures with Applications to Exponential Dispersion Models,

Olga Furman and Edward Furman
Pricing Equity-Indexed Annuities under Stochastic Interest Rates Using Copulas,

Patrice Gaillardetz
Local Likelihood Density Estimation and Value-at-Risk, Christian Gourieroux and Joann Jasiak
Zengas New Index of Economic Inequality, Its Estimation, and an Analysis of Incomes in
Italy, Francesca Greselin, Leo Pasquazzi, and Ricardas Zitikis
Risk Navigator SRM: An Applied Risk Management Tool, Dana L. K. Hoag and Jay Parsons
Estimating L-Functionals for Heavy-Tailed Distributions and Application, Abdelhakim Necir

and Djamel Meraghni
POT-Based Estimation of the Renewal Function of Interoccurrence Times of Heavy-Tailed

Risks, Abdelhakim Necir, Abdelaziz Rassoul, and Djamel Meraghni
Estimating the Conditional Tail Expectation in the Case of Heavy-Tailed Losses,
Abdelhakim Necir, Abdelaziz Rassoul, and Ricardas Zitikis
An Analysis of the Influence of Fundamental Values Estimation Accuracy on Financial

Markets, Hiroshi Takahashi
Hindawi Publishing Corporation
doi:10.1155/2010/392498
Editorial
Actuarial and Financial Risks: Models, Statistical
Inference, and Case Studies
Ricardas Zitikis,1 Edward Furman,2 Abdelhakim Necir,3

Johanna Neslehova,4 and Madan L. Puri5, 6
1
Department of Statistical and Actuarial Sciences, University of Western Ontario, London,
ON, Canada N6A 3K7
2
Department of Mathematics and Statistics, York University, Toronto, ON, Canada M3J 1P3
3
Laboratory of Applied Mathematics, Mohamed Khider University, Biskra 07000, Algeria
4
Department of Mathematics and Statistics, McGill University, Motreal, QC, Canada H3A 2K6
5
Department of Mathematics and Statistics, King Fahd University of Petroleum and Minerals,
Dhahran 31261, Saudi Arabia
6
Department of Mathematics, Indiana University, Bloomington, Indiana 47405, USA
Correspondence should be addressed to Ricardas Zitikis, [email protected]
Received 10 June 2010; Accepted 10 June 2010
Copyright q 2010 Ricardas Zitikis et al. This is an open access article distributed under the
Creative Commons Attribution License, which permits unrestricted use, distribution, and
reproduction in any medium, provided the original work is properly cited.
Understanding actuarial and financial risks poses major challenges. The need for reliable
approaches to risk assessment is particularly acute in the present context of highly uncertain
financial markets. New regulatory guidelines such as the Basel II Accord for banking and
Solvency II for insurance are being implemented in many parts of the world. Regulators
in various countries are adopting risk-based approaches to the supervision of financial
institutions.
Many researchers in a variety of areas have been dealing with nontrivial and highly
multifaceted problems in an attempt to answer seemingly plain questions such as how to
assess and quantify risks Crouhy, Galai, and Mark 1. The present issue of the Journal
of Probability and Statistics provides a glimpse of how challenging such problems are,
both philosophically and mathematically, through a collection of papers that cover a large
spectrum of applied and theoretical problems.
A number of ideas concerning measuring risks stem from the economic theory and
in particular from the classical utility theory Neumann and Morgenstern 2 as well as
from the prospect theory Kahneman and Tversky 3, which were subsequently developed
into the anticipated, also known as rank-dependent or generalized expected, utility theory
Quiggin 4, and most recently into a ground-breaking theory of choice under uncertainty
that allows for the presence of catastrophic risks Chichilnisky 5.
2 Journal of Probability and Statistics
This special Journal of Probability and Statistics issue oers many articles based
on such economic theories and their extensions. G. Chichilnisky develops foundations
for dealing with catastrophic risks, called black swans, which require tools beyond the
classical -additive probability theory. M. Finke, E. Belasco, and S. Huston review household
property risk management theory in order to compare optimal risk retention to conventional
practice. Aided with ideas of behavioral economics and finance, H. Takahashi investigates
the forecast accuracy of fundamental values in financial markets and clarifies issues such as
price fluctuations. F. Greselin, L. Pasquazzi, and R. Zitikis develop statistical inference for
Zengas index of economic inequality, whose construction brings to mind the relative nature
of notions such as small and large, poor and rich.
To make us aware of the scope and complexity of the problem, several authors have
contributed papers tackling risks within and beyond the financial sector. G. Chichilnisky and
P. Eisenberger have written a far-reaching article on asteroid risks. They convince us about the
critical importance of these risks, for which very little research has been carried out, and they
provide an interesting methodology for comparing asteroid risks with the risks of climate
change to make better decisions about research and development. For further information on
this and other related topics, we refer to the web site http://www.chichilnisky.com/.
W. J. Braun, B. L. Jones, J. S. W. Lee, D. G. Woolford, and M. Wotton examine the
risk assessment of forest fires in order to generate burn probability risk maps, concentrating
on the district municipality of Muskoka in Ontario, Canada, as an illustrative example.
D. L. K. Hoag and J. Parsons discuss their new program, Risk Navigator SRM, which
greatly expands the number of managers that can address risks in agriculture. The
program lays down solid foundations and provides state-of-the-art practical support tools,
including a web site http://www.risknavigatorsrm.com/, a risk management simulator
http://www.agsurvivor.com/, and a book that puts it all together Hoag 6.
Specific insurance risk-related problems are tackled by a number of authors.
P. Gaillardetz develops an elaborate evaluation approach to equity-linked insurance products
under stochastic interest rates. O. Furman and E. Furman propose layer-based counterparts of
a number of risk measures and investigate their properties and analytic tractability, especially
within the framework of exponential dispersion models.
One of the basic measures of risk is the so-called value-at-risk, which amounts to a
high quantile from the statistical point of view. It is arguably one of the most challenging
measures to estimate and to work with in practice. This risk measure was advocated by
the Basel Committee on Banking Supervision http://www.bis.org/bcbs/ and implemented
worldwide. C. Gourieroux and J. Jasiak suggest and develop a novel nonparametric method
for estimating the conditional value-at-risk and illustrate its performance on real-life portfolio
returns. G. Dmitrasinovic-Vidovic, A. Lari-Lavassani, X. Li, and A. Ware explore the
conditional capital-at-risk measure in the context of portfolio optimization and oer optimal
strategies.
This special Journal of Probability and Statistics issue also includes papers that
develop statistical inference for distortion risk measures and related quantities in the case
of heavy-tailed distributions. A. Necir and D. Meraghni deal with the estimation of L-
functionals, which are generalizations of the distortion risk measure, and which naturally
arise in the aforementioned anticipated utility theory. A. Necir, A. Rassoul, and D. Meraghni
develop a theory for estimating the renewal function of interoccurrence times of heavy-tailed
risks. A. Necir, A. Rassoul, and R. Zitikis introduce a new estimator of the conditional tail
expectation, which is one of the most important examples of the distortion risk measure and
Journal of Probability and Statistics 3
demonstrate the performance of the new estimator within the framework of heavy-tailed
risks.
We, the editors of this special issue, most sincerely thank three groups of people,
without whom this special issue would not have reached its fruition: first and foremost,
the authors who have shared with us their research achievements; second, the helpful and
ecient Hindawi Publishing Corporation sta; third, the professional and diligent referees
whose eorts resulted in useful feedback incorporated into often several rounds of revisions
of the herein published papers. We are also grateful to various granting agencies the
Actuarial Foundation, the Actuarial Education and Research Fund, the Society of Actuaries
Committee on Knowledge Extension Research, and the Natural Sciences and Engineering
Research Council of Canada for their support of our research.
Ricardas Zitikis
Edward Furman
Abdelhakim Necir
Johanna Neslehova
Madan L. Puri
References
1 M. Crouhy, D. Galai, and R. Mark, The Essentials of Risk Management, McGraw-Hill, NY, USA, 2006.
2 J. von Neumann and O. Morgenstern, Theory of Games and Economic Behavior, Princeton University
Press, Princeton, NJ, USA, 1944.
3 D. Kahneman and A. Tversky, Prospect theory: an analysis of decision under risk, Econometrica, vol.
47, pp. 263291, 1979.
4 J. Quiggin, Generalized Expected Utility Theory, Kluwer Academic, Dodrecht, The Netherlands, 1993.
5 G. Chichilnisky, An axiomatic approach to choice under uncertainty with catastrophic risks, Resource
and Energy Economics, vol. 22, no. 3, pp. 221231, 2000.
6 D. L. Hoag, Applied Risk Management in Agriculture, CRC Press, Boca Raton, Fla, USA, 2009.
doi:10.1155/2010/823018
Research Article
Forest Fire Risk Assessment: An Illustrative
Example from Ontario, Canada
W. John Braun,1 Bruce L. Jones,1 Jonathan S. W. Lee,1

Douglas G. Woolford,2 and B. Mike Wotton3
1
Department of Statistical and Actuarial Sciences, The University of Western Ontario, London,
Ontario, ON, Canada N6A 5B7
2
Department of Mathematics, Wilfrid Laurier University, Waterloo, ON, Canada N2L 3C5
3
Faculty of Forestry, The University of Toronto, Toronto, ON, Canada M5S 3B3
Correspondence should be addressed to W. John Braun, [email protected]
Received 2 October 2009; Revised 6 March 2010; Accepted 27 April 2010
Academic Editor: Ricardas Zitikis
Copyright q 2010 W. John Braun et al. This is an open access article distributed under the Creative
Commons Attribution License, which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.
This paper presents an analysis of ignition and burn risk due to wildfire in a region of Ontario,
Canada using a methodology which is applicable to the entire boreal forest region. A generalized
additive model was employed to obtain ignition risk probabilities and a burn probability map
using only historic ignition and fire area data. Constructing fire shapes according to an accurate
physical model for fire spread, using a fuel map and realistic weather scenarios is possible with
the Prometheus fire growth simulation model. Thus, we applied the Burn-P3 implementation of
Prometheus to construct a more accurate burn probability map. The fuel map for the study region
was verified and corrected. Burn-P3 simulations were run under the settings related to weather
recommended in the software documentation and were found to be fairly robust to errors in the
fuel map, but simulated fire sizes were substantially larger than those observed in the historic
record. By adjusting the input parameters to reflect suppression eects, we obtained a model which
gives more appropriate fire sizes. The resulting burn probability map suggests that risk of fire in
the study area is much lower than what is predicted by Burn-P3 under its recommended settings.
1. Introduction
Fire is a naturally occurring phenomenon on the forested landscape. In Canadas boreal forest
region, it plays an important ecological role. However, it also poses threats to human safety
and can cause tremendous damage to timber resources and other economic assets.
Wildfires have recently devastated parts of British Columbia, California, and several
other locations in North America, Europe, and Australia. The economic losses in terms of
suppression costs and property damage have been staggering, not to mention the tragic
loss of human life. Many of these fires have taken place at the wildland-urban interface
predominantly natural areas which are increasingly being encroached upon by human
habitation. As the population increases in these areas, there would appear to be potential
for increased risk of economic and human loss.
A wildland-urban interface is defined as any area where industrial or agricultural
installations, recreational developments, or homes are mingled with natural, flammable
vegetation 1. The Province of Ontario has several areas which could be classified as
wildland-urban interface. These areas include the Lake of the Woods region, the Thunder
Bay region, the region surrounding Sault St. Marie, and North Bay among others. One of the
most significant of these is the District of Muskoka which is a popular recreational area. This
district, located in Southern Ontario Figure 1, is commonly referred to as cottage country.
It spans 6,475 square kilometers and contains over 100,000 seasonal properties or cottages.
Many of these properties are nestled in forested areas, which make up most of the region.
This concentration of values is of particular interest to the Canadian insurance industry due
to the risk of claims from damage caused by wildfire.
Unlike British Columbia and California where topography plays a major role in the
rate of spread of wildfire, Ontario is relatively flat but is dominated geographically by the
Boreal and Taiga forests, where some of the largest fires in Canada have burned 2. The
Boreal forest has a large percentage of coniferous trees which are susceptible to high intensity
crown wildfires. The Muskoka region is on the southern edge of the Boreal forest, and thus
there is potential for substantial property damage from fires originating further north.
We are focusing on the Muskoka region to provide an illustration of how the tools
that have been developed by the forest management community can be applied to assess fire
risk. The methods described here can be adapted easily to other wildland-urban interface
locations. The Muskoka area presents some technical challenges which do not exist to the
same degree in most other wildland-urban interface settings.
Although there have not yet been substantial losses due to wildfire in the Muskoka
area, it is important to assess the risk because of what is being observed elsewhere e.g.,
British Columbia and California and because of possible climate change eects which could
ultimately lead to increased fire activity across Canada.
Wildfires usually start from point ignitions, either by people or by lightning, and if
not detected immediately, they can spread rapidly under appropriate weather conditions.
Approximately half of the forest fires in Canada are ignited by lightning. Such fires account
for approximately 80 percent of area burned 3.
The spread of a wildfire in a particular area depends on many factors but most
importantly, it is influenced by local weather, vegetation, and geography 2. Of these three
factors, the geographical features remain static, while vegetation changes gradually over
time. In addition, changes in human land use patterns, such as industrial forestry, or urban
expansion can lead to changes in vegetation. Weather is the most dynamic factor aecting
fire risk. The unpredictable nature of weather makes modelling forest fire spread a dicult
task. Nonetheless, the risk of wildfire in a region can be estimated using the methodology
described in this paper.
In Canada, the fire season can last from early April through October each year. During
this period, the probability of fire ignition and fire spread potential changes depending on
the time of year, primarily influenced by seasonal weather patterns. Each year an average of
2.5 million hectares are burned by 8,500 individual wildfires.
Most regions which are within the Boreal and Taiga zones have very accurate and
up-to-date fuel information because the provincial fire management agencies maintain these
records rigorously. The forest resource inventory information in our study area, and hence
the fuel map which is based upon it, is not updated as frequently by the Ontario Ministry of
Figure 1: Location of the District of Muskoka within the Province of Ontario.
Natural Resources in this region because there is a higher proportion of private land under
municipal fire protection agreements with the province and relatively little area under forest
management planning. Thus, it was necessary for us to validate the fuel map by doing a field
survey. To apply the methodology in other instances would be straightforward, not requiring
this kind of fieldwork.
The remainder of this paper will proceed as follows. The next section provides a
description of the study area and the fire data for that region. Section 3 contains results of
an ignition risk assessment which uses historic fire data only. This section also contains a
crude burn risk assessment.
In Section 4, we briefly describe the Prometheus fire growth model 4 and how it
is used in the Burn-P3 simulator 5 to generate a burn probability map. This section also
provides a description of the required data inputs and the procedure that was used to obtain
and verify this data. In Section 5, the results of the analysis are presented along with a
summary of the limitations of this study.
2. The Data and Study Area

2.1. Study Area
Of the properties in the Muskoka District, the most expensive are concentrated along the
shores of the three major lakes: Lake Joseph, Lake Muskoka, and Lake Rosseau. A 25 35 km
rectangular study area that encompasses a large portion of these lakes was selected Figure 2
for the our study. In order to reduce possible biases near the boundaries of this region, we also
considered a 5-km wide buer zone which surrounds the study area. Fires originating in
this zone could spread into the study area, and this possibility needs to be accounted for.
2.2. Description of Historic Fire Data

Fire data for over 12,200 fires from 1980 through 2007 were obtained for a region
encompassing the study area. For each fire, a number of covariates were recorded including
kilometers
0 2.5 5 10
Figure 2: Map illustrating the 25 35 km study area which is enclosed in the red box as well as the buer
zone used in the Burn-P3 modelling denoted by the blue box.
the date, ignition location, and final area. Figure 3 shows an estimate of the density of the
natural log-transformed fire sizes of escaped fires from this dataset. Here, we use the Ontario
Ministry of Natural Resources definition of an escaped fire: any fire where final area exceeds 4
hectares. The fuel composition, weather conditions, and fire suppression capabilities for this
region are relatively homogeneous and hence are representative of our smaller study area.
Within this dataset, 319 fires were located in the study area. Figures 4 and 5 show locations
of human-caused and lightning-caused ignitions
3. Ignition and Burn Probability Modelling Using Generalized

Additive Models
3.1. Ignition Modelling
Brillinger et al. 6 provide a method for assessing fire risk in a region using generalized
additive models. Their technique uses pixellated data on a fine scale where each pixel is
assigned a 1 or a 0 depending on whether or not a fire was ignited at that location. To be
precise, they considered temporal eects as well, while our focus will be to produce only a
spatial risk map. The resulting data set is very large with an overwhelming proportion of
1 1 km pixels sites without a fire ignition between 1980 and 2007, indicated by a value of 0.
However, a simple random sample of these 0-sites can be analyzed in the same way as the full
dataset with the addition of an oset of the form log1/s . Here s denotes the constant
inclusion probability for site s s1 , s2 , where s1 and s2 refer to the easting and northing
geographic coordinates, respectively.
0.6
0.5
0.4
Density
0.3
0.2
0.1
0 2 4 6 8
ln area in hectares
Figure 3: Estimated density of natural log-transformed fire sizes of historic escaped fires 19802007.
N
kilometers
0 2.5 5 10
Figure 4: Map of human ignition locations in study area 19802007.
We have explored our data set with a simple model from within this family of models:

1
logit ps fs1 , s2 log , 3.1
s
N
kilometers
0 2.5 5 10
Figure 5: Map of lightning ignition locations in study area 19802007.
where ps probability of ignition at site s and fs1 , s2 is a penalized tensor product spline
smoother using the cubic B-spline basis in each dimension see 19, Chapter 4. We have taken

0.01 if sites do not contain an ignition i.e., a 0
s 3.2
1 if sites contain an ignition i.e., a 1.
We chose this value for s in order to have a manageable data set which has sucient
covariate information for inference. The resulting ignition risk map is shown in Figure 6. We
note that there is a relatively high risk of ignition in the southeast region. This is the region
closest to the town of Gravenhurst. The rest of the region is less heavily populated, and thus
less likely to be subject to human-caused ignitions.
3.2. Simple Burn Probability Map

We also used the above modelling approach to assess the probability of burning by applying
the same methodology but instead of assigning a value of 1 to a pixel that had an ignition, we
assign a value of 1 to pixels that have burned, either directly by an ignition or spread from an
ignition point. Unfortunately, actual final fire shape was not available in the database, so we
made a crude approximation based on the observed final burned areas. The resulting burn
probability map is pictured in Figure 7. Notice the decreased fire risk near the town and the
increase in fire risk to the north and to the west. Because of the proximity to town, it may be
that the ignited fires in the southeast may be suppressed relatively quickly, leading to smaller
burned area. This phenomenon has been well documented e.g., 2.
N
kilometers
0 1 2 4
0.000050.00043 0.002340.00271 0.004230.0046
0.000440.00081 0.002720.00309 0.004610.00498

0.000820.00119 0.00310.00346 0.004990.00536
0.00120.00157 0.003470.00384 0.005370.00574

0.001580.00195 0.003850.00422 0.005750.00612
0.001960.00233
Figure 6: Model of ignition risk using generalized additive models with historic ignition data.
In addition to the loss of accuracy due to incorrect fire shape, the presence of relatively
large lakes in the study area causes some diculties for the smoother; essentially, boundary-
like eects are introduced into the interior of the region. Furthermore, vegetation type and
presence of other nonfuel fire barriers is not accounted for in this model.
For these reasons, we are motivated to consider a dierent modelling approach which
is based partially on a physical model for wildfire growth and which incorporates fuel and
fuel breaks. However, this map, based on historic records, can serve as a partial check on the
reasonableness of the model we will propose next.
4. Burn Probability Modelling Using a Fire Growth Model

4.1. The Prometheus Fire Growth Model
Another method of forest fire risk assessment is based on computer simulation of fires, taking
account of fuel information and local weather patterns. To model fire growth, we will employ
the Prometheus Fire Growth Model 4.
N
kilometers
0 1 2 4
00.00043 0.002560.00298 0.004690.0051
0.000440.00085 0.002990.0034 0.005110.00553

0.000860.00128 0.003410.00383 0.005540.00595
0.001290.0017 0.003840.00425 0.005960.00638

0.001710.00213 0.004260.00468 0.006390.0068
0.002140.00255
Figure 7: A simple burn probability map using generalized additive models with historic ignition data.
The evolution of a fire front simulated by Prometheus relies on the theory developed
by Huygens for wave propagation: each point of a fire front at a given time acts as an ignition
point for a small fire which grows in the shape of an ellipse based at that point. The size
and shape of each ellipse depend on fuel composition information, weather, and various fire
growth parameters as well as the time duration. The envelope containing all of the ellipses is
taken to be the fire perimeter at the next time step Figure 8.
In the absence of topographic variation, the orientation of each ellipse is aligned with
the direction of the wind. The shapes of the ellipses at each time step are calculated from
empirical models based on the Canadian Fire Behaviour Prediction FBP system which is
described in the next subsection. The length of each ellipse is related to a local estimate of
the forward rate of spread plus an analogous estimate of the back rate of spread, while the
width of an ellipse is related to a local estimate of the flank rate of spread. These local rates of
spread are, in turn, inferred from the empirical FBP models which relate spread rate to wind
speed, fuel moisture, and fuel type. The measurements required for this calculation are based
on local estimates of the weather conditions which have been extrapolated from the nearest
reliable weather station. Diurnal changes in fuel moisture as it is aected by temperature
and relative humidity and wind speed are also incorporated into the model.
4.2. Canadian Fire Behaviour Prediction System and Fire

Weather Index System
In Canada, forest fire danger is assessed via the Canadian Forest Fire Danger Rating
System CFFDRS. As described by Natural Resources Canada 7, the current form of this
system has been in development since 1968. The structure of the CFFDRS is modular and
currently consists of four subsystems. Two of these subsystems are of interest in our study:
the Canadian Fire Weather Index FWI System and the Canadian Forest Fire Behaviour
Prediction FBP System, both of which are fully documented and are used operationally
across Canada.
Many parts of the CFFDRS rely on information obtained using the FWI System. This
system is comprised of six components which summarize aspects of the relative fire danger
at its midafternoon peak 8. All calculations are based on locally observed weather readings
recorded at local noon: temperature, relative humidity, wind speed usually a 10-minute
average, and rainfall over the last 24 hours. Three fuel moisture codes are calculated, each
representing the dryness in a dierent layer of the forest floor. Three fire behaviour indices,
estimating the risk of fire spread, the fuel available for combustion, and the potential intensity
of a fire, are also calculated. For a recent exposition on the CFFDRS, see the account by Wotton
9 which is a review designed for modellers who require an understanding of this system
and how it is to be interpreted.
The FWI System is used to estimate forest fire potential. Its outputs are unitless
indicators of aspects of fire potential and are used for guiding fire managers in their decisions
about resource movements, presuppression planning, and so forth. However, this is only
a part of fire management. There is also the need, once fires have begun, to estimate
characteristics of fire behaviour at a point on the landscape; this is done with the FBP System.
Given inputs that fall into one of five categoriesfuels, weather, topography, foliar
moisture content, and type and duration of predictionthe FBP System can be used to
estimate fire behaviour quantitatively 10. The FBP System calculations yield four primary
and eleven secondary outputs as fire behaviour indices. It gives estimates which can be used
as the basis for predictions.
The primary outputs are Rate of Spread, Fuel Consumption as either the surface or
crown consumption, or total, Head Fire Intensity, and a fire description code crown fraction
burned and fire type. The secondary outputs are Flank and Back Fire Rate of Spread; Flank
and Back Fire Intensity; Head, Flank, and Back Fire Spread Distances; Elliptical Fire Area;
Fire Perimeter; Rate of Perimeter Growth; and Length-to-Breadth Ratio. The primary outputs
are based on a fire intensity equation and the secondary outputs are determined by assuming
elliptical fire growth. All underlying models and calculations are based on an extensive 30-
year field experimental burning program and are fully documented 10.
4.3. Burn-P3 Simulation Model

A burn risk probability map can be generated using the Burn-P3 simulation model software
developed by Marc Parisien of the Canadian Forest Service 5. P3 stands for Probability,
Prediction, and Planning. Burn-P3 runs repeated simulations of the Prometheus Fire Growth
a b
Figure 8: Illustration of fire perimeter growth under uniform burning conditions for homogenous fuels a
and nonhomogenous fuels b 11.
Model, under dierent weather scenarios, to give estimates of the probability distribution of
locations being burned during a single fire season.
In each iteration of a Burn-P3 simulation, a pseudorandom number is generated
and used to sample a number from the empirical distribution of the annual number of
escaped fires in the region. This empirical distribution is based on historic data. This number
represents the number of simulated fires for one realization of one fire season.
For each of these fires, a random cause, season, and ignition location combination
is selected from an ignition table. Burn-P3 creates an ignition table by combining ignition
grids for each cause/season combination. Ignition grids partition the study area into coarse
cells and represent the relative likelihood of a fire occurrence of an ignition in each cell.
This spatial distribution can be empirically based on historic ignition patterns or it can be
a uniform distribution, for example. The probability of selecting a certain row in the ignition
table is proportional to the ignition probability of that particular cell specified in the matching
ignition grid.
The duration of each simulated fire is also randomly drawn from an empirical
fire duration distribution based on historic data. Given the location and fuel conditions,
the Prometheus program is then used to simulate the growth of each fire individually
given a random weather stream consisting of conditions conducive to fire growth from
the appropriate season. All simulated fires in a single iteration are collectively used as an
independent realization of a fire season.
Repeatedly simulating such fire seasons allows for construction of a burn probability
map. Specifically, dividing the number of times each cell in the rasterized map of the study
region has been burned by the number of simulations run gives an estimate of the probability
that the particular cell will burn in a single fire season. See Figure 9 for a step-by-step
illustration of this process.
The version of Burn-P3 used in this paper is not programmed to handle vectorized
fuel breaks, that is, features in the landscape which tend to prevent fire from spreading. All
fuel breaks such as roads are rasterized in Burn-P3 which sometimes leads to anomalous
behaviour where a simulated fire passes between grid cells connected only at a single vertex.
By using a small grid cell size, we can avoid this problem.
1 fire 2 fires
a b
3 fires 5 fires
c d
10 fires 30 fires
e f
Figure 9: Step-by-step illustration of 30 iterations of the Burn-P3 simulation model. Darker colours indicate
areas that have been burned more often. Green areas are unburned fuels. White areas represent nonfuel.
a The yellow patch represents a single fire. b The two yellow patches represent two fires occurring in
two dierent years. c Yellow patches denote areas burned by one of 3 fires occurring in dierent years.
The orange patch represents an overlap of 2 of these fires. d The red patch represents an area burned by
fires in 3 or more dierent years.
N
kilometers
0 2.5 5 10
Fuel type
BS C-5 O-1
C-1 C-6 S-1
C-2 D-1 Water

C-3 M-1 Bog
C-4 M-3 Nonfuel
Figure 10: Corrected vector fuel map.
4.4. Inputs
4.4.1. Fuel Map
The FBP System classifies vegetation into 16 distinct fuel types Table 1, that can further be
grouped into the five categories: coniferous, deciduous, mixed wood, slash, and open 10.
A map of the fuel types in the District of Muskoka was obtained from the Ontario Ministry
of Natural Resources. The fuel map was created manually from aerial photography in 1994.
Fieldwork was conducted over the course of 7 days to verify and correct a subsample of the
fuel map which appears in Figure 10.
4.4.2. Verification in the Field

Regardless of the accuracy of the fuel map at the time of its creation, fuel types and extents
change over time due to land use changes, urban expansion, and natural causes such as
forest succession. For example, in the study area, a large area of fuel mapped as C-6 Conifer
Table 1: Fire Behavior Prediction System fuel types 10.
Group/Identifier Descriptive name

Coniferous
C-1 Spruce-lichen woodland
C-2 Boreal spruce
C-3 Mature jack or lodgepole pine
C-4 Immature jack or lodgepole pine
C-5 Red and white pine
C-6 Conifer plantation
C-7 Ponderosa pine-Douglas-fire
Deciduous
D-1 Leafless aspen
Mixedwood
M-1 Boreal mixedwood-leafless
M-2 Boreal mixedwood-green
M-3 Dead balsam fir mixedwood-leafless
M-4 Dead balsam fir mixedwood-green
Slash
S-1 Jack or lodgepole pine slash
S-2 White spruce-balsam slash
S-3 Coastal cedar-hemlock-Douglas-fir slash
Open
O-1 Grass
Plantation was found to be harvested and hence, was reclassified Figure 11. Not all areas
were accessible by public roads and thus could not be verified by fieldwork. Consequently,
satellite imagery was used to further supplement our fieldwork to help confirm such areas.
To get an estimate of the accuracy of the fuel map, multistage cluster sampling
procedure was carried out in the field. First, 20 roads were selected at random with
probability proportional to the length of the road Figure 12. For each of these roads,
observations were taken at the beginning, the end, and at various points along the road.
The number of observations taken was randomly generated from a Poisson distribution
with rate equal to the length of the road in kilometers Table 2. The exact locations of these
observations were randomly selected from a uniform distribution from the beginning to the
end of the road.
At each observation location, the width of the road including shoulder was measured
and recorded. One person without prior knowledge of the given fuel classification gave his
best assessment of fuel classification of the fuels on either side of the road, making sure to
look beyond the immediate vegetation at the tree line. Both the assessed fuel classification
and original fuel classification were recorded.
Three of these selected roads were privately owned or not maintained enough to be
traversable; these were not included in the sample. A summary of results is given in Table 3.
The subjective nature of fuel classification can be seen in the 81.1% misclassification rate
Figure 11: Area mapped as C-6 that has since been harvested and required reclassification.
assuming no tolerance for classification error. However, not all dierences in observations
can be considered practically important. For example, an area assessed as 10% mixed wood
originally classified as 20% mixed wood was not updated because the resulting change in fire
behaviour is very slight; the rate of spread changes marginally and the direction of spread
would not be aected at all. On the other hand, if a nonfuel was incorrectly classified as
some form of fuel in the original map, the correction was made because the dierence in fire
behaviour could be substantial. Using this criterion, the misclassification rate was found to
be to 22.7% in our sample: most of the fuel types were close to what we assessed them to
be.
4.4.3. Unmapped Private Properties

The main properties of interest are those located along the waterfront because they represent
the highest concentration of values at risk. Unfortunately, on the fuel map obtained from the
Ontario Ministry of Natural Resources, such areas are almost always mapped as nonfuels
because this is a private land not included in the Forest Resource Inventories on which the
fuels classification is based. From what was observed in the field, most of these properties
are located very close to fuels and could in fact be treated as a separately defined fuel type
Figure 13. The fuel types in these properties are almost always similar to what is located on
the opposite side of the road further inland. For the purpose of this paper, we will assume
that forest stands are continuous across roads into waterfront properties mapped as nonfuel.
Some properties have isolated buildings separated from surrounding fields e.g., by a
well-kept lawn or driveway. Although these buildings are unlikely to be damaged directly
by a wildfire, they are still at risk to ignitions caused by spotting. Fire spotting is the
situation where firebrands are transported long distances by the wind to start new fires. On
waterfront properties, isolated buildings are not common. Thus, we further assume that all
such properties have the same fuel type as the surrounding area.
LA RD
SK
CEM
EL
HEKK
ETE
ET
RY R
D
ON
LA
KE
RO
LU
A
CK
2
D
RD
EY
E
3
AK
EL
RD
MIL
EE
THR
RDNORTHS
Y HORE RD
RD
PU DEERWOOD DR
DAW
SON D
RD ER
SR
D NLIN
OW
L ET
AL C TRIC
H LE BEA
A
RW
AT
ER
SH
O
RE RD R RD
CR
T O KA
S IN ZIS
A
CK
BL PO
N
S
BE
V
D ZY A
RR
N
EE D
Y
BR
ASHFOR
LA
RD
FOGO ST N
MU E
SKO
KA
CRE
TH RD
S CE
NTR
D
(kilometeres)
0 2.5 5 10
Figure 12: Map of the 20 randomly selected roads to be sampled.
Table 2: Names of the 20 randomly selected roads to be sampled, the number of observations taken on
each road, and the location of each observation.
Name of Road Number of observations Locations along road km

Cemetery Road 3 0, 0.5, 0.8
Hekkla Road 5 0, 1.0, 2.3, 2.9, 4.3
Deerwood Drive 2 0, 0.7
Purdy Road 4 0, 0.1, 1.4, 2.4
North Shore Road 8 0, 0.4, 0.6, 3.0, 3.7, 3.9, 4.9, 5.3
Skeleton Lake Road 3 2 0, 1.2
Three Mile Lake Road 2 3 0, 2.5, 3.3
Walkers Road 10 0, 1.9, 2.7, 3.6, 3.7, 4.4, 5.1, 5.5, 5.6, 6.5
Luckey Road 7 0, 1.0, 1.4, 2.4, 2.5, 3.1, 4.0
Clearwater Shores Blvd 2 0, 1.5
Halls Road 2 0, 0.8
Dawson Road 8 0, 0.1, 1.7, 2.0, 2.1, 2.5, 2.6, 2.9
Beatrice Townline 9 0, 1.0, 2.1, 5.2, 7.1, 7.5, 7.6, 8.0, 8.4
Cranberry Road 4 0, 0.6, 1.9, 2.1
Fogo Street 3 0, 1.3, 1.8
Breezy Point Road 5 0, 0.5, 0.9, 1.3, 2.6
Ashworth Road 4 0, 0.3, 1.2, 1.7
Ziska Road 8 0, 0.9, 1.9, 3.3, 4.6, 4.9, 5.2, 5.3
Rock Sand Lane 3 0, 0.2, 0.7
Muskoka Crescent Road 2 0, 6.1
Inaccessible roads.
Table 3: Summary of fuel map accuracy sampling procedure.
Tolerance Margin % Misclassified

Zero tolerance 81.1%
Little practical dierence 22.7%
Figure 13: A structure embedded in a region of continuous fuel.
4.4.4. Fuel Breaks

As we have seen, fuel breaks are wide regions of what are essentially nonfuels that have the
potential to prevent a fire from spreading across. The most common fuel breaks are roads,
water bodies, and power lines.
The fuel map classifies major roads as nonfuels. Smaller private roads have not been
included on the map. In some cases, they are located within larger private property regions
mapped as nonfuel. We traced such roads from road data provided by the National Road
Network. These data contain the locations of almost all roads as well as the number of lanes
per road. We took a random sample of roads of varying numbers of lanes to measure their
eective widths. Not only are lane widths not uniform on smaller roads, the size of the
shoulder and distance to the tree line are highly variable. We used an average of sampled
road widths for roads for which the widths are unknown or cannot be reasonably estimated.
Bodies of water such as lakes and rivers are clearly and accurately identified on the
fuel map. However, bogs and swamps are problematic, since they are occasionally classified
as water. Although a suciently dry bog could potentially become a fuel source in the heat
of summer, the current fire growth model does not account for such a phenomenon.
Power lines introduce a further diculty, since they are not identified on the fuel map.
When power lines are built in a forested area, a path is cleared and growth underneath
the line is regularly maintained. The width of this clearing and the amount of growth
directly underneath the power line vary depending on how regular such maintenance occurs.
Without a map of the smaller power lines, power lines are assumed to be negligible as fuel
breaks.
4.4.5. Rasterization
The corrected vector fuel map must be converted to a raster map before it can be used by
the Burn-P3 program. In doing so, detail at resolutions smaller than the grid cell resolution
of the raster fuel map may be lost. However, refining the resolution of the raster fuel map
N
kilometers
0 2.5 5 10
Fuel type
BS C-5 O-1
C-1 C-6 S-1
C-2 D-1 Water

C-3 M-1 Bog
C-4 M-3 Nonfuel
Figure 14: Rasterized fuel map at 25 m resolution.
directly increases computation time. In this assessment, a 25 m resolution Figure 14 was
used. A coarser resolution of 150 m was also tested, but we have not included the results of
this rasterization, because important features such as fuel breaks were not respected.
4.4.6. Historic Weather

We obtained weather data from surrounding weather stations. The most complete weather
record also happens to be the station closest to the study region so only weather data from
this station was used in our analysis. This weather record begins with the 1980 fire season.
Only the extreme weather days are used for input to the Burn-P3 program. These days,
in which there is potential for substantial fire growth, are referred to as spread event days.
Days when the initial spread index ISI is less than 7.5 are considered to be nonextreme
and are deleted from the weather stream. The resulting data set consists of 232 cases, each
representing to a single day. Given the large number of simulations to be run, these weather
conditions are sampled frequently.
4.4.7. Seasons and Causes

To properly simulate the growth of fires for an entire year, fires need to be classified by cause
and by season in which they occur.
In regions with a mixture of deciduous and coniferous vegetation, there are often two
distinct fire subseasons each year, the first immediately following snowmelt, and the second
in summer. The spring fire subseason is a period of increased fire risk because leaves have
not appeared on the deciduous trees leading to drier surface conditions. Since there is limited
lightning activity during this period, ignitions are primarily due to people.
After the leaves appear, there is often a brief interval with few fires. Fire occurrence
increases with temperature increase and as lightning activity increases. Thus, during the
summer fire subseason, ignitions are due to people and lightning. This results in dierent
spatial ignition patterns depending on time of year.
Operationally, early to mid June is typically taken as the transition date between
the spring and summer seasons when classifying fires. However, this date is inferred from
observations taken in the northern boreal forest. The District of Muskoka is further south
and experiences a slightly warmer climate. Consequently, it is reasonable to assume that the
actual transition date occurs approximately a week earlier. Looking at the distribution of
human caused fires Figure 15, we can see a dip on June 11th in human caused ignitions
which is used as an estimate of the transition date and is highlighted with a vertical line in
the figure.
All fire ignitions can be classified into either human- or lightning-caused fires. Human-
caused fires can be further subdivided into eight specific causes: recreational, residential,
railway, forestry industrial, nonforestry industrial, incendiary, miscellaneous, and unknown.
Each of these causes is associated more strongly with either the spring or summer season
Figure 16. Between 1996 and 2005 inclusive, across the province, there were a total of 12,974
wildfires resulting in over 1.5 million hectares burned. Of these ignitions, nearly 7,000 can be
attributed to lightning.
4.4.8. Ignition Grids

For each season and cause combination, ignition grids were created to represent the relative
likelihood of an ignition in a certain cell. To create the ignition grids for human caused fires,
a grid twenty times coarser than the fuel grid was created and assigned a value of 1. For each
historic fire, the cell in which the ignition occurred had its value incremented by 1. Lightning
ignitions in the region appear to be uniformly random and so a uniform ignition grid was
used for lightning-caused ignitions.
4.4.9. Estimation of Spread Event Days

Burn-P3 models fire growth only on spread event days. The number of spread event days is
not necessarily equal to the total duration of a fire; there may be days for which a fire does
not spread either due to suppression or nonconducive weather. The fire data do not contain
information on fire spread days so a best estimate is made by counting the number of days
from the fire start date to the date in which the fire was reported as held. A held fire is one
that has been completely surrounded by fire line i.e., fuel breaks constructed by suppression
eorts as well as naturally occurring fuel breaks. The historic data on the distributions of
spread event days per fire, as defined previously, is displayed in Figure 17.
Human caused fires

Distribution of human caused fires
500
0.008
Density
Frequency
300
0.004
100
0 0
50 100 150 200 250 300 350 0 100 200 300 400
Julian date Julian date
a b
Figure 15: Distribution of human caused forest fires during the year. The vertical line indicates the
minimum of the dip, corresponding to June 11th.
The FBP Systems rate of spread is based on peak burning conditions, which are
assumed to occur in the late afternoon, generally specified as 1600 hours 10. Consequently,
any fire growth models based on this system are eectively simulating spread event days
12.
5. Analysis and Results

We will begin this section by discussing the application of the Burn-P3 simulator to the
corrected fuel map when the recommended settings are used. We will then discuss the
results of a sensitivity analysis in which the eects of fuel break misclassification on the
burn probabilities are studied. This will provide an indication of the uncertainty induced
by possible inaccuracies in the fuel map. We next compare the burn probability map obtained
from Burn-P3 with the map obtained using generalized additive models, and finally compare
simulated fire size distributions with the historical record; this ultimately guides us to what
we believe is a more accurate burn probability map.
A uniform ignition grid for lightning caused fires is normally recommended for use
in Burn-P3. For human-caused fires, the ignition grid is also uniform, but with increased
probability at locations where previous ignitions occurred. The distribution of spread event
days is based on historic weather data where the weather stream has been adjusted so that it
only contains extreme weather, conducive to fire growth. Using these recommended settings,
we obtain the burn probability map and fire size distribution as shown in the left panel of
Figure 18.
In that figure, it can be seen that the fire risk is higher in the north. This result seems
plausible, since there are larger forest stands in that region, and we have already conjectured
on the possibility of large fires spreading into the study region from further north.
In order to assess the uncertainty induced by possible inaccuracies in the fuel map,
we conducted the same simulation but with 20% of randomly selected nonfuel grid cells
converted to the M-1 20% mixed wood fuel type. By making this kind of change, we should
observe the largest range of realistic fire behaviour in the study region, since much of the
forest in the area is of M-1 type, and changes within this categorization have minimal eect
on fire behaviour. By contrast, changes from nonfuel to any kind of of fuel can have relatively
Residential Railway Recreational
800 100 1000

800
Frequency
80
Frequency
Frequency
600
60 600
400
40 400
200 20 200
0 0 0
50 150 250 350 100 150 200 250 300 50 150 250 350
Julian date Julian date Julian date
a b c
Miscellaneous Industrial non-forestry Unknown
250 30
40
200 25
Frequency
Frequency
Frequency
20 30
150
15 20
100 10
50 5 10
0 0 0
100 150 200 250 300 100 150 200 250 300 100 150 200 250 300
d e f
Incendiary Industrial forestry Lightning

6 400
60 5
300
Frequency
Frequency
4
Frequency
40 3 200
2
20 100
1
0 0 0
50 150 250 350 50 100 150 200 250 300 100 150 200 250 300
g h i
Figure 16: Histograms of forest fires during the year by cause. Vertical line indicates June 11th, the
estimated transition date used to separate spring and summer subseasons.
dramatic eects on fire behaviour, since nonfuel regions often serve as fuel breaks; replacing
parts of such regions with fuel allows for the possibility of a fire breach where it would
not otherwise have been possible. As expected, the burn probability map Figure 18b
exhibits larger regions of relatively high probability than in the original map, especially in
the eastern region as well as in the north. Note that regions where the burn probability was
already relatively high do not see a substantial gain in burn probability when the nonfuels
are perturbed. Rather, we see somewhat more substantial increases in burn probability in
those areas where the probability was much less. An additional simulation was run with
only 10% of the nonfuel randomly converted to fuel; we have not shown the resulting map
because of its similarity to the map in the right panel of Figure 18. We conclude that gross
misclassification of fuel as nonfuel could lead to a moderate underestimate of the area at
elevated risk.
Estimated spread event days
5000
Frequency
3000
1000
0
0 5 10 15
Number of spread days
Figure 17: Histogram of estimated spread event days.
Comparing the map obtained from Burn-P3 using the corrected and unperturbed
fuel map Figure 18 with the burn probability map obtained using the generalized additive
model Figure 7, we see some similarities. Both maps exhibit elevated burn risk in the north,
but the prevalence of ignitions in the southeast seems to figure more prominently in the
map obtained from the generalized additive model. It should be noted that the latter map
is based on accurate fire sizes but incorrect fire shapes, while the former map is based on
what are possibly more realistic fire shapes, but with a fire shape and size distribution that is
determined by the weather and fuels.
We can then use the historic fire size distribution as a check on the accuracy of the
Burn-P3 output. Figure 19 shows the estimated density of fire sizes in the study region solid
black curve on the natural logarithmic scale. The dashed curve represents the estimated
density of the simulated log fire sizes under the recommended settings, and the dotted curve
corresponds to the perturbed nonfuel simulation. Both densities fail to match that of the
historic record. Modal log fire sizes are close to 2 in the historic record, while the simulations
give modes exceeding 5. Note that, in accordance with our earlier observations regarding the
nonfuels, the fire sizes indeed increase when fuel breaks are removed.
In order to find a model which matches the historic record more closely, we could
introduce additional fuel breaks, but we have no way of determining where they should be
located without additional substantial fieldwork, and the earlier sensitivity study indicates
that even fairly substantial errors in the fuel map will lead to only modest discrepancies in the
fire size distribution. Instead, it may be more important to consider the eects due to weather.
To investigate this, we have run four additional Burn-P3 simulations under dierent settings.
The resulting burn probability maps appear in Figure 20. We now proceed to describe these
simulations and their resulting fire size distributions.
First, we replaced the spread event day distribution with a point mass at 1 day. In
other words, we made the assumption that even if fires in the area burn for several days,
there would only be one day in which the fire would burn a nonnegligible amount. All other
simulation settings remain as before. The fire size distribution for this situation is pictured in
Figure 19 as the long-dashed curve, having a mode near 4. This is closer to the historic record,
but still unsatisfactory.
N N
kilometers 0 1 2 4 kilometers 0 1 2 4
00.000181 0.0014450.001625 00.000181 0.0014450.001625
0.0001820.000361 0.0016260.001806 0.0001820.000361 0.0016260.001806

0.0003620.000542 0.0018070.001986 0.0003620.000542 0.0018070.001986
0.0005430.000722 0.0019870.002167 0.0005430.000722 0.0019870.002167
0.0007230.000903 0.0021680.002347 0.0007230.000903 0.0021680.002347

0.0009040.001083 0.0023480.002528 0.0009040.001083 0.0023480.002528
0.0010840.001264 0.0025290.002708 0.0010840.001264 0.0025290.002708
0.0012650.001444 0.0027090.002889 0.0012650.001444 0.0027090.002889

a b
Figure 18: a Burn probability map of simulated fires using recommended Burn-P3 settings. b Burn
probability map of simulated fires using recommended settings and 20% of nonfuels randomly converted
to M-1 fuel type.
The next simulation made use of the entire weather record, dispensing with the notion
of spread event days completely. Fire durations were sampled from historic fire duration
distribution. Again, all other simulation settings were the same as before. The resulting fire
size distribution is displayed in Figure 19 as the dashed-dotted curve, having a mode near
3a substantial improvement, but still not satisfactory. The dierence between this result
and the earlier simulations which depend only on extreme weather calls such practice into
question.
In the succeeding simulation run, the duration of the fires was reduced to a single day,
again sampling from the full weather stream. The resulting density estimate is displayed in
Figure 19 as the long-dashed-dotted curve, having a mode near the historic mode, although
its peak is not nearly as pronounced. An additional simulation was conducted using the
same settings but with an ignition grid based on the generalized additive model for ignitions
obtained in Section 3.1. The resulting fire size distribution is also pictured in Figure 19 and is
very similar to the result of the preceding simulation.
We conclude that the use of the full weather stream and that limiting the duration of
the fires to one day give more accurate fire size distributions. Use of the uniform ignition grid
is slightly less accurate than the use of the modelled grid based on historic ignitions.
0.6
0.5
0.4
Density
0.3
0.2
0.1
0
0 2 4 6 8
log area in hectares
Historic 1 spread event day

Default Full weather 1 spread event day
Perturbed non fuels Full weather 1 spread event day
Full weather GAM ignition grid
Figure 19: Estimated density functions of log-transformed simulated and observed fire sizes under various
scenarios. The historic fire size density is based on data from an area encompassing the study region 1980
2007. The other fire curves are based on Burn-P3 simulations, using the default setting with the corrected
fuel map as well as a fuel map with 20% of the nonfuel randomly selected to be changed to mixed wood-
type fuel. The Full weather curve is based on a simulation using all weather data but with the same spread
event day distribution as before.
6. Discussion
We have shown how to estimate a burn probability map which could be used by insurers
to estimate expected losses due to wildfire risk in the region under study. We found that
substantial perturbation of the fuel map, converting nonfuels to fuels, gives rise to moderate
changes in the fire risk.
We have also used historic fire size distribution information as a check on its accuracy
and found that the recommendation to use a spread event day distribution for fire duration
overestimates the fire size distribution. The use of spread event days in the Burn-P3 model
could be degrading the probability estimates. The spread event day distribution may be
biased since it is based on the time between when a fire was first reported and when it was
declared as being successfully held by suppression activities. A fire would not necessarily be
spreading rapidly during this entire period.
Note that fire suppression is not accounted for directly in Burn-P3. This could account
for the dierence between the simulated and observed fire size distributions. By using the
full weather stream and a one day fire duration, the simulated fire size distribution comes
closer to matching the historic record. In fact, using the 1 day fire duration may be realistic
because of suppression eects; it is unlikely that most fires are allowed to burn for more than
1 day in this region without being attacked. If allowed to burn longer, it would not be under
extreme weather conditions, and such fires would not be spreading fast.
Thus, there is some justification for our approach. We note, however, that there is still
a discrepancy between the simulated fire size distribution and the historic record. As we saw,
fuel/nonfuel misclassification can have a modest eect on the fire size distribution estimates.
It is possible that some of the small roads that are not classified as fuel breaks may in fact
N N
a b
N N
00.000057 0.0004550.00051 00.000057 0.0004550.00051
0.0000580.000113 0.0005110.000567 0.0000580.000113 0.0005110.000567

0.0001140.00017 0.0005680.000624 0.0001140.00017 0.0005680.000624
0.0001710.000227 0.0006250.000681 0.0001710.000227 0.0006250.000681
0.0002280.000284 0.0006820.000737 0.0002280.000284 0.0006820.000737

0.0002850.00034 0.0007380.000794 0.0002850.00034 0.0007380.000794
0.0003410.000397 0.0007950.000851 0.0003410.000397 0.0007950.000851
0.0003980.000454 0.0008520.000907 0.0003980.000454 0.0008520.000907
c d
Figure 20: Burn probability maps under various simulation scenarios. a Using full weather stream. b
Using one spread event day. c Using full weather stream and one spread event day. d Using full weather
stream and one spread event day with ignition grid from GAM model.
be serving as fuel breaks; this kind of error could explain the discrepancy in the fire size
distributions.
Our final model indicates a relatively low burn probability across the region.
Depending on the ignition grid used, the higher risk areas are either in the north uniform
grid or to the east and southeast historic ignition grid. The latter indicates a somewhat
more serious risk for the more heavily populated area, and may be the more realistic
scenario, since there is no reason to believe that the pattern of human-caused ignitions which
are largely responsible for the fires here will change in the future without some form of
intervention.
However, there are several limitations to this approach. The results obtained in this
assessment need to be interpreted with some care. We should note that by using the fire size
distribution as our standard for accuracy checking, we are assuming that the fuel distribution
and composition is similar to how it was in the past and that the climate has not changed
substantially. Future research in which this model is run under various realistic climate
change scenarios will be important. Any changes in fire management in the area, either past
or future, have not been factored into our estimates of fire risk.
The Prometheus Fire Growth Model has been used extensively in Canadian fire
management operations. It has been found to be reasonably accurate under many
circumstances, especially under peak burning conditions, which is where the FBP system
is at its most accurate. Indeed, predictions from the FBP System, which forms the foundation
of fire growth in the Prometheus model and consequently in Burn-P3, are used as a regular
and important part of forecasting the active growth of fires and the consequent planning
of fire suppression resource needs. While the FBP has limitations see discussion below, it
constitutes the best available system for predicting fire ignition and growth in the forests of
Canada. Consequently, the Prometheus fire growth model as well as Burn-P3 have been used
in a number of research studies in a wide variety of locations in Canada 5, 1316.
However, the size of fires may be overstated under moderate weather conditions
12. Since Prometheus is based on the FBP system which is, in turn, based on empirical
observations, the process under which these empirical observations were collected influences
model performance. Some of these observations were from controlled burns, so spread rates
of wildfires in dierent fuel types may be quite dierent, at least during the acceleration
phase. The reason for this is that the prescribed fires were started with a line ignition
under somewhat variable weather conditions. Because of the line ignition, the estimated
spread rate may be biased upwards, since point fire ignitions are more common in naturally
occurring fires. Spread rates for mixed wood fuel types were not empirically developed from
observed fire behaviour; instead, they were calculated as weighted averages of spread rates
of coniferous and deciduous fuel types.
The Burn-P3 simulation model is also limited in that it is not programmed to handle
vectorized fuel breaks so any fuel breaks smaller than the chosen cell resolution do not
prevent a fire from spreading. Furthermore, inputs for Burn-P3 are based on empirical
observations which makes an assumption that what will be observed in future fire seasons is
similar to what has happened in the past.
Acknowledgments
Funding was provided by the Institute for Catastrophic Loss Reduction, the MITACS
Accelerate internship program, and GEOIDE. The authors would also like to thank Marc
Parisien for assistance in the use of the Burn-P3 simulation model, Jen Beverly and Cordy
Tymstra for a number of helpful discussions, and the Ontario Ministry of Natural Resources
for providing the fuel map of the area as well as the fire and weather data. Comments from
two anonymous referees are also gratefully acknowledged.
References
1 S. McCarey, Thinking of wildfire as a natural hazard, Society and Natural Resources, vol. 17, no. 6,
pp. 509516, 2004.
2 B. J. Stocks, J. A. Mason, J. B. Todd et al., Large forest fires in Canada, 19591997, Journal of
Geophysical Research D, vol. 108, no. 1, pp. 5.15.12, 2003.
3 D. G. Woolford and W. J. Braun, Convergent data sharpening for the identification and tracking of
spatial temporal centers of lightning activity, Environmetrics, vol. 18, no. 5, pp. 461479, 2007.
4 C. Tymstra, R. W. Bryce, B. M. Wotton, and O. B. Armitage, Development and structure of
Prometheus: the Canadian wildland fire growth simulation model, Information Report NOR-X-
417, Natural Resources Canada, Canadian Forestry Service, Northern Forestry Centre, Edmonton,
Canada, 2009.
5 M. A. Parisien, V. G. Kafka, K. G. Hirsch, J. B. Todd, S. G. Lavoie, and P. D. Maczek, Using the Burn-P3
simulation model to map wildfire susceptibility, Information Report NOR-X-405, Natural Resources
Canada, Canadian Forest Service, Northern Forestry Centre, Edmonton, Canada, 2005.
6 D. R. Brillinger, H. K. Preisler, and J. W. Benoit, Risk assessment: a forest fire example, in Statistics
and Science: A Festschrift for Terry Speed, D. R. Goldstein, Ed., vol. 40 of IMS Lecture Notes Monograph
Series, pp. 177196, Institute of Mathematical Statistics, Beachwood, Ohio, USA, 2003.
7 Natural Resources Canada, Canadian Forest Fire Danger Rating System, NRC, August 2009,
http://fire.nofc.cfs.nrcan.gc.ca/en/background/bi FDR summary e.php .
8 C. E. van Wagner, Development and structure of the Canadian forest fire weather index system,
Forest Technical Report 35, Canadian Forest Service, Ottawa, Canada, 1987.
9 B. M. Wotton, Interpreting and using outputs from the Canadian Forest Fire Danger Rating System
in research applications, Environmental and Ecological Statistics, vol. 16, no. 2, pp. 107131, 2009.
10 Forestry Canada Fire Danger Group, Development and structure of the Canadian forest fire
behavior prediction system, Information Report ST-X-3, Forestry Canada, Science and Sustainable
Development Directorate, Ottawa, Canada, 1992.
11 FARSITE, 2008, http://www.firemodels.org/.
12 J. J. Podur, Weather, forest vegetation, and fire suppression influences on area burned by forest fires in Ontario,
Ph.D. dissertation, Graduate Department of Forestry, University of Toronto, Toronto, Canada, 2006.
13 C. Tymstra, M. D. Flannigan, O. B. Armitage, and K. Logan, Impact of climate change on area burned
in Albertas boreal forest, International Journal of Wildland Fire, vol. 16, no. 2, pp. 153160, 2007.
14 R. Suing, A. Grant, and R. Feick, Modeling prescribed burns to serve as regional firebreaks to allow
wildfire activity in protected areas, Forest Ecology and Management, vol. 256, no. 11, pp. 18151824,
2008.
15 M. A. Parisien, C. Miller, A. Ager, and M. Finney, Use of artificial landscapes to isolate controls on
burn probability, Landscape Ecology, vol. 25, no. 1, pp. 7993, 2010.
16 J. L. Beverly, E. P. K. Herd, and J. C. R. Conner, Modeling fire susceptibility in west central Alberta,
Canada, Forest Ecology and Management, vol. 258, no. 7, pp. 14651478, 2009.
doi:10.1155/2010/838240
Research Article
The Foundations of Probability with Black Swans
Graciela Chichilnisky
Departments of Economics and Mathematical Statistics, Columbia University, 335 Riverside Drive, New
York, NY 10027, USA
Correspondence should be addressed to Graciela Chichilnisky, [email protected]
Received 8 September 2009; Accepted 25 November 2009
Copyright q 2010 Graciela Chichilnisky. This is an open access article distributed under the
We extend the foundation of probability in samples with rare events that are potentially
catastrophic, called black swans, such as natural hazards, market crashes, catastrophic climate
change, and species extinction. Such events are generally treated as outliers and disregarded.
We propose a new axiomatization of probability requiring equal treatment in the measurement
of rare and frequent eventsthe Swan Axiomand characterize the subjective probabilities that
the axioms imply: these are neither finitely additive nor countably additive but a combination of
both. They exclude countably additive probabilities as in De Groot 1970 and Arrow 1971 and
are a strict subset of Savage 1954 probabilities that are finitely additive measures. Our subjective
probabilities are standard distributions when the sample has no black swans. The finitely additive
part assigns however more weight to rare events than do standard distributions and in that
sense explains the persistent observation of power laws and heavy tails that eludes classic
theory. The axioms extend earlier work by Chichilnisky 1996, 2000, 2002, 2009 to encompass the
foundation of subjective probability and axiomatic treatments of subjective probability by Villegas
1964, De Groot 1963, Dubins and Savage 1965, Dubins 1975 Purves and Sudderth 1976
and of choice under uncertainty by Arrow 1971.
1. Introduction
Black swans are rare events with important consequences, such as market crashes, natural
hazards, global warming, and major episodes of extinction. This article is about the
foundations of probability when catastrophic events are at stake. It provides a new axiomatic
foundation for probability requiring sensitivity both to rare and frequent events. The
study culminates in Theorem 6.1, that proves existence and representation of a probability
satisfying three axioms. The last of these axioms requires sensitivity to rare events, a property
that is desirable but not respected by standard probabilities. The article shows the connection
between those axioms and the Axiom of Choice at the foundation of Mathematics. It defines
a new type of probabilities that coincide with standard distributions when the sample is
populated only by relatively frequent events. Generally, however, they are a mixture of
countable and finitely additive measures, assigning more weight to black swans than do
normal distributions, and predicting more realistically the incidence of outliers, power
laws, and heavy tails 1, 2.
The article refines and extends the formulation of probability in an uncertain world.
It provides an argument, and formalization, that probabilities must be additive functionals
on L U where U is a -field of events represented by their indicator bounded and real
valued functions, that are neither countably additive nor finitely additive. The contribution
is to provide an axiomatization showing that subjective probabilities must lie in the full space
L rather than L1 as the usual formalization Arrow, 3 forcing countable additivity implies.
The new axioms refine both Savages 4 axiomatization of finitely additive measures, and
Villegas 5 and Arrows 3 that are based on countably additive measures, and extend both
to deal more realistically with catastrophic events.
Savage 4 axiomatized subjective probabilities as finitely additive measures repre-
senting the decision makers beliefs, an approach that can ignore frequent events as shown
in the appendix. To overcome this, Villegas 5 and Arrow 3 introduced an additional
continuity axiom called Monotone Continuity that yields countably additivity of the
measures. However Monotone Continuity has unusual implications when the subject is
confronted with rare events, for example, it predicts that in exchange for a couple of
cents, one should be willing to accept a small risk of death measured by a countably
additive probability, a possibility that Arrow called outrageous 3, Pages 4849. This
article defines a realistic solution: for some, very large, payos and in certain situations,
one may be willing to accept a small risk of deathbut not in others. This means that
Monotone Continuity holds in some cases but not in others, a possibility that leads to the
axiomatization proposed in this article and is consistent with the experimental observations
reported by Chanel and Chichilnisky 6, 7. The results are as follows. We show that
countably additive measures are insensitive to black swans: they assign negligible weight to
rare events, no matter how important these may be, treating catastrophes as outliers. Finitely
additive measures, on the other hand, may assign no weight to frequent events, which is
equally troubling. Our new axiomatization balances the two approaches and extends both,
requiring sensitivity in the measurement of rare as well as frequent events. We provide an
existence theorem for probabilities that satisfy our axioms, and a characterization of all that
do.
The results are based on an axiomatic approach to choice under uncertainty and
sustainable development introduced by Chichilnisky 810 and illuminate the classic issue
of continuity that has always been at the core of subjective probability axioms Villegas,
5, Arrow 3. To define continuity, we use a topology that tallies with the experimental
evidence of how people react to rare events that cause fear Le Doux 11, Chichilnisky 12,
previously used by Debreu 13 to formalize a markets Invisible Hand, and by Chichilnisky
9, 12, 14 to axiomatize choice under uncertainty with rare events that inspire fear. The new
results provided here show that the standard axiom of decision theory, Monotone Continuity,
is equivalent to De Groots Axiom SP4 that lies at the foundation of classic likelihood theory
Proposition 2.1 and that both of these axioms underestimate rare events no matter how
catastrophic they may be. We introduce here a new Swan Axiom Section 3 that logically
negates them both, show it is a combination of two axioms defined by Chichilnisky 9, 14
and prove that any subjective probability satisfying the Swan Axiom is neither countably
additive nor finitely additive: it has elements of both Theorem 4.1. Theorem 6.1 provides
a complete characterization of all subjective probabilities that satisfy linearity and the Swan
Axiom, thus extending earlier results of Chichilnisky 1, 2, 9, 12, 14.
There are other approaches to subjective probability such as Choquet Expected Utility
Model CEU, Schmeidler, 15 and Prospect Theory Kahneman and Tversky, 16, 17. They
use a nonlinear treatment of probabilities of likelihoods see, e.g., Dreze, 18, or Bernstein,
19, while we retain linear probabilities. Both have a tendency to give higher weight to
small probabilities, and are theoretical answers to experimental paradoxes found by Allais in
1953 and Ellsberg in 1961, among others refuting the Independence Axiom of the Subjective
Expected Utility SEU model. Our work focuses instead directly on the foundations of
probability by taking the logical negation of the Monotone Continuity Axiom. It is striking
that weakening or rejecting this axiomrespectively, in decision theory and in probability
theoryends up in probability models that are more in tune with observed attitudes when
facing catastrophic events. Presumably each approach has advantages and shortcomings. It
seems that the approach oered here may be superior on four counts: i it retains linearity
of probabilities, ii it identifies Monotone Continuity as the reason for underestimating
the measurement of catastrophic events, an axiom that depends on a technical definition of
continuity and has no other compelling feature, iii it seems easier to explain and to grasp,
and therefore iv it may be easier to use in applications.
2. The Mathematics of Uncertainty

Uncertainty
Uncertainty is described by a set of distinctive and exhaustive possible events represented

by a family of sets {U }, N, whose union describes a universe U U . An event
U U is identified with its characteristic function U : U R where U x 1 when
x U and U x 0 when x / U. The subjective probability of an event U is a real number
WU that measures how likely it is to occur according to the subject. Generally we assume
that the probability of the universe is 1 and that of the empty set is zero W 0. In
this article we make no dierence between subjective probabilities and likelihoods, using
both terms intercheangeably. Classic axioms for subjective probability resp. likelihoods
are provided by Savage 4 and De Groot 20. The likelihood of two disjoint events is
the sum of their likelihoods: WU1 U2 WU1 WU2 when U1 U2 ; a
property called additivity. These properties correspond to the definition of a probability
or likelihood as a finite additive measure on a family -algebra of measurable sets of U,
which is Savages 4 definition of subjective probability. W is countably additive when

W i1 Ui i1 WUi whenever Ui Uj if i
/ j. A purely finitely additive probability is one
that is additive but not countably additive. Savages subjective probabilities can be purely
finitely additive or countably additive. In that sense they include all the probabilities in this
article. However as seen below, this article excludes probabilities that are either purely finitely
additive, or countably additive, and therefore our characterization of a subjective probability
is strictly finer than that of Savages 4, and dierent from the view of a measure as a
countably additive set function e.g. De Groot , 21 The following Axioms were introduced
by Villegas 5; and others for the purpose of obtaining countable additivity.
Monotone Continuity Axiom (MC) (Arrow [3])

For every two events f and g with Wf > Wg, and every vanishing sequence of events

{E }1,2... defined as follows: for all , E1 E and 1 E there exists N such that
altering arbitrarily the events f and g on the set Ei , where i > N, does not alter the subjective
probability ranking of the events, namely, Wf

> Wg
, where f
and g
are the altered
events.
This axiom is equivalent to requiring that the probability of the sets along a vanishing
sequence goes to zero. Observe that the decreasing sequence could consist of infinite intervals
of the form n, for n 1, 2 . . . . Monotone continuity therefore implies that the likelihood
of this sequence of events goes to zero, even though all its sets are unbounded. A similar
example can be constructed with a decreasing sequence of bounded sets, 1/n, 1/n for n
1, 2 . . . , which is also a vanishing sequence as it is decreasing and their intersection is empty.
De Groots Axiom SP4 (De Groot, [20], Chapter 6, page 71)

If A1 A2 is a decreasing sequence of events and B is some fixed event that is less likely

than Ai for all i, then the probability of the intersection i Ai is larger than that of B.
The following proposition establishes that the two axioms presented above are one
and the same; both imply countable additivity.
Proposition 2.1. A relative likelihood (subjective probability) satisfies the Monotone Continuity
Axiom if and only if it satisfies Axiom SP4 . Each of the two axioms implies countable additivity.
Proof. Assume that De Groots axiom SP4 is satisfied. When the intersection of a decreasing

sequence of events is empty i Ai and the set B is less likely to occur than every
set Ai , then the subset B must be as likely as the empty set; namely, its probability must
be zero. In other words, if B is more likely than the empty set, then regardless of how
small is the set B, it is impossible for every set Ai to be as likely as B. Equivalently, the
probability of the sets that are far away in the vanishing sequence must go to zero. Therefore
SP4 implies Monotone Continuity. Reciprocally, assume that MC is satisfied. Consider a
decreasing sequence of events Ai and define a new sequence by substracting from each set

the intersection of the family, namely, A1 i Ai , A 2 i Ai , . . . . Let B be a set that is
more likely than the empty set but less likely than every Ai . Observe that the intersection

of the new sequence is empty, i Ai i Ai and since Ai Ai1the new sequence
is, by definition, a vanishing sequence. Therefore by MC limi WAi i Ai 0. Since

WB > 0, B must be more likely than Ai A for some i onwards. Furthermore,
i i
Ai Ai A A and A A
A , so that WA i > WB is
i i
i i i i i i i

equivalent to WAi i Ai W i Ai > WB. Observe that W i Ai < WB would

contradict the inequality WAi WAi i Ai W i Ai > WB, since as we saw

above, by MC, limi WAi i Ai 0, and WAi i Ai W i Ai > WB. It follows

that W i A i > WB, which establishes De Grootss Axiom SP 4 . Therefore Monotone
Continuity is equivalent to De Groots Axiom SP4 . A proof that each of the axioms implies
countable additivity is in Villegas 5, Arrow 3 and De Groot 20.
The next section shows that the two axioms, Monotone Continuity and SP4 , are biased
against rare events no matter how catastrophic these may be.
3. The Value of Life

The best way to explain the role of Monotone Continuity is by means of an example provided
by Arrow 3, Pages 4849. He explains that if a is an action that involves receiving one cent,
b is another that involves receiving zero cents, and c is a third action involving receiving
one cent and facing a small probability of death, then Monotone Continuity requires that the
third action involving death and one cent should be preferred to the action with zero cents
if the probability of death is small enough. Even Arrow says of his requirement this may
sound outrageous at first blush. . . Arrow 3, Pages 4849. Outrageous or not, Monotone
Continuity MC leads to neglect rare events with major consequences, like death. Death is a
black swan.
To overcome the bias we introduce an axiom that is the logical negation of MC:
this means that sometimes MC holds and others it does not. We call this the Swan Axiom,
and it is stated formally below. To illustrate this, consider an experiment where subjects
are oered a certain amount of money to choose a pill at random from a pile, which is
known to contain one pill that causes death. It was shown experimentally Chanel and
Chichilnisky 7 that in some cases people accept a sum of money and choose a pill provided
that the pile is large enoughnamely, when the probability of death is small enoughthus
satisfying the Monotone Continuity axiom and determining the statistical value of their lives.
But there are also cases where the subjects will not accept to choose any pill, no matter
how large is the pile. Some people refuse the payment of one cent if it involves a small
probability of death, no matter how small the probability may be Chanel and Chichilnisky,
6, 7. This conflicts with the Monotone Continuity axiom, as explicitly presented by Arrow
3.
Our Axiom provides a reasonable resolution to this dilemma that is realistic and
consistent with the experimental evidence. It implies that there exist catastrophic outcomes
such as the risk of death, so terrible that one is unwilling to face a small probability of death to
obtain one cent versus nothing, no matter how small the probability may be. According to our
Axiom, no probability of death may be acceptable when one cent is involved. Our Axiom also
implies that in other cases there may be a small enough probability that the lottery involving
death may be acceptable, for example if the payo is large enough to justify the small risk.
This is a possibility discussed by Arrow 3. In other words: sometimes one is willing to take
a risk with a small enough probability of a catastrophe, in other cases one is not. This is the
content of our Axiom, which is formally stated as follows.
The Swan Axiom

This axiom is the logical negation of Monotone Continuity: There exist events f and g with
Wf > Wg, and for every vanishing sequence of events {Ei }i1,2... an N > 0 such that
altering arbitrarily the events f and g on the set Ei , where i > N, does not alter the probability
ranking of the events, namely, Wf
> Wg
, where f
and g
are the altered events. For
other events f and g with Wf > Wg, there exist vanishing sequence of events {Ei }i1,2...
where for every N, altering arbitrarily the events f and g on the set Ei , where i > N, does
alter the probability ranking of the events, namely Wf
< Wg
, where f
and g
are the
altered events.
Definition 3.1. A probability W is said to be biased against rare events or insensitive to rare events
when it neglects events that are small according to Villegas and Arrow; as stated in Arrow
3, page 48: An event that is far out on a vanishing sequence is small by any reasonable
standards Arrow 3, page 48. Formally, a probability is insensitive to rare events when
given two events f and g and any vanishing sequence of events {Ej }, N Nf, g > 0,
such that Wf > Wg Wf
> Wg
for all f
, g
satisfying f
f and g
g a.e. on
Ejc R when j > N, and Ec denotes the complement of the set E.
Proposition 3.2. A subjective probability satisfies Monotone Continuity if and only if it is biased
against rare events.
Proof. This is immediate from the definitions of both 3, 12.
Corollary 3.3. Countably additive probabilities are biased against rare events.
Proof. It follows from Propositions 2.1 and 3.2 9, 12.
Proposition 3.4. Purely finitely additive probabilities are biased against frequent events.
Proof. See example in the appendix.
Proposition 3.5. A subjective probability that satisfies the Swan Axiom is neither biased against rare
events, nor biased against frequent events.
Proof. This is immediate from the definition.
4. An Axiomatic Approach to Probability with Black Swans

This section proposes an axiomatic foundation for subjective probability that is unbiased
against rare and frequent events. The axioms are as follows:
Axiom 1. Subjective probabilities are continuous and additive.
Axiom 2. Subjective probabilities are unbiased against rare events.
Axiom 3. Subjective probabilities are unbiased against frequent events.
Additivity is a natural condition and continuity captures the notion that nearby
events are thought as being similarly likely to occur; this property is important to ensure
that sucient statistics exist. Nearby has been defined by Villegas 5 and Arrow 3 as
follows: two events are close or nearby when they dier on a small set as defined in Arrow
3, see previous section. We saw in Proposition 3.2 that the notion of continuity defined by
Villegas and Arrownamely, monotone continuityconflicts with the Swan Axiom. Indeed
Proposition 3.2 shows that countably additive measures are biased against rare events. On
the other hand, Proposition 3.4 and the Example in the appendix show that purely finitely
additive measures can be biased against frequent events. A natural question is whether there
is anything left after one eliminates both biases. The following proposition addresses this
issue.
Theorem 4.1. A subjective probability that satisfies the Swan Axiom is neither finitely additive nor
countably additive; it is a strict convex combination of both.
Proof. This follows from Propositions 3.2, 3.4 and 3.5, Corollary 3.3 above, and the fact that
convex combinations of measures are measures. It extends Theorem 6.1 of Section 6 below,
which applies to the special case where the events are Borel sets in R or in an interval a, b
R.
Theorem 4.1 establishes that neither Savages approach nor Villegas and Arrows
satisfy the three axioms stated above. These three axioms require more than the additive
subjective probabilities of Savage, since purely finitely additive probabilities are finitely
additive and yet they are excluded here. At the same time the axioms require less than the
countably subjective additivity of Villegas and Arrow, since countably additive probabilities
are biased against rare events. Theorem 4.1 above shows that a strict combination of both
does the job.
Theorem 4.1 does not however prove the existence of likelihoods that satisfy all three
axioms. What is missing is an appropriate definition of continuity that does not conflict with
the Swan Axiom. The following section shows that this can be achieved by identifying an
event with its characteristic function, so that events are contained in the space of bounded
real-valued functions on the universe space U, L U, and endowing this space with the sup
norm.
5. Axioms for Probability with Black Swans, in R or a, b

From here on events are the Borel sets of the real line R or the interval a, b. This is a
widely used case that make the results concrete and allows to compare the results with the
earlier axioms on choice under uncertainty of Chichilnisky 9, 12, 14. We use a concept of
continuity based on a topology that was used earlier by Debreu 13 and by Chichilnisky
1, 2, 9, 10, 12, 14: observable events are in the space of measurable and essentially bounded
functions L L R with the sup norm f ess supxR |fx|. This is a sharper and more
stringent definition of closeness than the one used by Villegas and Arrow, since two events
can be close under the Villegas-Arrow definition but not under ours, see the appendix.
A subjective probabiliy satisfying the classic axioms by De Groot 20 is called a
standard probability, and is countably additive. A classic result is that for any event f L
a standard probability has the form Wf R fx xd, where L1 R is an integrable
function in R.
The next step is to introduce the new axioms, show existence and characterize all the
distributions that satisfy the axioms. We need more definitions. A subjective probability W :
L R is called biased against rare events, or insensitive to rare events when it neglects events
that are small according to a probability measure on R that is absolutely continuous with
respect to the Lebesgue measure. Formally, a probability is insensitive to rare events when
given two events f and g f, g > 0, such that Wf > Wg Wf
> Wg
for
all f
, g
satisfying f
f and g
g a.e. on A R and Ac < . Here Ac denotes the
complement of the set A. W : L R is said to be insensitive to frequent events when given
any two events f, g f, g > 0 that Wf > Wg Wf
> Wg
for all f
, g
satisfying
f
f and g
g a.e. on A R and Ac > 1 . W is called sensitive to rare respectively
frequent events when it is not insensitive to rare respectively frequent events.
The following three axioms are identical to the axioms in last section, specialized to
the case at hand.
Axiom 1. W : L R is linear and continuous.
Axiom 2. W : L R is sensitive to frequent events.
Axiom 3. W : L R is sensitive to rare events.

The first and the second axiom agree with classic theory and standard likelihoods
satisfy them. The third axiom is new.
Lemma 5.1. A standard probability satisfies Axioms 1 and 2, but it is biased against rare events and
therefore does not satisfy Axiom 3.

Proof. Consider Wf R fxxdx, R xdx K < . Then

W f W g fxxdx gxxdx
R R
5.1

fx gx xdx W f g ,
R
since f and g are characteristic functions and thus positive. Therefore W is linear. W is
continuous with respect to the L1 norm f1 R |fx|xd because f < implies

W f fx xdx fx xdx xdx K. 5.2
R R
Since the sup norm is finer than the L1 norm, continuity in L1 implies continuity with
respect to the sup norm Dunford and Schwartz, 22. Thus a standard subjective probability
satisfies Axiom 1. It is obvious that for every two events f, g, with Wf > Wg, the
inequality is reversed namely Wg
> Wf
when f
and g
are appropriate variations of
f and g that dier from f and g on sets of suciently large Lebesgue measure. Therefore
Axiom 2 is satisfied. A standard subjective probability is however not sensitive to rare events,
as shown in Chichilnisky 1, 2, 9, 10, 12, 14, 23.
6. Existence and Representation

Theorem 6.1. There exists a subjective probability W : L R satisfying Axioms 1, 2, and 3. A
probability satisfies Axioms 1, 2 and 3 if and only if there exist two continuous linear functions on L ,
denoted 1 and 2 and a real number , 0 < < 1, such that for any observable event f L

W f fx1 xdx 1 2 f 6.1
xR
where 1 L1 R, defines a countably additive measure on R and 2 is a purely finitely additive
measure.
Proof. This result follows from the representation theorem by Chichilnisky 9, 12.
Example 6.2 Heavy Tails. The following illustrates the additional weight that the new
axioms assign to rare events; in this example in a form suggesting heavy tails. The finitely
additive measure 2 appearing in the second term in 6.1 can be illustrated as follows. On
the subspace of events with limiting values at infinity, L
{fL : limx x < },
define 2 f limx fx and extend this to a function on all of L using Hahn Banachs
theorem. The dierence between a standard probability and the likelihood defined in 6.1
is the second term 2 , which focuses all the weight at infinity. This can be interpreted as a
heavy tail, a part of the distribution that is not part of the standard density function 1 and
gives more weight to the sets that contain terminal events, namely sets of the form x, .
Corollary 6.3. In samples without rare events, a subjective probability that satisfies Axioms 1, 2, and
3 is consistent with classic axioms and yields a countably additive measure.
Proof. Axiom 3 is an empty requirement when there are no rare events while, as shown above,
Axioms 1 and 2 are consistent with standard relative likelihood.
7. The Axiom of Choice

There is a connection between the new axioms presented here and the Axiom of Choice
that is at the foundation of mathematics Godel, 24, which postulates that there exists a
universal and consistent fashion to select an element from every set. The best way to describe
the situation is by means of an example, see also Dunford and Schwartz 22, Yosida 25, 26,
Chichilnisky and Heal 27, and Kadane and OHagan 28.
Example 7.1 illustration of a purely finitely additive measure. Consider a possible measure
satisfying the following: for every interval A R, A 1 if A {x : x > a, for some
a R}, and otherwise A 0. Such a measure would not be countably additive, because

the family of countably many disjoint sets {Vi }i0,1,... defined as Vi i, i 1 i 1, i,

satisfies Vi Vi when i / j, and i0 Vi i0 i, i1 i1, i R, so that
i0 Vi 1,

while i0 Vi 0, which contradicts countable additivity. Since the contradiction arises
from assuming that is countably additive, such a measure could only be purely finitely
additive.
One can illustrate a function on L that represents a purely finitely additive measure
if we restrict our attention to the closed subspace L
of L consisting of those functions fx
in L that have a limit when x , by the formula f limx fx, as in Example 6.2
of the previous section. The function can be illustrated as a limit of a sequence of delta
functions whose supports increase without bound. The problem however is to extend the
function to another defined on the entire space L . This could be achieved in various ways
but as we will see, each of them requires the Axiom of Choice.
One can use HahnBanachs theorem to extend the function from the closed
subspace L
L to the entire space L preserving its norm. However, in its general
form HahnBanachs theorem requires the Axiom of Choice Dunford and Schwartz, 22.
Alternatively, one can extend the notion of a limit to encompass all functions in L including
those with no standard limit. This can be achieved by using the notion of convergence along
a free ultrafilter arising from compactifying the real line R as by Chichilnisky and Heal 27.
However the existence of a free ultrafilter also requires the Axiom of Choice.
This illustrates why any attempts to construct purely finitely additive measures, requires
using the Axiom of Choice. Since our criteria include purely finitely additive measures, this
provides a connection between the Axiom of Choice and our axioms for relative likelihood.
It is somewhat surprising that the consideration of rare events that are neglected in standard
statistical theory conjures up the Axiom of Choice, which is independent from the rest of
mathematics Godel, 24.
Appendix
Example A.1 Illustration of a probability that is biased against frequent events. Consider the
function Wf lim infxR fx. This is insensitive to frequent events of arbitrarily large
Lebesgue measure Dunford and Schwartz, 22 and therefore does not satisfy Axiom 2. In
addition it is not linear, failing Axiom 1.
Example A.2 two approaches to closeness. Consider the family {Ei } where Ei i, ,

i 1, 2, . . . . This is a vanishing family because for all i, Ei Ei1 and i1 Ei . Consider
now the events f i t K when t Ei and f i t 0 otherwise, and g i t 2K when t Ei
and g i t 0 otherwise. Then for all i, supEi |f i t g i t| K. In the sup norm topology this
implies that f i and g i are not close to each other, as the dierence f i g i does not converge
to zero. No matter how far along the vanishing sequence Ei the two events f i , g i dier by
K. Yet since the events f i , g i dier from f 0 and g 0 respectively only in the set Ei , and
{Ei } is a vanishing sequence, for large enough i they are as close as desired according to
Villegas-Arrows definition of nearby events.
The Dual Space L : Countably Additive and Finitely Additive Measures

The space of continuous linear functions on L with the sup norm is the dual of L , and
is denoted L . It has been characterized, for example, in Yosida 25, 26. L consists of the
sum of two subspaces
i L1 functions
g that define countably additive measures on R by
the rule A A gxdx where R |gx|dx < so that is absolutely continuous with respect
to the Lebesgue measure, and ii a subspace consisting of purely finitely additive measures.
A countable measure can be identified with an L1 function, called its density, but purely
finitely additive measures cannot be identified by such functions.
Example A.3. Illustration of a Finitely Additive Measure that is not Countably Additive
See Example 7.1 in Section 7.
Acknowledgments
This research was conducted at Columbia Universitys Program on Information and
Resources and its Columbia Consortium for Risk Management CCRM. The author
acknowledges support from Grant no 5222-72 of the US Air Force Oce of Research directed
by Professor Jun Zheng, Arlington VA. The initial results Chichilnisky 8 were presented
as invited addresses at Stanford Universitys 1993 Seminar on Reconsideration of Values,
organized by Professor Kenneth Arrow, at a 1996 Workshop on Catastrophic Risks organized
at the Fields Institute for Mathematical Sciences of the University of Toronto, Canada, at
the NBER Conference Mathematical Economics: The Legacy of Gerard Debreu at UC Berkeley,
October 21, 2005, the Department of Economics of the University of Kansas National Bureau
of Economic Research General Equilibrium Conference, September 2006, at the Departments
of Statistics of the University of Oslo, Norway, Fall 2007, at a seminar organized by the late
Professor Chris Heyde at the Department of Statistics of Columbia University, Fall 2007,
at seminars organized by Drs. Charles Figuieres and Mabel Tidball at LAMETA Universite
de Montpellier, France December 19 and 20, 2008, and by and at a seminar organized
by Professor Alan Kirman at GREQAM Universite de Marseille, December 18 2008. We
are grateful to the above institutions and individuals for supporting the research, and for
helpful comments and suggestions. An anonymous referee provided insightful comments

that improved this article.
References
1 G. Chichilnisky, The Limits of Econometrics: Non Parametric Estimation in Hilbert Spaces,
Econometric Theory, vol. 25, no. 4, pp. 10701086, 2009.
2 G. Chichilnisky, The Work and Legacy of Kenneth Arrow, in Encyclopedia of Quantitative Finance, R.
Cont, Ed., pp. 7682, John Wiley & Sons, Chichester, 2010.
3 K. J. Arrow, Essays in the Theory of Risk-Bearing, North-Holland, Amsterdam, The Netherlands, 1970.
4 L. J. Savage, The Foundations of Statistics, Dover, New York, USA, revised edition, 1972.
5 C. Villegas, On quantitative probability algebras, The Annals of Mathematical Statistics, vol. 35,
pp. 17891800, 1964.
6 O. Chanel and G. Chichilnisky, The influence of fear in decisions: experimental evidence, Journal of
Risk and Uncertainty, vol. 39, no. 3, pp. 271298, 2009.
7 O. Chanel and G. Chichilnisky, The value of life: theory and experiments, Working Paper, GREQE,
Universite de Marseille and Columbia University, New York, NY, USA, 2009.
8 G. Chichilnisky, An axiomatic approach to sustainable development, Social Choice and Welfare, vol.
13, no. 2, pp. 231257, 1996.
9 G. Chichilnisky, An axiomatic approach to choice under uncertainty with catastrophic risks,
Resource and Energy Economics, vol. 22, no. 3, pp. 221231, 2000.
10 G. Chichilnisky, Updating von Neumann Morgenstern axioms for choice under uncertainty, in
Proceedings of a Conference on Catastrophic Risks, the Fields Institute for Mathematical Sciences, Toronto,
Canada, 1996.
11 J. Le Doux, The Emotional Brain, Simon and Schuster, New York, NY, USA, 1996.
12 G. Chichilnisky, The topology of fear, Journal of Mathematical Economics, vol. 45, no. 12, pp. 807816,
2009.
13 G. Debreu, Valuation equilibrium and Pareto optimum, Proceedings of the National Academy of
Sciences of the United States of America, vol. 40, pp. 588592, 1954.
14 G. Chichilnisky, Catastrophic risk, in Encyclopedia of Environmetrics, A. H. El-Shaarawi and W. W.
Piegorsch, Eds., vol. 1, pp. 274279, John Wiley & Sons, Chichester, UK, 2002.
15 D. Schmeidler, Subjective probability and expected utility without additivity, Econometrica, vol. 57,
no. 3, pp. 571587, 1989.
16 D. Kahneman and A. Tversky, Prospect theory: an analysis of decisions under risk, Econometrica,
vol. 47, pp. 263291, 1979.
17 A. Tversky and D. Kahneman, Advances in prospect theory: cumulative representation of
uncertainty, Journal of Risk and Uncertainty, vol. 5, no. 4, pp. 297323, 1992.
18 J. H. Dreze, Essays on Economic Decisions under Uncertainty, Cambridge University Press, Cambridge,
UK, 1987.
19 P. L. Bernstein, Against the Gods: A Remarkable Story of Risk, John Wiley & Sons, New York, NY, USA,
1996.
20 M. H. DeGroot, Optimal Statistical Decisions, McGraw-Hill, New York, NY USA, 1970.
21 M. H. DeGroot, Optimal Statistical Decisions, Wiley Classics Library, John Wiley & Sons, Hoboken, NJ,
USA, 2004.
22 N. Dunford and J. T. Schwartz, Linear Operators, Part I, Interscience, New York, NY, USA, 1958.
23 G. Chichilnisky and H.-M. Wu, General equilibrium with endogenous uncertainty and default,
Journal of Mathematical Economics, vol. 42, no. 4-5, pp. 499524, 2006.
24 K. Godel, The Consistency of the Continuum Hypothesis, Annals of Mathematics Studies, No 3, Princeton
University Press, Princeton, NJ, USA, 1940.
25 K. Yosida and E. Hewitt, Finitely additive measures, Transactions of the American Mathematical
Society, vol. 72, pp. 4666, 1952.
26 K. Yosida, Functional Analysis, vol. 12 of Die Grundlehren der mathematischen Wissenschaften, Springer,
New York, NY, USA, 4th edition, 1974.
27 G. Chichilnisky and G. Heal, Social choice with infinite populations: construction of a rule and
impossibility results, Social Choice and Welfare, vol. 14, no. 2, pp. 303318, 1997.
28 J. B. Kadane and A. OHagan, Using finitely additive probability: uniform distributions on the
natural numbers, Journal of the American Statistical Association, vol. 90, no. 430, pp. 626631, 1995.
doi:10.1155/2010/954750
Research Article
Asteroids: Assessing Catastrophic Risks
Graciela Chichilnisky and Peter Eisenberger

Columbia University, New York, NY 10027, USA
Correspondence should be addressed to Graciela Chichilnisky, [email protected]
Received 28 January 2010; Accepted 9 May 2010
Copyright q 2010 G. Chichilnisky and P. Eisenberger. This is an open access article distributed
under the Creative Commons Attribution License, which permits unrestricted use, distribution,
and reproduction in any medium, provided the original work is properly cited.
We evaluate two risk profiles: i global warming risks and ii collisions with asteroids that can
cause the extinction of our species. The expected values computed for these two risks suggest that
no action will be taken to avoid extinction. The result is somewhat counterintuitive, but it is typical
of the results of using classic decision theory to evaluate catastrophic risks in the distant future, see
the study by Posner 2004. We establish why expected value is insensitive to catastrophic risks see
the study by Chichilnisky 1996, and use another criterion to evaluate risk based on axioms for
choice under uncertainty that update the classic Von Neumann theory and require equal treatment
for rare and frequent events. Optimizing according to the new criterion is shown to be equivalent to
optimizing expected utility with a restriction on the worst outcome in the case of a catastrophe. The
evaluation obtained from the new criterion seems more intuitively plausible, and suggests a more
practical and realistic approach to catastrophic risks: optimizing expected value while minimizing
losses in the case of a catastrophe.
1. Asteroids
Sixty five million years ago, an asteroid crashed into earth. Global winds distributed the dust
throughout the atmosphere, blocking sunlight, and many life forms that relied on the sun
eventually perished. In a short period of time, experts believe, the mighty dinosaurs that
dominated our planet went extinct. Realistically the same fate awaits us. Over 99.99% of the
species that have ever existed are now extinct 1, 2. If our species survives long enough,
we will be exposed to an asteroid and could suer the same fate as the dinosaurs. The data
suggests that asteroids of that caliber will hit our planet on average once every 100 million
years 2. The last one was 65 million years ago. Under current conditions, when the next one
hits the earth, humans and many other species could go extinct.
What should we do about this threat to our survival and others like it? And if the issue
is serious, why is this issue getting so little attention whereas the less catastrophic threat of
global warming is in the news almost daily?
The purpose of this paper is to provide answers to these questions. We examine

systematically how to deal with catastrophic risks such as asteroid impacts, which are small-
probability events with enormous consequences, events that could threaten the survival
of our species, and compare their treatment with risks like global warming that are more
imminent and familiar but possibly less catastrophic.
The task is not easy. Classic tools for risk management are notoriously poor
for managing catastrophic risks, see Posner 2 and Chichilnisky 3, 4. There is an
understandable tendency to ignore rare events, such as an asteroid impact, which are unlikely
to occur in our lifetimes or those of our families 2, 5. Yes this is a questionable instinct
at this stage of human evolution where our knowledge enables to identify such risks.
Standard decision tools make this task dicult. We show using the existing data that a major
disturbance caused by global warming of less than 1% of GDP overwhelms in expected
value the costs associated with an asteroid impact that can plausibly lead to the extinction
of the human species. We show that the expected value of the loss caused by an asteroid
that leads to extinctionis between $500 million and $92 billion. A loss of this magnitude is
smaller than that of a failure of a single atomic plantthe Russians lost more than $140 billion
with the accident at Chernobylor with the potential risks involved in global warming that
is between $890 billion and $9.7 trillion 2. Using expected values therefore we are led
to believe that preventing asteroid impacts should not rank high in our policy priorities.
Common sense rebels against the computation we just provided. The ability to anticipate
and plan for threats that have never been experienced by any current or past member of the
species and are unlikely to happen in our lifespans, appears to be unique to our species. We
need to use a risk management approach that enables us to deal more eectively with such
threats 2. To overcome this problem this paper summarizes a new axiomatic approach to
catastrophic risks that updates current methods developed initially by John Von Neumann,
see Chichilnisky 3, 4, 69, and oers practical figures to evaluate possible policies that
would protect us from asteroid impacts. Our conclusion is that we are underinvesting in
preventing the risk of asteroid like threats. Much can and should be done at a relatively small
cost; this paper suggests a methodology and a range of dollar values that should be spent to
protect against such risks to help prevent the extinction of our species.
2. Catastrophes and the Survival of the Species

A catastrophe is a rare event with enormous consequences. In a recent book, Posner 2
classifies catastrophes into various types, each of which threats the survival of our species.
He uses a classic approach to value the importance of a risk by quantifying its expected value,
namely, the product of the probability times the loss. For example, the expected value of an
event that occurs with ten percent probability and involves $1 billion loss is $109 101 $100
million. This approach is used by actuaries to price the cost of life insurance policies, and is
also by law the measure used in US Congress when evaluating budget plans with uncertain
outcomes. The notion of expected value started with Von Neumann and Morgenstern about
60 years ago 10, and it is based on their axioms or principles for decision making under
uncertainty formalized in 11, 12. Posner 2 uses the concept of expected value to evaluate
risks but warns the reader about its weaknesses for evaluating catastrophic risks see Posner
2, Chapter 3, pages 150154. This weakness is exposed in the case of asteroids, when we
ask how much we should invest in preventing the impact of an asteroid that can destroy all
of the earths economic value forever. Posner 2 argues that expected value does not capture
the true impact of such a catastrophe; that something else is at stake. Because of his loyalty
to the concept of expected value, which does not work well in these cases, Posner appears
to be arguing that rationality does not work in the case of catastrophes, that we cannot deal
rationally with small probabilities events that cause such large and irreversible damage.
Perhaps the problem is not one of rationality. There may be a dierent rationality
needed when considering the long-range future of the species. It could be that expected value
is a good measure for evaluating risks that have a good chance to occur in our lifetime, but
not for evaluating risks that are important but have essentially a zero chance to occur while
we are alive. For such risks we may need another approach overall, for both the present and
the future. In our current state of evolution it would seem useful to oppose a human tendency
based on our hunter-gatherer origins to give preference to immediate outcomes as opposed to
more distant ones; see the study by McClure et al. 5. When using expected value the response
we obtain seems to clash with our intuition because the probabilities involved are so small
that they render the computation almost meaningless, as seen numerically in the examples
provided below. The experimental evidence summarized below provides further support for
this view.
3. Experimental Evidence and Alternative Approaches

Expected utility optimization derives from Von Neumanns NM axioms, but it is well
known for sometime that it conflicts with the experimental evidence on how humans choose
under uncertainty; for example, see the studies by Allais 13, Machina, 14, 15, Tversky
and Wakker 16. Problems arise when there are infrequent events involved; examples are
weather catastrophes like hurricanes or mass extinctions. Similar types of conflicts appear
when using the standard criterion of present value optimization for choosing among projects
that evolve through time. Other problems arise when the plans involve very long time
horizons as in the disposal of nuclear waste Heal 17. While the problem areas mentioned
above are quite dierent, they all share a common mathematical root: the relative insensitivity
of the classic axioms of choice towards 1 small-probability events 2 the far away future
3, 4, 6, 17. The mathematical structure of the problem is the same in all cases: it arises
from the use of normal distributions; the bell curves to describe the frequency with
which we expect everything to occur from weather events to returns on investments or
corporate profits. Normal distributions arise when many independent events contribute to
some outcome. However when there are unexpected interconnections or catastrophic events,
normal distribution can understate 1 the role of small-probability events 2 the role of
events that are very far into the future. We formalize this problem below and provide a
solution.
Taking a leaf from Von Neumann and Morgensterns book, Chichilnisky 3, 4, 6
reconsidered the foundations of the expected value approach, which are the VNM axioms
for choice under uncertainty, see also Arrow 11 and Hernstein and Milnor 12. A first
step is to show that classic axioms can be biased against small-probability events, as was
established by Chichilnisky in 6. She introduced new axioms for choice under uncertainty
that require more symmetry in the treatment of small and large probability events 3, 4. The
new axioms coincide with those of Von Neumann and Morgenstern when the events involved
have normal probabilities, for example, when they are likely to occur in our lifetime. But
the new axioms give rise to a rather dierent decision making criterion when the probabilities
involved are extremely small, or equivalently when the events are only likely to occur in
a very long future. The two sets of axioms are consistent with each other for normal events
while they are quite dierent on catastrophic events. How can this be?
A somewhat far-fetched analogy is the relationship between classical mechanics and
general relativity. The former applies to normal scales that are closer to own reality on earth,
while the latter applies to large-scale phenomena involving astral bodies. Both are correct in
their respective spheres, and neither contradicts the other. The same could be the case with
the Von Neumann-Morgenstern and the Chichilniskys axioms. The next section presents the
new axioms. It has been shown empirically and theoretically Posner 2 and Chichilnisky
3 that standard tools of decision making under uncertainty are ill suited to evaluate such
risks, more on this below.
In sum: the standard approaches do not provide a satisfactory answer and we provide
here an alternative approach to risk management that seems better suited to the management
of catastrophic risks and risks that are most likely to occur in the very distant future. This
approach has an axiomatic treatment that parallels Von Neumanns theory of choice under
uncertainty, but extends it requiring equal treatment for frequent and rare events.
The next section provides empirical motivation for the new approach by comparing it
with expected utility in two risk profiles: asteroids and global warming risks.
4. Two Risk Profiles: Asteroids and Global Warming

In September 16, 2000, the British Task Force on Potentially Hazardous Near Earth Objects
NEOs produced a report classifying asteroids risks by their size, energy yield, and average
interval between impacts. Large-mass extinctionsfor example, the Cretaceous Terciary
Geological boundaryfollow from the impact of asteroids of 16 km in diameter, which occur
on the average once every 100 million years, and threaten the survival of all advanced life
forms on the planet of which 99.9% have already gone extinct 18. Below we compare the
risk profile presented by asteroids with global warming risks.
i An asteroid impact of this magnitude occurs on average once every 100 million
years.
ii It produces damage of about $120 trillion 2, obliterating all human-produced
value in the planet.
iii The damage is permanentit continues annually for about 1 billion years, the
expected lifetime of our planet before it is destroyed by our sun becoming a red
star.
iv Existing observations indicate that such an impact will not take place in the next 30
years.
Below we compare this risk with the risk of global warming with the following
simplified profile.
i The probability of global warming is 1, namely, it is happening.

ii The best estimate is that it will produce damage that is calculated for the
catastrophic case to bring in a permanent loss of about $2 trillion a year loss in
the US and globally about $8 trillion a year; see, for example, the study by Posner
in 2. There is no consensus on whether the gradual or the catastrophic case for
global warming is more likely.
Before examining the two risk profiles, we explain the connection between rare events
and events in the distant future.
5. Decisions over Time and under Uncertainty

We adopt a simple approach that views choices over time and choices under uncertainty
as two aspects of one and the same phenomenon.
The probability of an event is viewed as the frequency with which this event is likely to
occur through time. A word of warning is in order. This is not the only approach to defining
probabilitiesindeed many people object to it because of it views reality as an experiment
that can repeat itself. Yet for our purposes here the approach has an important advantage in
that it simplifies matters and at the same time generalizes the results. For example, drawing
heads with a fair coin is an event with probability 0.50 because in repeated flipping of the
coin, tails tend to occur 50% of the time. The appearance of heads is thus a relatively
frequent event, one that will on average occur one out of every two trials. If we flip a coin
every year, for example, heads will occur almost surely in our lifetime. In this context, high
frequency over time translates into high-probability and vice versa. Low-frequency events
translate into low probability and vice versa. This way we treat time and uncertainty. as
two aspects of the same phenomenon.
We saw that high-probability events are those that occur frequently in time. In the case
of asteroids, we know with certainty that at some point in time one will hit the earth and
destroy most life on the planet unless we take action. The probability of such destructive
event sometime in the future is one, although the event is so infrequent that the probability
is essentially zero in any persons lifetime.
A catastrophe has been defined in 2, 6 as an event with enormous negative
consequencessuch as the extinction of the speciesan event that may occur with
probability one sometime in the distant future, but has nearly zero probability of occurring
during the lifetime of any one individual. There is basically zero risk that the catastrophe
will occur in our lifetime, although we know for sure it will occur at some point in the future.
It is possible although very unlikely that we will pick up the papers tomorrow and read
that some astronomer has discovered a massive comet taking deadly aim at our planet. With
this definition, it becomes clear that dealing with catastrophes is the same as dealing with
events that take place in the distant future. It is well known that the losses from events that
take place in the very long-run future are discounted in an exponential manner to obtain
a present discounted value, and that this exponential discount renders long-term losses
almost meaningless. Having connected choices over time and choices under uncertainty,
this explains why expected value, which is the parallel to present discounted value, leads
to counterintuitive evaluation of catastrophic losses or for that matter distort outcomes
generally. Next we provide a formal framework.
6. The Mathematics of Risk

A system is in one of several possible states that can be described by real numbers. To each
state s R there is an associated outcome, so that one has xs RN , N 1. A description
of outcomes across all states is called a lottery x : R RN . The space of all lotteries
is therefore a function space L identified in the following by the space of all essentially
bounded real-valued functions with the sup norm, denoted L . Under conditions of
uncertainty one makes decisions by ranking lotteries in L.
Von Neumann-Morgenstern NM axioms provided a mathematical formalization of
how to rank or order lotteries. Optimization according to such an order is called expected
utility EU and defines standard decision making under uncertainty. The main result from
the NM axioms is that the decision procedure is obtained by optimizing a function of the
following form:

Wx uxsdx, 6.1
sR
where the real line R is the state space, x : R RN is a lottery, u : RN R is a utility

function describing the utility provided by the outcome of the lottery in each state s, us,
and where dx describes a countably additive measure over sets of states or events in
R that determines their relative frequency. Arrow 11 explains why utility functions must
be essentially bounded so as to overcome the St. Petersburg paradox. This implies that the
space L of utility values provided by lotteries is L , namely, it is the space of measurable and
essentially bounded functions An essentially bounded function is a measurable function
that is bounded except perhaps on a set of measure zero. on the line R. Using the EU
criterion, a lottery x is ranked above another y if and only if W assigns to x a larger real
number:

x y Wx > W y , 6.2
where W satisfies 6.1. The optimization of expected utility EU is a widely used procedure
for evaluating choices under uncertainty. The Euler-Lagrange equations are typically used
to characterize optimal solutions.
In the following examples we consider the space of lotteries to be the space of all
continuous linear real-valued functions, L with the sup norm, and the dual of L , denoted
L , consists of all continuous real-valued functions on L . L includes integrable functions
on R as well as purely finite additive measures 1921 that are not representable by functions,
for example a measure that assigns measure zero to all bounded measurable subsets of the
line. Other examples are provided below. An example of a purely finitely additive measure
is the continuous linear real-valued function : L R defined by f lims fs on
all functions that have such a limit, and extended by using Hahn-Banach theorem to the rest.
6.1. Catastrophic Risks

A catastrophic risk is a rare event leading to major widespread losses. Expected utility
undervalues such risks: Chichilnisky 3, 4, 6 showed formally that using expected utility
criteria underestimates catastrophic risks, and by doing so conflicts with the observed
evidence of how humans evaluate such risks if they are likely to occur in their lifetimes.
In order to formalize the problem we need some definitions.
Definition 6.1. A ranking W : L R is said to be insensitive to rare events when Wx >
Wy Wx > Wy for any two lotteries x and y that are obtained by modifying
arbitrarily x and y on any set of states S R with an arbitrarily small-probability x, y.
Formally,
W is insensitive to rare events if x, y x, y : Wx > Wy Wx > Wy for
all x , y satisfying

y y a.e. on Sc R, x x a.e. on Sc R, where S < . 6.3
Otherwise, W is said to be sensitive to rare events.
Definition 6.2. W is insensitive to frequent events if

x, y x, y : Wx > W y W x > W y x , y s.t.
6.4
x x, y y a.e. on Sc R : S > 1 .
Otherwise, W is called sensitive to frequent events.
Mathematically, the problem with expected utility is that it is insensitive to rare events
no matter how catastrophic these may be.

Proposition 6.3. Any expected utility function sR ucssds, where s L1 R is
insensitive to rare events.
For a proof see the theorem in Chichilnisky 6.
6.2. New Axioms of Choice Under Uncertainty

Introduced in 3, 6 the following axioms contrast with Von-Neumann Morgensterns axioms
NM in that they treat symmetrically rare and frequent risks. They postulate that the ranking
of lotteries W : L R must satisfy the following.
Axiom 1. Sensitivity to rare events.
Axiom 2. Sensitivity to frequent events.
Axiom 3. Linearity and continuity of the ranking W.
Axioms 2 and 3 are standard; they satisfied, for example, by expected utility EU.
However Axiom 1 is not satisfied by EU Proposition 6.3 above.
To clarify the meaning of these axioms, the following are examples of rankings W that
do not satisfy our axioms.
Example 6.4. Consider a criterion of choice W : L R that ranks lotteries assigning measure
zero to any bounded set in R 1921. Such functionals are ruled out by Axiom 2 which
requires sensitivity to frequent events.
Example 6.5. Expected utility maximization is ruled out, as is shown in Proposition 6.3 above
see also 6, because it does not satisfy Axiom 1.
Like the NM axioms, the new Axioms 1, 2, and 3 lead to a representation theorem.
Theorem 6.6. There exist ranking criteria : L R that satisfy all three axioms. Any criterion
that satisfies the axioms is a convex combination of an expected utility plus a purely finitely additive
measure focused on catastrophic events, for example:

ucss
cs, 6.5
R
where cs describes the value of the lottery in state s R, s is an integrable real-valued
function on the line R, for example, s es , and u : R R is a bounded utility function.
The first term is thus an expected utility with an L1 R density function s, and the
second term is a purely finitely additive measure such as s limc cs for lotteries
that have such a limit and extended otherwise to all lotteries by Hahn-Banachs theorem. For
a proof see the study by Chichilnisky in 6. The optimization of functionals such as is
not amenable to standard tools of calculus of variations, which must be developed in new
directions; see, for example, the studies by Chichilnisky 3, 6 and Heal 17.
A recent result established that the new criterion in 6.5 is a way to formalize the
notion of optimizing expected utility while bounding the worst outcome in the case of a
catastrophe:
Theorem 6.7. Optimizing

the ranking criterion in 6.5 is equivalent to optimizing an expected
utility function R ucss subject to a constraint on the possible loss in case of a catastrophe.
For a proof see 4.

An interpretation of his theorem is as follows: the first term s
s1 uxs is an
s
integral operator with a countably additive kernel { }sZ which emphasizes the weight
of frequent events in the ranking of a lottery x l . The second purely finitely additive part
uxs assigns positive weight to rare events, which have small-probability according to .
Both parts are present, so is sensitive to small and to large probability events. Catastrophic
risks are therefore ranked more realistically by such functionals.
Next section applies this framework for evaluating risks to the two risk profiles
described above, asteroids and global warming.
7. Evaluating Asteroid Impacts

We first use standard methods for calculating the value of the damage caused by a risk
calculated by its present value, which brings future value into the present Posner 2 for
both, and in particular page 150 for the latter. The reader should refer to the risk profile of
an asteroid impact presented above.
i In the case of an asteroid as described above, the expected cost of the loss in any
given year is obtained by multiplying $120 trillion or $120 1012 by the probability of the
loss, which occurs on average once every 100 million years, and is therefore 108 . Once the
loss occurs, however, it is assumed to be permanent. Therefore the expected value of the loss
in year N is

120 1012 108 N
t 120 1012 108 N t
t1 t1
7.1

8
120 10 10
12 N
,
1
where is the time discount factor, 0 < < 1, and 1 is the discount rate.
If the risk does not occur in year N, then it can occur in year N
1, and if not in year
N
2, N
3, and so forth, and each time it occurs, it lasts permanently. Therefore the total
risk is the sum of the risk of it occurring in year 30, plus the risk of it occurring in year 31,
plus the risk of it occurring in year 32, and so forth, namely,

2
120 1012 108 N j 120 1012 108 N . 7.2
1 j1 1
At a 5% discount rate 0.95, and the total expected discounted value of the loss from such
as asteroid is

2
95 30 95/100
120 1012 108 9.2982 107 or about $92 million. 7.3
100 1 95/100
At a 10% discount rate, that value is
90 30 7.4
120 1012 108 81 4.1204 106 or about $4 million.
100
At a 3% discount rate, the value is
97 30 7.5
120 1012 108 322 : 4.9276 108 or about $500 million.
100
These values pale by comparison with the estimated value of other losses such as
global warming, which are estimated to be in the tens of trillions, as shown below. In all
cases, therefore, it appears to makes sense to allocate more funding to the global warming
problem than to the problem of preventing asteroid impacts; more on this below.
7.1. Evaluating Global Warming

The United Nations Intergovernmental Panel on Climate Change IPCC found in 1996 that
human-induced global warming is already occurring. There are two main scenarios for the
damages that global warming will cause: 1 catastrophic global warming eects, and 2
a more gradual buildup of damages
Scenario 1 Catastrophic Global Warming. The catastrophic scenario is described by Posner

in 2, page 181, as follows. There is rapid increase in temperature, which produces a damage
that is calculated to bring in a permanent loss of about $2 trillion a year in the US and
globally about $8 trillion a year. Considering the risk profile already established above cf.
also Posner, 1992, page 181, the present discounted value of such a disaster at a 3% discount
rate is
97
2 1012 2 1012 6.4667 1013 or about $65 trillion. 7.6
1 3
At a 5% discount rate the number is
95
2 1012 : 3.8 1013 or about $38 trillion, 7.7
5
and at a 10% discount rate the number is
2 1012 9 : 1.8 1013 or about $18 trillion. 7.8
Scenario 2 Gradual Buildup of Damages. In the second scenario global warming is also
here today, but temperature increases slowly and its damages increase for about 100 years
to reach 1% of the planets GDP. Global GDP is calculated to be about $120 trillion then
the same number used in the asteroid risk. After we reach maximum damage, we consider
various possibilities going forward: 1 the annual damage remains the same a perpetuity, and
2 damages decrease slowly and disappear 100 years later. Let us compute using standard
technique the present discounted value of the losses.
In the first case using a 3% discount rate we obtain

1012 97 i
i 1.077 8 1013 , which is about $10 trillion 7.9
i1
100 100
using a 5% discount rate

1012 95 i
i 3.8 1012 which is about $3.8 trillion, 7.10
i1
100 100
and using a 10% discount rate,

1012 9i
i : 9.0 1011 , which is about $900 billion. 7.11
i1
100 10
In the second case, when the damage gradually decreases until it vanishes after 100
years following its maximum impact we have using a 3% discount rate

100
1012 97 i 100
1012 97 100
i
i
100 i 1 9.7456 1012 $9.7 trillion,
i1
100 100 i2
100 100
7.12
using a 5% discount rate

100
1012 95 i
100
1012 95 100
i
i
100 i 1 3.7506 1012 $3.7 trillion
i1
100 100 i2
100 100
7.13
and using a 10% discount rate

100
1012 9i 100
1012 9 100
i
i
100 i 1 8.9993 1011 $890 billion.
i1
100 10 i2
100 10
7.14
As was indicated above, in all cases, and with all three discount rates, 3%, 5%, and 10%,
the global warming problem overwhelms in terms of present discounted values the costs
involved with asteroid impacts. This is despite the fact that even in the noncatastrophic case
global warming decreased GDP by a small fraction, only 1% and only after 100 years.
8. Comparing Global Warming and Asteroid Impacts

Using expected values we are led to believe that preventing asteroid impacts should not rank
high in our policy priorities. The results from the numerical computations provided above
to evaluate the risks of asteroids and global warming seem counterintuituve. How can it be
that a major disturbance caused by global warmingeven when we take very conservative
estimated losses of less than 1% of GDP building up slowly over 100 yearsoverwhelm the
costs associated with an asteroid impact that can plausibly lead to the extinction of the human
species? The expected value of the loss caused by an asteroid that leads to extinctionis
between $500 million and $92 billion, as seen above. A loss of this magnitude is smaller that
of a failure of a single atomic plantthe Russians lost more than $140 billion with the accident
at Chernobylor with the potential risks involved in global warming that is between $890
billion and $9.7 trillion. Common sense rebels against the computation we just provided.
Let us use other examples for comparison. In year 2004, the profits of the 10 biggest oil
companies were about $100 billion. It seems therefore unreasonable to think of losses from
asteroid impacts as valued between $500 million and $92 billion. At this rate, it seems dicult
to believe that we will ever do anything about averting human extinction, since current
priorities will always outweighed such infrequent events, no matter how important they
may be. Is there anything wrong with this argument? The alternative is to use the evaluation
criterion arising from Axioms 1, 2, and 3 above. In view of the representation theorem, the
next section utilizes a ranking function as defined in 6.5, and to make the computation
explicit, provides a plausible number for the parameter that appears in the definition of
above.
9. Catastrophes and the Survival of the Species

The axioms proposed here lead us to evaluate catastrophic risks by a formula that adds to the
present expected value a second term that focuses on rare events. To focus on catastrophes
that involve extinction of the human species, we evaluate the cost of an event using a sum of
the presented expected value plus the cost of extinction of the species. The most conservative
scenario for the cost of extinction of the species Posner 2 is when everyone alive today
dies without warning, and at probabilities that are so small that the value of a human life
is computed according to experimental evidence at about $50, 000. Recall Posner 2 that
the value of life decreases with the probability of death, so this number corresponds to events
with probabilities lower that 1 in 100 million. At such small probabilities, with the current
population, the species extinction event amounts to $600 trillion. We may therefore assume
that, in the following mathematical expression for our criterion,

x s uxs
uxs, 9.1
s1
the expression uxs $600 trillion.

To use this approach, we need to specify now a value for , which is the weight
one gives to extremely small-probability events in ones decision making. It is clear that the
weight to give to the second termwhich addresses the value of the low probability and
thus likely distant eventis somewhat subjective. Those concerned about the long-term fate
of our species may argue that we should consider it equally to the more familiar currentand
likely more frequent threat like global warming. Others may argue that our capability to deal
with such a threat will improve with time and therefore we should not spend too much on
a threat that will most likely only occur in the distant future. For this discussion we will just
note that many people would say that it is a mistake to spend money on something that will
occur millions of years from now devalue the threat for an incorrect reason. They incorrectly
conclude that something that occurs only every 100 million years on average will not occur
in their lifetime. But it can, in fact the likelihood is the same every year.
Example 9.1. Valuating the Parameter : An Illustration.
Consider the criterion defined in 6.5 above. In the studies by Chichilnisky 4, 79,
we assume that we give the catastrophic event a weight of only 1 in 100,000, namely,
1 105 , or equivalently 1 105 . 9.2
On this basis we can compare the global warming scenario with the asteroid collision
scenario. One takes into consideration the following.
1 Neither of the two cases of global warmingabrupt or gradualinvolve human
extinction.
2 The asteroid impact considered here does involve extinction of the human species.
Under those conditions, the total cost involved in global warming is approximately
$66 trillion at 3% discount rates, as shown above, while the total cost involved in an asteroid
impact neglecting the presented discounted value which is no larger than $500 million as
shown above, is about
$600 1012 105 6 109 6 billion. 9.3
Under the conditions, therefore, the yearly investment in prevention of asteroid impacts
should be about 1/10 of the yearly investment in prevention of global warming, which
is currently $1.7 billion Posner 2, page 182 leading to $170 million, while the current
expenditures are instead $3.9 million, requiring therefore to be increased by a factor of about
60.
Our rational decision maker who values the future of the species and understands
what probabilities really mean, could go through the following simple analysis. For any
value of even close to one-half the expected value we have calculated makes asteroids
more threatening than global warming that is attracting all the attention of policy makers
and the public today. In one sense this is satisfying since we would like to believe that we
would give great value to prevent our extinction. However, we used the number of US$300
trillion 1/2 for the expected value and argued that it is what we should spend to
defend against extinction. This does not seem intuitively correct for many reasons, not the
least of which is that we would have no resources left to do anything else. The answer to this
dilemma is to recognize that what we are really interested in is utility loss from extinction
rather than expected value for the dollars we allocate. This view can help us achieve an
intuitively pleasing answer that we should spend as much money today on defenses against
extinction as can be usefully transferred into improved protection. In the case of asteroids
based on current estimates many experts believe this might be only about 10 times what
we are now spending which is about US$30 million dollars. This is a small number and
the corrected valuation of the risk is high enough that we should need no further analysis
to decide to increase our eorts now and when new opportunities become available in the
future.
10. Conclusions
We believe that the above analysis is the beginning of a much more extensive assessment and
research about our response to all kinds of catastrophic risks. Recent results provide ways
to enhance our subjective judgments about the value of , which is approximated by the
marginal utility of avoiding extinction near the catastrophe, see the study by Chichilnisky in
4.
Other methods could include the application of Bayesian analysis involving experts
who understand the nature of the threats as well as the correct meaning of low probability
events. A Bayesian approach can be helpful to determine both the true risk profile and
the most plausible utility function for the use of resources to combat a given threat. Such
evaluations identify not only high expected value but also high utility. If there are very
expensive things we can do to prevent the risk the the allocations of a large amount of
resources may be warranted and the problem becomes more complicated. Our political
leaders will need to make the more dicult choices between meeting todays needs compared
with the need to defend against distant catastrophic threats. This is not a new challenge since
we and other nations spend a significant part of our resources to defend against the threat of
nuclear war or the nuclear winter that would follow it. What is new is that now we recognize
that many serious threats like those arising from glaciation, asteroid impact, and biodiversity
loss are unlikely to occur within our lifetimes, yet we do not want to wake up one day and
find that we are facing the impact of what was an avoidable catastrophic risk. Furthermore
the same type of deficiency in our approach also exists for very rare events like tsunamis and
earthquakes also leading to a poor allocation of resources, as was likely the case for the 2005
Asian tsunami. This work provides a framework to address these threats in a way that agrees
with our intuition. We would like to allocate resources in a way that can be useful in reducing
the catastrophic threats we face.
In conclusion we oer another perspective that might also be useful for understanding
why it is now that we are confronting the dilemmas. An analogy might help. Early on nobody
spent a lot of money on personal insurance to protect him/herself. As we gained more
knowledge of the risks we face and as we became auent enough we decided to spend
increasing amounts of money on insurance. In a similar way our species only recently has
obtained the knowledge of some of the catastrophic risks we face and developed ways to
cope with them. For the moment we are seriously underinsured so any way that we can do
useful things to reduce our risk we should do so. Someday in the future we may be challenged
as we were doing the cold war to decide between present risks and future ones.
Acknowledgments
This paper reports the research performed at the Columbia Consortium for Risk Management
CCRM, Columbia University, New York; see Columbia University Program on Information
and Resources PIR, http://columbiariskmanagement.org/leadership.php. Some of the
references are available at http://www.chichilnisky.com/. The authors are indebted to
Duncan Foley for valuable comments and suggestions. The authors acknowledge support
from the US Air Force, research grant FA9550-09-1-0467 Optimal Statistical Decisions
with Catastrophic Risks from AFOSR, Arlington VA to Columbia Consortium for
Risk Management CCRM directed by Graciela Chichilnisky and Columbia Universitys
Department of Economics, New York, NY 10027.
References
1 D. M. Raup, Extinction: Bad Genes or Bad Luck? W. W. Norton & Company, New York, NY, USA, 1st
edition, 1992.
2 R. Posner, Catastrophes, Risk and Resolution, Oxford University Press, Oxford, UK, 2004.
3 G. Chichilnisky, An axiomatic approach to choice under uncertainty with catastrophic risks,
Resource and Energy Economics, vol. 22, no. 3, pp. 221231, 2000.
4 G. Chichilnisky, Avoiding extinction: equal treatment of the present and the future, Economics, vol.
3, no. 2009-32, 2009.
5 S. M. McClure, D. I. Laibson, G. Loewenstein, and J. D. Cohen, Separate neural systems value
immediate and delayed monetary rewards, Science, vol. 306, no. 5695, pp. 503507, 2004.
6 G. Chichilnisky, Updating von Neumann Morgenstern axioms for choice under uncertainty with
catastrophic risks, in Proceedings of the Conference on Catastrophic Risks, The Fields Institute for
Mathematical Sciences, Toronto, Canada, June 1996.
7 G. Chichilnisky, Avoiding extinction: the future of economics, International Journal of Green
Economics, vol. 3, no. 1, pp. 118, 2009.
8 G. Chichilnisky, The foundations of probability with black swans, Journal of Probability and Statistics,
vol. 2010, Article ID 838240, 11 pages, 2010.
9 G. Chichilnisky, The foundations of statistics with black swans, Mathematical Social Sciences, vol. 59,
no. 2, pp. 184192, 2010.
10 J. Von Neumann and O. Morgenstern, Theory of Games and Economic Behavior, Princeton University
Press, Princeton, NJ, USA, 3rd edition, 1980.
11 K. J. Arrow, Essays in the Theory of Risk Bearing, North-Holland, Amsterdam, The Netherlands, 1970.
12 I. N. Herstein and J. Milnor, An axiomatic approach to measurable utility, Econometrica, vol. 21, pp.
291297, 1953.
13 M. Allais, The general theory of random choices in relation to the invariant cardinal utility function
and the specific probability function, in Risk Decision and Rationality, B. R. Munier, Ed., pp. 233289,
Reidel, Dordrecht, The Netherlands, 1988.
14 M. J. Machina, Expected utility analysis without the independence axiom, Econometrica, vol. 50, no.
2, pp. 277323, 1982.
15 M. Machina, Dynamic consistency and non-expected utility models of choice under uncertainty,
Journal of Economic Literature, vol. 22, pp. 16221668, 1989.
16 A. Tversky and P. Wakker, Risk attitudes and decision weights, Econometrica, vol. 63, no. 6, pp.
12551280, 1995.
17 G. M. Heal, Valuing the Future, Columbia University Press, New York, NY, USA, 2002.
18 A. G. Sanfey, J. K. Rilling, J. A. Aronson, L. E. Nystrom, and J. D. Cohen, The neural basis of economic
decision-making in the Ultimatum Game, Science, vol. 300, no. 5626, pp. 17551758, 2003.
19 K. Yosida and E. Hewitt, Finitely additive measures, Transactions of the American Mathematical
Society, vol. 72, pp. 4666, 1952.
20 K. Yosida, FunctionalAnalysis, Classics in Mathematics, Springer, New York, NY, USA, 4th edition,
1995.
21 A. Tversky and P. Wakkler, Risk attitudes and decision weights, Econometrica, vol. 63, pp. 12251280,
1995.
doi:10.1155/2010/976371
Research Article
Continuous Time Portfolio Selection under
Conditional Capital at Risk
Gordana Dmitrasinovic-Vidovic, Ali Lari-Lavassani,

Xun Li, and Antony Ware
The Mathematical and Computational Finance Laboratory, University of Calgary, Calgary,
AB, Canada T2N 1N4
Correspondence should be addressed to Gordana Dmitrasinovic-Vidovic, [email protected]
Received 29 October 2009; Revised 4 February 2010; Accepted 26 March 2010
Copyright q 2010 Gordana Dmitrasinovic-Vidovic et al. This is an open access article distributed
under the Creative Commons Attribution License, which permits unrestricted use, distribution,
and reproduction in any medium, provided the original work is properly cited.
Portfolio optimization with respect to dierent risk measures is of interest to both practitioners and
academics. For there to be a well-defined optimal portfolio, it is important that the risk measure be
coherent and quasiconvex with respect to the proportion invested in risky assets. In this paper we
investigate one such measureconditional capital at riskand find the optimal strategies under
this measure, in the Black-Scholes continuous time setting, with time dependent coecients.
1. Introduction
The choice of risk measure has a significant eect on portfolio investment decisions.
Downside risk measuresthat focus attention on the downside tail of the distribution of
portfolio returnshave received considerable attention in the financial world. Value at risk
VaR is probably the most famous among these measures, having featured heavily in various
regulatory frameworks. It can be defined for a random variable X and a confidence level by
VaRX EX q , where q is the -quantile of X see e.g., 1, Equation 1.2 Another
common definition is that the VaR of a loss distribution L is the smallest number x such
that P L > x . This is equivalent to the definition given here if we define the loss of
the portfolio X to be given by L EX X, and identify x VaRX.. A closely-related
downside risk measure is capital at risk CaR, defined in 2 see also 3, 4 as the dierence
between the riskless investment and the quantile q .
Quantile-based risk measures such as VaR and CaR suer from several shortcomings.
First, while they measure the best of the worst outcomes at the 1001 % confidence
level, they do not answer the question of how severe the loss can be. Also, one of the
most important concerns is that these measures are not in general subadditive; that is,
when used to measure risk, they do not always satisfy the notion that the diversification
should not create more risk 5. Finally, as illustrated in 6, VaR can exhibit multiple local
extrema.
These issues were addressed in the widely cited article by Artzner at al. 7, where the
authors define coherent risk measures by four conditions that such measures should satisfy.
The article motivated a number of authors 5, 813 to propose and investigate dierent types
of coherent risk measures, all of which are tail mean-based risk measures.
One such measure, that does not suer from the critical shortcomings of VaR and CaR,
is conditional capital at risk CCaR. This is defined in 14 as the dierence between the
riskless investment and the conditional expected wealth, under the condition that the wealth
is smaller than the corresponding quantile, for a given risk level. As such, this measure
provides an indication of the likely severity of the loss in the event that the loss exceeds a
given quantile. In this paper we prove that CCaR is strongly quasiconvex as a function of the
portfolio, which is an essential property for optimization. We investigate conditional capital
at risk in a multiasset Black-Scholes setting, in continuous time, and with time-dependent
coecients. We generalize and extend the optimization approach of Emmer at al. see 2, 14
to the continuous-time setting.
The outline of this paper is as follows. In Section 2, we give the notation and define
the portfolio process and CCaR. Section 3 provides the proof that CCaR is a coherent risk
measure and that it satisfies the property of strong quasiconvexity. In Section 4, we derive an
analytical solution for the minimal CCaR problem, up to a scalar constant which has to be
evaluated numerically. Section 5 is devoted to the derivation of an analytical strategy for the
maximal expected wealth, subject to constrained CCaR. Section 6 provides some numerical
examples, and Section 7 concludes this paper.
2. Preliminaries
We introduce the following notation. The m-dimensional column vector with each
component equal to 1 is denoted by e, the Euclidean norm of a matrix or vector by ,
and the space of Rn -valued, square-integrable functions defined on 0, t by L2 0, t, Rn , or
just L2 . The natural inner product of this space is denoted by , t , and the corresponding
norm by t .
We work under the following assumptions.
Assumption 2.1. i The securities are perfectly divisible.

ii Negative positions in securities are possible.
iii Rebalancing of the holdings does not lead to transaction costs.
Assumption 2.2. i m 1 assets are traded continuously over a finite horizon 0, T .
ii m of these assets are stocks that follow the generalized Black-Scholes dynamics:

m
dSi t Si tbi tdt ij tdW j t, t 0, T , Si 0 > 0, i 1, . . . , m, 2.1
j1
where W j t, j 1, . . . , m, are independent standard Brownian motions.

iii One of the assets is a bond, whose price S0 t, t 0, evolves according to the
dierential equation:
dS0 t rtS0 tdt, t 0, T , S0 0 S0 > 0, 2.2
where rt> 0 is the interest rate of the bond. Throughout this work, we assume that
borrowing in the bond is unconstrained.
iv The volatility matrix t with ijth element ij t, its inverse 1 t, the drift
vector bt : b1 t, . . . , bm t , and the interest rate rt are deterministic, Borel measurable,
bounded functions over 0, T , so that they belong to the appropriate L2 spaces.
v t satisfies the nondegeneracy condition:
x tt x x x, t 0, T , x Rm , 2.3
where > 0 is a given constant.
We note that, under the above assumptions, the market is complete.

At any time t, Ni t shares are held in the asset Si t, leading to the wealth X t

Ni tSi t. The m 1-dimensional vector-valued function Nt N0 t, . . . , Nm t is
called the trading strategy. We denote the fraction of the wealth X t invested into the risky
asset Si t by
Ni tSi t
i t , i 1, . . . , m, 2.4
X t
and call t : 1 t, . . . , m t Rm the portfolio. The fraction held in the bond is 0 t
1 te. Under the assumption that the trading strategy is self-financing, the wealth process
follows the dynamics

dX t X t rt Bt t dt ttdWt , X0 X0 , 2.5
where X0 is the initial wealth, and the risk premium vector Bt is defined by
Bt : bt rte, t 0, T . 2.6
Proceeding as in 3, to ensure a minimal tractability of the optimization problems

which we solve through the following sections, we restrict our attention in this work to
the class Q of portfolios which are Borel measurable, deterministic, and bounded over
0, T . Such portfolios are called admissible. Note that for an admissible portfolio , 2.5 is
guaranteed to have a strong solution X see 15, Theorem 5.2.9. Note that, by allowing
for time-dependent coecients, this generalises the class of portfolios considered in 14.
Under condition 2.3 the market price of risk is uniquely defined by
t t1 Bt. 2.7

It will be shown throughout this work that the magnitude of the L2 -norm of the market price
of risk is the determining criterion for optimal investment strategies, which turn out to be the
weighted averages of the bond and Mertons portfolio defined by
M t : tt1 Bt. 2.8
This result is an illustration of the mutual fund theorem in complete markets.

We now recall the definitions of the quantiles and quantile-based risk measures for
X t.
Definition 2.3. Suppose that Ft x is the cumulative distribution function of the wealth X t,
at time t 0, T . For a risk level 0, 0.5, the -quantile of X t is defined as
q X0 , , t : inf {x R | Fx }. 2.9
The tail mean or expected shortfall of the wealth process X t, which we denote by
TM X t, is the expected value of X t conditional on X t q X0 , , t, that is,

TM X t : E X t | X t q X0 , , t . 2.10
The conditional capital at risk, which we denote by CCaRX0 , , t, is defined to be the
dierence between the riskless investment and the tail mean, that is,
CCaRX0 , , t : X0 Rt TM X t, 2.11
where Rt is defined as
t
Rt exp rsds . 2.12
0
Remark 2.4. Note that, with CCaR defined in this way, we get the following.
i An investment in a riskless asset corresponds to zero CCaR.

ii An increase in the tail mean corresponds to a decrease in CCaR. Thus, negative
CCaR corresponds to the case when the tail mean is above the riskless return, that
is, a desired result, while positive CCaR corresponds to the case when the tail mean
is below the riskless return which we try to reverse by minimizing CCaR.
It was shown in 3 that the -quantile of the wealth process X t can be written as

1

2

q X0 , , t X0 Rt exp B, t t |z | t , 2.13
2
where, for a given risk level , z denotes the corresponding -quantile of the standard
normal distribution note that this is negative when < 0.5. In the following, we let and
denote the density and cumulative distribution functions of the standard normal random
variable. We have the following proposition.
Proposition 2.5. The tail mean of the wealth process X t solving 2.5 can be expressed as
1
TM X t X0 Rt expB, t |z | t , 2.14

and the expected value is given by
EX t X0 Rt expB, t . 2.15
The proof is given in the appendix. From Proposition 2.5 we get the following
corollary.
Corollary 2.6. The conditional capital at risk of the solution to 2.5 can be written as

1
CCaRX0 , , t X0 Rt 1 expB, t |z | t . 2.16

Throughout the paper we will use the following function:

g , t : B, t ln |z | t ln , 2.17
which transforms the expression for conditional capital at risk into

CCaRX0 , , t X0 Rt 1 exp g , t . 2.18
We now turn our attention to investigating the properties of CCaR.
3. Coherency and Quasiconvexity of CCaR

The notion of coherent risk measures has been widely discussed in the recent literature see
5, 713 e. g.. Below we recall the definition of coherency, confirm that CCaR is a coherent
risk measure, and prove that it is quasiconvex.
Definition 3.1. Let V be the set of real-valued random processes on a complete probability
space , F, P , with its natural filtration {Ft }t0,T , satisfying ESt < for all St V ,
and for t 0, T . Then : V R is a coherent risk measure if it satisfies the following
properties.
1 Subadditivity: is subadditive if, for all random processes St,Y t V ,
St Y t St Y t for t 0, T . 3.1
2 Positive Homogeneity: is positive homogeneous if, for all St V and constant
c > 0,
cSt cSt for t 0, T . 3.2
3 Monotonicity: is monotone if, for all St,Y t V , such that St Y t almost
everywhere, and S0 Y0 ,
St Y t for t 0, T . 3.3
4 Translation Invariance: is translation invariant if, for all St V and c R,
St c St for t 0, T . 3.4
To prove that CCaRX0 , , t of a wealth process X t V is a coherent risk measure

we refer to 8, Definition 2.6, where the expected shortfall of a random process Xt, with
the corresponding -quantile q , is defined as

ES Xt 1 E XtIXtqx qx P Xt qx 3.5
and is shown to be coherent 8, Proposition 3.1. We should note that the definition of
coherency used in 8 involves translation invariance in the sense that Xt c Xtc,
for all Xt V and c R.
In order to relate the result in 8 to the coherency of CCaR in the sense used here,
we first note that, if Xt is a random process with a continuous probability distribution,
we have ES Xt TM Xt, so that with a slight abuse of notation CCaRXt
X0 Rt ES Xt. If we consider the shifting of the portfolio value by an amount c, we have
CCaRXt c X0 Rt c ES Xt c
X0 Rt c ES Xt c 3.6
X0 Rt ES Xt CCaRXt,
and so we have the following result.
Corollary 3.2. Conditional capital at risk CCaRX0 , , t of a wealth process X t V , at time
t 0, T , for a risk level 0, 0.5, is a coherent risk measure.
Remark 3.3. i While properties 2 and 3 are quite natural, subadditivity has the obvious, yet
important consequence that diversification does not create more risk.
ii Properties 1 and 2 together are very important as they guarantee the convexity of
the risk measure which is essential for optimization.
iii In our definition of a coherent risk measure we say that a risk measure is
translation invariant, if Xt c Xt, for all Xt V , and c R. While in article
8, It is said that arisk measure is translation invariant, if Xt c Xt c, for all
Xt V and c R It is our belief that the definition of the translation invariance, as given
in this paper, corresponds more with the intuition behind the notion of translation invariance
than the definition given in 8.
Remark 3.4. In the above remarks we investigated the impact of diversification to the portfolio
risk, as measured by CCaR. We saw that the axiom of subadditivity requires that the risk
of the sum of two risky processes, say two stock price processes, be less than or equal to
the sum of the individual risks associated with these stock price processes. This means that
diversification does not create more risk, as measured by CCaR, and that, as long as we
diversify, we expect risk reduction.
However, if the portfolio consists of all market assets, the diversification is completed.
Then the risk can be further reduced by optimization, that is, by rebalancing the positions
across these assets. We then look at the risk measure CCaR as a function of the portfolio ,
and prove that it is strongly quasiconvex in which further implies the uniqueness of the
corresponding optimization problems solutions.
We now turn to the notion of strong quasiconvexity, which we note has not been
discussed in the context of portfolio optimization in any of the quoted references except in
3, 4. Its usefulness lies in its role in establishing the existence of unique solutions to portfolio
optimization problems.
We first recall see 16, Definition 3.5.8 that a function f : U Rm R is said to be
strongly quasiconvex if

f 1 < max f, f , U,
/ , 0, 1. 3.7
In the following theorem we prove that CCaR has this important property, when viewed as a
function of the portfolio weights.
Theorem 3.5. For all distinct , Q (where the set {t 0, T | t /
t} has a positive
Lebesgue measure), and for all 0, 1
CCaRX0 , 1 , T < max{CCaRX0 , , T , CCaRX0 , , T }. 3.8
Proof of Theorem 3.5. We suppose, without loss of generality, that
CCaRX0 , , T > CCaRX0 , , T , 3.9
in which case, from 2.18, we have
g , T < g , T . 3.10
Let 0, 1. We claim that
CCaRX0 , 1 , T < CCaRX0 , , T , 3.11
which is equivalent to the statement
g 1 , T > g , T . 3.12
For ease of notation, we define the function
x : ln|z | x. 3.13
Since g , T B, T ln T , 3.10 is equivalent to

B, T T T > 0, 3.14
while 3.12 becomes

1 B, T 1 T T > 0. 3.15
Clearly, from 3.14, the left-hand side of 3.15 is greater than

1 T T 1 T , 3.16
and the theorem will be proved if we can establish that this is nonnegative. In order to do
this, we make use of the following lemma, the proof of which is given in the appendix.
Lemma 3.6. The function kx ln x is decreasing and strictly concave for all x > 0.
Since x k|z | |x|, one can make use of Lemma 3.6 to establish the following:

1 T T 1 T
3.17
T 1 T .
The proof is complete.
This theorem has an immediate, important consequence. Namely, from 16, Theorem
3.5.9, if a function f : U Rm R is strongly quasiconvex, then its local minimum is its
unique global minimum. Therefore, the following corollary is true.
Corollary 3.7. If CCaRX0 , , T has a local minimum at Q, then is its unique global
minimum.
This guarantees that minimization of CCaRX0 , , T with respect to gives the

true, global minimum. To illustrate strong quasiconvexity, we give the plot of CCaR for
Quasiconvexity of CCaR
25
20
15
CCaR 10
5
5 4 3 2 1 0 1 2 3 4 5
Stock investment
Figure 1: The graph of CCaR as a function of the fraction of wealth invested in the stock, for 5, 5,
T 16, S0 10, r 0.05, 0.2, b 0.15. Observe that, in this case with a long-time horizon
and constantly high-performing stocks, the global minimum lies in the interval 0, 1 and satisfies
CCaR < 0.
the following example where the parameters were chosen to represent rather extreme
conditions, for the purposes of illustration.
Example 3.8. We consider a market consisting of one stock following the SDE
dSt St0.15dt 0.2dWt, t 0, 16, S0 10, 3.18
and the bond with the constant interest rate r 0.05. The graph of CCaR is given in
Figure 1.
4. Minimal Conditional Capital at Risk

The first portfolio selection problem we consider is to minimize risk as measured by CCaR,
that is, to determine its minimal value over all Q. The problem can be stated as
min CCaRX0 , , T . 4.1

Q
Using 2.18, problem 4.1 transforms into

min X0 RT 1 exp g , T . 4.2
Q
Since RT is a constant, problem 4.2 is equivalent to
max g , T . 4.3
Q
We now introduce the fundamental dimension reduction procedure, used throughout this
work. Following 2, 14, we project the optimization problems considered in this paper onto

the family of surfaces Q { L2 : 2T 2 }, and note that Q 0 Q .
We denote by g the restriction of g to Q , so that
g , T B, T ln|z | ln . 4.4
Taking into account the definition of Q , problem 4.3 can be stated as
max max g , T . 4.5

0 Q
We deal with this problem in two stages. First, fixing reduces the problem to
max g , T . 4.6
Q
If denotes the unique maximising portfolio for this problem, then 4.5 can be solved
through the one-dimensional problem:
max g , T . 4.7
0
It remains to solve the subproblem 4.6. Since is fixed, we see from 4.4 that 4.6 is
equivalent to the problem
maxB, T , 4.8
Q
the solution of which is given by the following proposition the proof of which is given in 6,
Proposition 2.1.
Proposition 4.1. For a fixed , the solution of problem 4.6 is attained by
1
t tt Bt, 4.9
T
where denotes the market price of risk, defined in 2.7.
We are now ready to solve problem 4.3, and we have the following theorem.
Theorem 4.2. Let be the market price of risk 2.7. g , T attains its maximum when
1
t : t 1
T tt Bt, 4.10
where, if T |z |/, is defined to be the unique solution of the equation
|z |
T , 4.11
|z |
and, if T < |z |/, 0. The corresponding minimum conditional capital at risk is

1
CCaRX0 , , T X0 RT 1 exp T |z |
4.12

(which is 0 if 0), and the expected wealth is
EX T X0 RT exp T . 4.13
Proof of Theorem 4.2. Using the definition of t and substituting 4.9 into 4.5 allows us to
rewrite 4.5 as
max g , T : T ln|z | ln . 4.14

0
If we define the function
f : g , T T ln|z | ln , 4.15
we get
|z | |z |
f T T ,
|z | |z |
|z | |z |
f 0 T T , 4.16
|z |
|z |
f 2
1 |z | |z | |z | .
|z |
We see that f has the same form as k x, where kx is defined in Lemma 3.6,
with |z | , instead of x, so that f 0. Since f is a decreasing function of for > 0,
we have two cases.
i If f 0 0, that is, T |z |/, then the equation
|z |
f T 0 4.17
|z |
has a unique solution which we denote by . Since f is a concave function, it

reaches its maximum at .
ii If f 0 < 0, that is, T < |z |/, then f < 0 for all 0, and so 4.17 has
no solution. This implies that f is decreasing for all 0, so that the optimal
solution to problem 4.14 is 0.
To complete the proof of Theorem 4.2, note that 4.9, with , is the optimal
solution of problem 4.3. One can then write
2
1 BT T , t 2 ,
2
B, T 4.18
T
leading to

1
CCaRX0 , , T X0 RT 1 exp T |z | ,
4.19

EX
t X0 RT exp T .
Remark 4.3. i Note that, if T |z |/, we can deduce by using A.12 that
|z | |z |
T |z |. 4.20
1 |z |
Therefore, the condition T |z |/, which has to be satisfied for investing into stocks
under conditional capital at risk, is stronger than the condition T |z |. The latter
condition is sucient in order to include stocks into the optimal strategy using capital at
risk as a risk measure. Otherwise stated, conditional capital at risk is a more conservative risk
measure than capital at risk, which is consistent with its definition.
ii Increasing the time horizon T leads to increasing the L2 norm of the market price
of risk, so that, in case ii, the optimal strategy changes from a pure bond strategy to a mixed
bond-stocks strategy. In case i, increasing the L2 norm T leads to increasing the expected
wealth EX T and decreasing conditional capital at risk CCaRX0 , , T .
iii As was noted in the preliminary remarks, the optimal portfolio is a weighted
average of Mertons portfolio and the bond, which is an illustration of the two-fund
separation theorem.
We note that the solution provided in Theorem 4.2 is not an explicit analytical solution,
but it is expressed in terms of the solution of the one-dimensional equation 4.11, whose
solution can be easily computed. Analytical upper and lower bounds for are given in
the following lemma, whose proof is given in the appendix.
Lemma 4.4. The unique solution of 4.17 satisfies the inequality

1
T 1 |z | T |z |, 4.21
|z |2
where a : max{a, 0}.

Remark 4.5. Note that, for 0.05, that is, at the 95% confidence level, 1 1/|z |2 0.63,
4.21 approximates the result from 14, which states

2
T |z | T |z | for < 0.15. 4.22
3
However, at a higher confidence level, that is, < 0.05, 4.21 gives a better approximation
for , that is, a smaller interval to which belongs.
5. Portfolio Optimization with Respect to Conditional Capital at Risk

We now turn to the problem of maximizing wealth subject to constrained CCaR, that is,
max EX T subject to CCaRX0 , , T C. 5.1

Q
Using 2.18, the above problem can be written in the form

max X0 RT expB, T subject to X0 RT 1 exp g , T C, 5.2
Q
which is equivalent to
C
maxB, T subject to exp g , T 1 . 5.3
Q X0 RT
Since CCaR, from its definition, is smaller than total wealth, to avoid trivial cases we only
consider C such that
C < X0 RT 5.4
Under condition 5.4, problem 5.3 can be written as
maxB, T subject to g , T c, 5.5

Q
where

C
c ln 1 . 5.6
X0 RT
Note that condition 5.4 guarantees that c is well defined. Using the dimension reduction
procedure of 2, 14 see Section 4, we can write problem 5.5 as a one-parameter
optimization problem:
max maxB, T subject to g , T c. 5.7

0 Q
The solution to 5.7 is given in the following theorem.
Theorem 5.1. Suppose that the constant risk level C satisfies the following condition:

1 |z |
X0 RT 1 exp T |z | C < X0 RT , if T

5.8
|z |
0 C < X0 RT , if T < ,

where is defined in Theorem 4.2. Then the optimal solution to problem 5.7 is
1
t tt Bt, 5.9
T
where c is defined in 5.6, and , is the unique solution of the equation
h : T ln|z | ln c 0. 5.10
The corresponding expected wealth is equal to
EX T X0 RT exp T , 5.11
with the corresponding conditional capital at risk:
CCaRX0 , , T C. 5.12
The ecient frontier is given by the following curve, whose first component is an increasing function
of C defined implicitly through 5.6, 5.10, and 5.11:
{C, EX T | C satisfies 5.8}. 5.13
Proof of Theorem 5.1. Proposition 4.1 states that the problem
maxB, T 5.14
Q
has the optimal solution
1
t t t Bt. 5.15
T
Substituting 5.15 into 5.7 transforms it into the problem
max T subject to g , T c. 5.16

0
Using 5.10, the above problem can be written as
max T subject to h T ln|z | ln c 0. 5.17

0
Clearly, the problem achieves its optimal solution for maximal for which the constraint is
satisfied. The solution to this problem is given in the following lemma, whose proof is given
in the appendix.
Lemma 5.2. Under condition 5.8, the equation h 0 has a maximal solution , .
Furthermore, h < 0, for > .
We substitute into 5.15 to get the optimal portfolio
1
t tt Bt, 5.18
T
which yields
EX T X0 RT exp T , CCaRX0 , , T C. 5.19
Finally, to prove that the expected wealth is an increasing function of the risk constant C, we
rewrite 5.10 in the form

C
h T ln|z | ln ln 1 0, 5.20
X0 RT
which defines as an implicit function of C. Dierentiating the above, we get
h d h
0, 5.21
dC C
so that
d h/C
. 5.22
dC h/
Since h/C 1/X0 RT C, and since C < X0 RT , it follows that h/C > 0. From
Lemma 5.2 it follows that dh/d < 0, for all > , and hence for , so that d /dC > 0.
Thus, is an increasing function of C. From 5.11, we see that the expected wealth is an
increasing function of , and consequently of C. This completes the proof of the theorem.
Remark 5.3. If a risk-averse investor decides to take the minimal risk which, in the first case,
means taking for the risk constant

1
C X0 RT 1 exp T |z | , 5.23

implying
c ln T ln|z | , 5.24
then the unique solution of the equation
T ln|z | ln c 0 5.25
is .
In the second case, for a risk averse investor, the minimal risk constant is C 0, which
implies c 0, leading to the equation
h T ln|z | ln 0, with h0 0, h 0, 5.26
so that the unique solution of the equation h 0 is 0.

Therefore, in both cases, choosing the minimal risk constant leads to the optimal
strategies coinciding with the strategies from the previous chapter, in which minimal
CCaRX0 , , T was determined.
6. Applications
To illustrate the results from previous sections, we give some numerical examples. We recall
that the system of SDEs that models the stocks prices is

m
dSi t Si tbi tdt ij tdW t,
j
Si 0 > 0, i 1, . . . , m, 6.1
j1
and that the stocks returns variance-covariance matrix, which we denote by t, is equal to
tt . We also recall that t can be decomposed as
t ttt, 6.2
where t is the stocks returns correlation matrix, and t is a diagonal matrix with the
entries equal to the stocks returns standard deviations. Therefore, from 6.2 we get
t tt ttt. 6.3
Although, theoretically, we assume that the vector of independent Brownian motions Wt
is {Ft }t0,T -adapted, that is, known at time t 0, T , it is a common practice that we only
observe t or, equivalently, t and t, but not t. From 6.3 we see that this leads to a
nonunique decomposition of t into the product tt . Despite that fact, the Euclidean
norm, and consequently, the L2 -norm of the market price of risk, is uniquely determined by
2

t2 t1 Bt Bt tt1 Bt, 6.4
or, in the terms of the standard deviation and the correlation matrix, as
t2 Bt t1 t1 t1 Bt. 6.5
To keep the exposition simple we consider the cases where the interest rate rt, as
well as the volatility matrix, is constant, and the number of stocks is m 3. In all numerical
computations and the corresponding plots we use an annual time scale for the drifts, standard
deviations, and the correlation matrix t of the 3 stocks. We model time dependency in the
drift bi t by assuming that it oscillates periodically around a constant level i . In order to
capture cycles in the economy or in the dynamics of the stocks we model the drifts as

bi t i i cos i t , i 1, 2, 3. 6.6
We note that the above model for stocks drifts was already used in 17, as the amplitude and
frequency coecients i and i allow a high degree of flexibility in adjusting the shape of this
time dependency. We also note that, when modeling real market data, it is quite easy to deal
with the above functional form and estimate these two parameters by maximum likelihood
techniques, rather than detrending the data.
We now look at four special cases with the following characteristics.
i We let 1 2 3 , with 0.75; that is, the economic cycles of all three stocks
are the same. We consider 1 0.75, 2 0.5, 3 0.25, with 0.015, which
corresponds to a 1.5% deviation around the constant values i .
ii We assume that the interest rate is r 0.05 and numerically explore the sensitivity
of the optimal strategies with respect to i and .
iii We assume that the stocks returns have constant standard deviations given as
follows:
1 20%, 2 25%, 3 30%. 6.7
To emphasize the importance of the diversification eect and, consequently, of the

market price of risk, for the optimal strategies, in Examples 6.2, 6.3, and 6.4 we keep the same
drift and volatility coecients and change the correlation matrix.
Example 6.1. We assume that 1 0.12, 2 0.10, and 3 0.08, and that the correlation
matrix is

1.0 0.6 0.8

0.6 1.0 0.5
. 6.8
0.8 0.5 1
In Figure 2, we show a the stocks drifts over a ten year period, with daily granularity,
and b the optimal strategy, under minimal CCaR.
In this example, we see the expected result, that is, stock 1, which has the largest
constant part in the drift, and the smallest volatility is present in the optimal portfolio in
0.14 3
0.13 1 t
b1 t 2
0.12
1 2 t
0.11 3 t
0.1 b2 t
0
0.09
b3 t 1
0.08
0.07 2
0.06
3 0 t
0.05 rt
0.04 4
0 2 4 6 8 10 0 2 4 6 8 10
Time years Time years
a b
Figure 2: Plot of the Stock Drifts a and portfolio weights b for Example 6.1.
0.13 1
0.12 b3 t
0.8
0.11 1 t
b2 t 0.6
0.1
0.09 0.4 2 t
b1 t
0.08 0.2 3 t
0.07
0
0.06
rt 0.2 0 t
0.05
0.04 0.4
0 2 4 6 8 10 0 2 4 6 8 10
Time years Time years
a b
Figure 3: Plot of the Stock Drifts a and portfolio weights b for Example 6.2.
the highest percentage. However, to assume that just drifts and volatilities determine the
optimal strategy would be misleading, as the following examples show.
Example 6.2. We assume that 1 0.08, 2 0.10, 3 0.12, and that the correlation matrix is
the same as in Example 6.1.
The stocks drifts and portfolio weights are given in Figure 3.

We see that, despite the fact that stock 1 has the smallest constant part of the drift,
it is still present in the optimal portfolio in the highest percentage, due to its high negative
correlations with both stocks 2 and 3.
3 t
4 2 t
0
1 t
2
6 0 t
8
0 2 4 6 8 10
Time years
Figure 4: Portfolio Weights under Minimal CCaR for Example 6.3.
Example 6.3. In this example we assume that 1 0.08, 2 0.10, and 3 0.12; that is, the
stocks have the same constant parts of the drifts as in Example 6.2 but a dierent correlation
matrix given by

1.0 0.3 0.5

0.3 1.0 0.9. 6.9
0.5 0.9 1
The optimal strategy, under minimal CCaR, is given in Figure 4.

In this example, stocks 2 and 3, which are highly negatively correlated, are present
in the optimal portfolio in percentages above 400%, while stock 1 and the bond are being
borrowed.
Example 6.4. In this example, we assume again that the stocks have the same constant parts
of the drifts as in Examples 6.2 and 6.3, while the correlation matrix is

1.0 0.2 0.3

0.2 1.0 0.1 . 6.10
0.3 0.1 1
The criterion for investing into stocks, under the minimal CCaR, is T |z |/.
For 0.05, |z | 1.645, so that we must have T 2.0620, in order to start investing into
stocks. In the case of low correlation coecients, such as in the above example, the market
price of risk increases very slowly, so that it takes the time horizon of 34 years to invest into
stocks with the risk level of 0.05. For 0.01, |z | 2.33, so that we must have T 2.6424
to start investing into stocks. If we choose the risk level to be 0.01, we have to assume the very
106
5.5
4.5
Ecient frontier
Expected wealth $

4
3.5
2.5
1.5
1
0 200 400 600 800
CCaR $
Figure 5: Ecient Frontier for Example 6.1.
Table 1
T 10 20 30 33 34 40 50 56 57
T 1.1420 1.5944 1.9344 2.0302 2.0705 2.2293 2.5030 2. 6424 2.6655
long time horizon, of 57 years, in order to include stocks into the optimal strategy, under
minimal CCaR. The relation between T and T can be found in Table 1.
Ecient Frontiers
Figure 5 shows the ecient frontier for Example 6.1 created using Theorem 5.1. The theorem
states that the expected wealth is an increasing function of C, which is bounded above by
Cmax X0 RT . 6.11
In order to avoid extremely risky strategies as C Cmax , and for the sake of more
transparency of the graphs, we restrict C to the interval 0, 0.5Cmax . We note that the ecient
frontiers for the other three examples are of the same exponential form, so that we omit their
graphs.
The graph given in Figure 5 illustrates the fact that the ecient frontiers are an
increasing function of the risk constant C which bounds CCaR.
Time-CCaR-Expected Wealth Surfaces

In Figure 6 we give a three-dimensional plot of the expected wealth as a function of time and
the risk constant C. From Theorem 5.1 we know that the expected wealth increases when the
risk constant C increases. We recall that the expected wealth is also an increasing function
of the time horizon T . The following figure illustrates these facts for Example 6.1. In order
106
5
Ecient surface
Expected wealth $

4
800
600 10
400 5
CC r s
aR
% 200 yea
0 T e
i m
Figure 6: Time-CCaR-Expected Wealth Surface for Example 6.1.
to avoid extremely risky strategies, and to get a better representation, we restrict the upper
bound for CCaR to the interval 0, 0.5Cmax .
7. Conclusion
In this work we investigated continuous time portfolio selection under the notion of
conditional capital at risk, within the Black-Scholes asset pricing paradigm, with time
dependent coecients. We showed that conditional capital at risk is a coherent risk measure
and proved that it satisfies an important property of strong quasiconvexity. Based on an
idea from 14, generalized in 3, we introduced the fundamental dimension reduction
procedure which transforms m-dimensional optimization problems into one-dimensional
problems, within the class of admissible portfolios, which are Borel measurable, bounded,
and deterministic. We further developed optimal strategies for portfolio optimization under
constrained conditional capital at risk. It is important to emphasize that we considered time
dependent portfolios, where the methods developed in 14 no longer work. We illustrated
the two-fund separation theorem, by showing that all optimal strategies are the weighted
averages of Mertons portfolio and the bond, and the weights depend on the investors risk
tolerance only. Finding optimal strategies requires solving nonlinear equations which include
the cumulative distribution function of the normal random variable, so that the weights
can be only found numerically. We provide several numerical examples which illustrate
the importance of diversification, given by the correlation matrix. The correlation matrix
significantly impacts the magnitude of the L2 norm of the market price of risk, which, in
turn, is the determining criterion for investment strategies.
Appendix
Proof of Proposition 2.5. Let
t rt Bt t, t t t. A.1

Then the dierential equation of the wealth process given by 2.5 can be written as

dX t X t tdt tdWt , X 0 X0 . A.2
For convenience, set Y t ln X t. Applying the multidimensional version of Itos Lemma,
it can be shown see the proof of Proposition 2.1 of 3 that Y t follows the dynamics

1 2
dY t t t dt tdWt, Y 0 lnX0 Y0 . A.3
2
The -quantile of Y t is equal to
t
1 2 t
q Y0 , , t Y0 s s ds |z | s2 ds, A.4
0 2 0
so that the -quantile of Xt is equal to q X0 , , t expq Y0 , , t. We will further
simplify the notation by introducing
qx q X0 , , t, qy q Y0 , , t. A.5
From
X tIX qx expY tIY qy , A.6
where I is the corresponding indicator function, and from Bayes theorem, we obtain

TM X t 1 E expY tIY qy . A.7
Choose t 0, T. We will evaluate A.7 using the characteristics of the distribution of Y t,
for fixed t. From A.3 we have that Y t is a normal random variable with the parameters
t
1
2

t
s2 ds. A.8
y Y0 s s ds, y
0 2 0
Taking into account A.7 and A.8, we can write
qy
1 1

TM X t E expY tIY qy

exp y y dy. A.9

Using the standard integration techniques, and taking into account that z < 0, we get

1
y2
TM X t exp y

|z | y . A.10
2
If we substitute the original notation from A.8 and A.1, we get 2.10. This completes the
proof of Proposition 2.5.
Proof of Lemma 3.6. Let x > 0. The function kx ln x is clearly decreasing. To prove
that it is strictly concave, we dierentiate twice to get
x
k x 2
1 xx x . A.11
x
To evaluate k x, we use the following standard inequality. For x > 0,
1 1 1 x 1
. A.12
x x3 x x
Applying the above, we get that k x 0 which means that kx is concave. To prove that
kx is strictly concave we need to show that
x1 x < x for x > 0. A.13
If we define the function
wx : x1 x x, A.14
it is an easy exercise to prove that wx < 0, for all x > 0; that is, the function kx is strictly
concave for all x > 0, which ends the proof of Lemma 3.6.
Proof of Lemma 4.4. We prove that the optimal solution of 4.17 satisfies 4.21. From 4.17
we have that
1 |z | 1

. A.15
|z | T
We use again the standard inequality A.12. From inequality A.12, with x |z | , we
have
1 1 1 1 |z | 1

. A.16
|z |
|z | 3 T |z | |z |
The right-hand side of the inequality implies
T |z |. A.17
From the left-hand side of the inequality we have

1
T 1 |z | . A.18
|z | 2
Since
1 1
, A.19
|z | 2 |z |2
inequality A.18 implies

1
T 1 |z | . A.20
|z |2
The fact that 0 and A.20 give the desired result, so that Lemma 4.4 is proved.
Proof of Lemma 5.2. In this proof we find a maximal solution of the equation
h T ln|z | ln c 0. A.21
Since h f , defined by 4.15, we will apply some arguments from the proof of
Theorem 4.2.
i Suppose that T |z |/, which means that h 0 f 0 0. Then the unique
maximum of h is achieved at . From condition 5.8 i we have that

C
c ln 1
X0 RT

1 A.22
ln 1 1 exp T |z |

ln T ln|z | .
Hence
h T ln|z | ln c
T ln|z | ln A.23
ln T ln|z | 0.
Therefore, h 0. We further distinguish the following two subcases.
a Suppose that C < 0. Then A.21 implies that h0 c < 0. From the proof
of Theorem 4.2, for > , h 0, h is concave, and since h 0, it
follows that the equation h 0 has at least one solution, with the bigger
solution , .
b For C 0, h0 c 0. Using the same arguments as in case a, we get that
A.21 has a unique solution , .
ii Suppose T < |z |/, that is, h 0 < 0. We recall that in this case 0. The
definition of h yields
h0 ln|z | ln c c. A.24
From condition 5.8 ii, we have that c 0, that is, h0 0. Using h f , and the
proof of Theorem 4.2, we have
|z |
h 0; h 0 T < 0. A.25

Therefore, h < 0 for all 0, with h0 0, so that h 0 has a unique solution
, . This completes the proof of Lemma 5.2.
Acknowledgments
This research was partially supported by the National Science and Engineering Research
Council of Canada and the Network Centre of Excellence, Mathematics of Information
Technology and Complex Systems.
References
1 P. Jorion, Value at Risk: The New Benchmark for Managing Financial Risk, McGraw-Hill, New York, NY,
USA, 2nd edition, 2000.
2 S. Emmer, C. Kluppelberg, and R. Korn, Optimal portfolios with bounded capital at risk,
Mathematical Finance, vol. 11, no. 4, pp. 365384, 2001.
3 G. Dmitrasinovic-Vidovic, A. Lari-Lavassani, X. Li, and T. Ware, Dynamic Portfolio Selection Under
Capital at Risk, University of Calgary Yellow Series, University of Calgary, Alberta, Canada, 2003,
Report 833.
4 G. Dmitrasinovic-Vidovic, Portfolio selection under downside risk measures, Ph.D. thesis, University of
Calgary, Alberta, Canada, 2004.
5 C. Acerbi, Spectral measures of risk: a coherent representation of subjective risk aversion, Journal of
Banking and Finance, vol. 26, no. 7, pp. 15051518, 2002.
6 G. Dmitrasinovic-Vidovic and T. Ware, Asymptotic behaviour of mean-quantile ecient portfolios,
Finance and Stochastics, vol. 10, no. 4, pp. 529551, 2006.
7 P. Artzner, F. Delbaen, J.-M. Eber, and D. Heath, Coherent measures of risk, Mathematical Finance,
vol. 9, no. 3, pp. 203228, 1999.
8 C. Acerbi and D. Tasche, On the coherence of expected shortfall, Journal of Banking and Finance, vol.
26, no. 7, pp. 14871503, 2002.
9 H. Follmer and A. Schied, Convex measures of risk and trading constraints, Finance and Stochastics,
vol. 6, no. 4, pp. 429447, 2002.
10 G. Pflug, Some remarks on the value-at-risk and the conditional value-at-risk, in Probabilistic
Constrained Optimization, vol. 49 of Methodology and Applications, Kluwer Academic Publishers,
Dordrecht, The Netherlands, 2000.
11 F. Riedel, Dynamic coherent risk measures, Stochastic Processes and Their Applications, vol. 112, no. 2,
pp. 185200, 2004.
12 T. Rockafellar and S. Uryasev, Conditional value at risk: optimization algorithms and applications,
Financial Engineering News, vol. 14, 2000.
13 D. Tasche and D. Bundesbank, Expected shortfall and beyond, in Statistical Data Analysis Based on
the L1-Norm and Related Methods, Statistics for Industry and Technology, Springer, Berlin, Germany,
2002.
14 S. Emmer, C. Kluppelberg, and R. Korn., Optimal Portfolios with Bounded Downside Risks, Working
Paper, Center of Mathematical Sciences, Munich University of Technology, Munich, Germany, 2000.
15 I. Karatzas and S. E. Shreve, Methods of Mathematical Finance, Applications of Mathematics, Springer,
New York, NY, USA, 1999.
16 M. S. Bazaraa and C. M. Shetty, Nonlinear Programming, John Wiley & Sons, New York, NY, USA, 1979.
17 A. Lari-Lavassani, A. A. Sadeghi, and T. Ware, Modeling and Implementing Mean Reverting Price
Processes in Energy Markets, Electronic Publications of the International Energy Credit Association,
2001, http://www.ieca.net/.
doi:10.1155/2010/805309
Research Article
Individual Property Risk Management
Michael S. Finke,1 Eric Belasco,2 and Sandra J. Huston1

1
Division of Personal Financial Planning, Texas Tech University, Lubbock, TX 79409, USA
2
Department of Agricultural and Applied Economics, Texas Tech University, Lubbock, TX 79409, USA
Correspondence should be addressed to Michael S. Finke, [email protected]
Received 2 October 2009; Accepted 21 January 2010
Copyright q 2010 Michael S. Finke et al. This is an open access article distributed under the
This paper reviews household property risk management and estimates normatively optimal
choice under theoretical assumptions. Although risk retention limits are common in the financial
planning industry, estimates of optimal risk retention that include both financial and human
wealth far exceed limits commonly recommended. Households appear to frame property losses
dierently from other wealth losses leading to wealth-reducing, excess risk transfer. Possible
theoretical explanations for excess sensitivity to loss are reviewed. Dierences between observed
and optimal risk management imply a large potential gain from improved choice.
1. Introduction
Property risk management, a fundamental aspect of individual financial planning, has
perhaps been subject to the least amount of rigor. While investment management draws
directly from a theoretical structure of modern portfolio theory, risk management often
involves only the identification of risk exposures and products available to eliminate these
exposures. A common method of ensuring consistency in choice among insurance products
is to retain all risks beneath a risk retention limit; however the practice literature oers little
insight into how much retention is appropriate. This paper uses expected utility theory to
estimate optimal risk retention limits for households given reasonable assumptions about risk
aversion, human and financial wealth, and cost of insurance. Estimated retention limits are
generally much larger than limits chosen by individuals or recommended by professionals.
This suggests that households are either overweighting losses in a manner consistent with
Kahneman and Tverskys 1 prospect theory or unaware of normatively ecient insurance
decision making.
Risky decision making involves consideration of the likelihood of expected outcomes
and the consequences of each outcome on expected well being. While the profit motive
of firms suggests a preference for risky decisions that have a positive net expected value,
households are willing to pay a greater premium to mitigate risk. More formally, consumers
make decisions that maximize expected utility U by sacrificing expected wealth to reduce
the variance of possible outcomes.
For example, a household can face m possible states of nature where the likelihood
of an event is generated from a Berj distribution where j is the likelihood of event
j occurring. Along with the probabilities associated with each state is the payout or return
R associated with each event. In this case, we can compute the expected payout associated
as
m
E Payout j Rj . 1.1
j1
If a household is risk neutral, they will make insurance decisions that maximize their
expected payout. Since the insurance product is costly, this assumption typically leads to
no new purchases of insurance. A more realistic scenario is one where households have some
aversion to risk. For example, if a household is asked whether they prefer an annual salary
of $50,000 with full certainty or either $30,000 with a 95% likelihood or $450,000 with a 5%
likelihood, more households are likely to take the certain salary even though the expected
income from the uncertain scenario is higher. This is an illustration of the Von Neumann-
Morgenstern utility function, which is generally assumed to be strictly concave with respect
to wealth W. If W were a random variable and utility uW is strictly concave, then
Jensens inequality results in the following relationship where
EUW < uEW. 1.2
This implies that when facing large positive payouts, the utility associated with expected
wealth, uEW, is greater than the expected utility associated wealth, EUW. The
dierence between these two points represents the welfare gain to the household as well
as the profit opportunity for the insurer. In other words, a household can achieve greater
utility when faced with uncertain outcomes that are less extreme and will be willing to give
up expected wealth in order to forego these potential losses. The amount that a household
is willing to pay to mitigate risk is dependent on their degree of risk aversion, which is
known empirically as the risk aversion parameter. Risk aversion parameters are embedded
into utility functions, where an individual with UW < 0 is risk averse, UW > 0 is risk
seeking, and UW 0 is risk neutral.
An actuarially fair premium rate is exactly equal to the expected loss or the product of
the expected loss and probability of loss:

rate Eloss loss E payout | loss . 1.3
Risk-averse agents have incentive to purchase insurance in order to mitigate their risk and
increase expected utility. Insurance firms are able to take on new risk due to their ability
to spread risks among a diversified insurance pool while taking advantage of their risk
neutrality. At the same time, consumers are willing to pay a fee for this protection. If insurance
firms charged exactly the actuarial fair premium rate, they would make zero profits if they
eectively diversified their insurance pool and have no moral hazard or adverse selection.
Insurance firms use load fees to pay for administrative costs and generate profits. The cost of
insurance to individuals is
cost 1 load Eloss, 1.4
where load is some percentage greater than zero and adds to the cost. A rational, risk-averse
agent might still have incentive to purchase this insurance product with a negative expected
payout.
Insurance is also subject to other costs that are included in the load. Employees need
to be hired to verify the claim that a negative state has occurred. Employees also need to
be hired to estimate the likelihood that an uncertain state will occur in order to calculate
the cost of the contingent claim actuaries. Additional employees need to be hired to sell
the contingent claims and manage the collection of fees charged for the claims. These costs,
incurred to provide insurance to households, impact appropriate use of contingent claims to
maximize welfare. They further increase how much wealth is sacrificed to decrease risk.
Household investment choice assumes that there is an optimal amount of portfolio
risk for each investor at the point where an additional unit of risk no longer provides greater
expected utility despite greater expected wealth. Insurance involves this same tradeo of
expected wealth for reduced risk. The next section focuses on calculating the point at which
an additional unit of insurance no longer provides an increase in expected utility.
2. Estimating Optimal Insurance

Estimating insurance needs is similar to computing an optimal investment portfolio. In a
simple model, we need only know the wealth of the household, its risk aversion, and the cost
of the contingent claim.
2.1. Human Wealth

Total household wealth consists of both net worth and human wealth discounted expected
future household earnings. A formula to estimate human wealth is provided by Ibbotson et
al. 2, where human capital HC is equal to the expected future earnings E from next year
until retirement, discounted each year by the discount rate r and a risk premium v:

n
Eht
HCx . 2.1
tx 1 1 r vtx
A good proxy for the discount rate r is the rate on taxable, low-risk corporate
bonds, since income streams are similar to a bond and fully taxable. The risk premium v
further discounts expected future income streams that may be more volatile. To illustrate the
importance of human wealth in a household portfolio, consider that a 25-year-old with an
income of $75,000 at a 6% discount rate has an estimated present value wealth of $1.1 million,
assuming no income growth and no volatility if the individual expects to work until age
65. Alternatively, a 55-year-old with an income of $100,000 has a human wealth of $640,000,
given the same assumptions. As we age, we transform our human wealth into income until it
is exhausted at retirement. Generally, the wealth of younger households will consist primarily
of human wealth.
Risks to net worth include investment risk and negative events which will reduce
the value of assets destruction of property, a lawsuit, etc.. Risks to human wealth include
uncertain events that decrease the expected value of future earnings, including disability,
death, illness, or a labor market shock. Some risks are insurable through contingent claims, for
example, disability and property destruction, while others must either be retained or insured
through the public sector.
2.2. Cost of the Contingent Claim

Investors are often induced to accept greater risk by the opportunity to realize greater
expected returns. Similarly, the decision to retain or to transfer risk is influenced by the cost of
the insurance product. More expensive insurance will provide a disincentive to transfer risk
and conversely a greater incentive to retain risk. Unfortunately, it is often dicult to estimate
with any degree of accuracy the actual cost of the contingent claim.
In general, insurance products with heavy sales, underwriting, claims, moral, and
morale hazard expenses will be more costly. These types of costs are endemic to many
property and casualty policies. Loss ratios, or claims paid to policyholders, range from
roughly 55% for homeowners insurance to 60% for automobile insurance to 75% for
group health insurance. Based on 2007 national insurance loss ratio statistics provided by
the Missouri Department of Insurance, Financial Institutions & Professional Registration,
available at http://insurance.mo.gov/reports/lossratio/.
A loss ratio of 60% for property insurance policies implies that the expected return on
the premium paid for the policies is 40%. Of every dollar paid for an insurance premium
the household can expect to receive 60 cents back in claims on average. A $2,000 autopolicy
will thus yield an expected annual loss in wealth of $800. This is the cost of risk transfer.
However, the policy prevents a wealth loss if a negative state occurs. By choosing to buy
the insurance policy the individual is revealing that the expected utility from preventing the
uncertain negative state is greater than the certain loss of $800 in expected wealth, assuming
full information.
It is also important to note that the ownership of any asset that may decline in
value due to a negative state peril implies an expected annual loss in wealth that is a
function of the magnitude and likelihood of this loss. It is the cost of insurance 1-loss ratio
that represents the loss in wealth above the expected loss inherent in asset ownership. For
example, the ownership of a $100,000 home with a 1 in 100 chance of complete loss from a
fire involves an implied cost of $1,000 per year in risk on average. A policy with a loss ratio of
50% would cost $2,000 expected payout/loss ratio implying an additional expected wealth
loss of $1,000 per year.
2.3. Calculation of Insurance Needs

The estimate of insurance needs analysis relies on the following assumptions.
1 Life is risky. It is possible that negative states may occur that will reduce wealth.
2 Individuals are risk averse and are willing to reduce expected wealth in order to
avoid a risk.
3 Insurance reduces expected wealth to the extent that premiums exceed expected
payouts.
4 Risk aversion and wealth determine optimal risk retention and transfer.
Optimal risk retention will occur at the point where the expected utility from retaining
an additional dollar of possible loss is equal to the expected utility of transferring risk to
prevent the loss. According to Hanna 3, if we assume that this is a pure risk with two
potential states loss or no loss, the decision to insure will involve a comparison of utility
over four possible outcomes. The amount of utility gained or lost is a function of wealth and
risk aversion.
Suppose that individuals are assumed to exhibit constant relative risk aversion
CRRA which can be written as
W 1r
U , 2.2
1r
where r is the coecient of relative risk aversion. While CRRA utility functions are
commonly used, decreasing relative risk aversion DRRA can also be assumed where the
relative risk coecient decreases for higher levels of wealth. Because a DRRA utility function
assumes that individuals are relatively less risk averse at higher levels of wealth, the optimal
premium rate may be lower relative to CRRA as individuals are more willing to take on
additional relative risk for higher levels of wealth. Other utility functions include more
flexible forms as well as a class of absolute relative risk aversion functions CARA, IARA, and
DARA. Notice that the Arrow-Pratt coecient r of relative risk aversion has the following
relationship in this scenario:

U w U x
r x r 2.3
U w x U
implying that the negative of the elasticity of utility with respect to wealth is constant. Also,
notice that a higher r implies greater risk aversion and an r closer to 1 implies greater risk
tolerance. A household with a coecient of relative risk aversion equal to 1 is indierent
between a 50/50 chance of a 100% increase in expected lifetime income or a 50% reduction
in income. A coecient of 4 implies indierence between a gamble whose outcome is either
a 100% increase in income or a 20% decline in income. Empirical estimates from Kaplow 4
suggest that most fall near a relative risk aversion coecient of 3 to 5.
As an illustration, if we assume a wealth of W and assume ones house is worth H
where is the probability of fire damage destroying the house where 0, 1, the choice to
insure requires a premium payment of and a deductible payment d. The loss ratio, which
is the expected loss in wealth to the agent relative to the insurance premium rate, can then be
expressed as the ratio of expected losses over the premium rate:

1 H
lr . 2.4

For example, in the case where W $250, 000 and H $100, 000 where the cost of
insurance was $2, 000 and d $0 with the likelihood of fire equal to 1% the resulting
loss ratio would be 50%.
If we compare only the expected wealth of each choice, given probabilities insurance
will never be the optimal choice since by definition it will require a decline in expected wealth
to be economically viable. To estimate the optimal choice of risk-averse investors dollar values
must be transformed using a utility function that incorporates the degree of relative risk
aversion.
In the case of insurance, two dierent wealth levels can be attained dependent upon
two possible states fire or no fire. In the case of fire, wealth WIF is equal to initial
wealth W0 minus the insurance premium and deductible d. In the case of no fire, no
deductible is paid so that WINF W0 . In the expected utility framework, we can express
expected utility as

EUI uWIF 1 uWINF
2.5
W0 d1r W0 1r
1 .
1r 1r
If we assume r 3, then we obtain EUI 8.14e 12 units of utility. Alternatively, if

the agent does not purchase insurance, WNIF W0 H and WNINF W0 . In the case of no
insurance, the expected utility EUNI is equal to 8.13e 12.
Because utility is an ordinal measure, we can conclude that even with a 50% loss ratio
an individual with the specified preference function is better o purchasing insurance. The
main reason is that the loss represents a very large share of wealthin this case a loss of
40% of wealth. With this magnitude of loss an individual would only consider not buying
insurance if the loss ratio fell beneath 45%. In the above example, if the initial wealth had
instead been $500,000, the maximum potential loss represents 20% of wealth. At 20% of
wealth a household would be willing to buy insurance if the loss ratio is 60% or more.
The optimal retention limit is found by computing the point at which expected utility
from retaining risk is exactly equal to the utility from transferring risk through insurance. At
this point, is set to the point where EUI EUNI . This can be rewritten as

EUI W d1r 1 W 1r

W H1r 1 W d1r 2.6
EUNI .
Even though an explicit analytical solution could not be determined for fW, d, r, , H,
the solution can be found numerically. With this example, the premium rate that solves the
above equation is 2, 193, which results in a loss ratio of 53.4% when the risk retention ratio
is 20%. A greater loss ratio or expected insurance claims from each premium dollar will
encourage greater transfer of risk as the cost of insurance is lower. More expensive insurance
will encourage risk retention. A larger or catastrophic potential loss will lead to a greater
willingness to insure. If the potential loss is small in terms of wealth, then the cost of insurance
would need to be very low to induce a household to buy insurance.
Figure 1 illustrates the optimal tradeo between risk transfer and risk retention given
loss as a percentage of wealth and loss ratios. As the loss ratio declines, insurance becomes
more expensive and individuals are only willing to insure if the loss represents a high percent
transfer as a percentage of net worth %

25
20
Minimum risk to
15
10
0
5 10 15 20 25 30 35 40 45 50
1-loss ratio %
Tolerant
Averse
Figure 1: Optimal Risk Retention Limits.
of wealth. Risk tolerant individuals are assigned a coecient of relative risk aversion of 3 and
the risk-averse individuals are assigned a relative risk aversion of 5.
At a 40% loss ratio, characteristic of many property policies, it is only rational for
individuals to insure losses that are about 1/7 of total wealth if they are risk averse. For the
more risk tolerant an optimal risk retention limit is closer to 1/5 of total wealth. In practical
terms these risk retention limits are enormous. A risk-averse household with a human wealth
and wealth of $2 million would retain all risks below 10% of total wealth if the policy paid
back 75% of premiums. With wealth this large products like comprehensive or collision
insurance on a vehicle would have no appeal. It would also not make sense to carry anything
other than the highest possible deductibles or to insure any assets whose value falls beneath
$200,000.
While the conclusions of the model may appear extreme, they are valid in the sense
that they are consistent with less intuitively extreme investment decision making. For
example, retaining risk on property insurance either through very high deductibles or by not
buying insurance provides an expected yield equal to the opportunity cost of insurance
which in the case of most property insurance is equivalent to a yield of at least 40%. The
downside is that the household may subject themselves to a loss of, in this case, up to $200,000
to earn this return. However, expected returns on equities in the U.S. have been roughly 10%
12% and the tradeo is exactly the samethe possibility of a large loss in wealth. In terms of
wealth, there is no dierence between a $100,000 loss due to property or casualty loss and a
$100,000 loss in an investment portfolio. Both losses were the result of risk borne to maximize
expected utility given uncertain outcomes.
Prudent risk management must also acknowledge the limitations of including human
wealth when estimating risk retention limits. A wealth shock that falls beneath the estimated
risk retention limit, for example, the loss of a $35,000 car to a 25-year-old, will be devastating
if it wipes out liquid savings and if the ability to borrow against human capital is limited.
Credit constraints in the face of a large loss can lead to a significant drop in consumption
and a loss to illiquid projects that require a constant stream of cash to maintain such as
mortgages, business expenses, and student loan payments. Liquidity and access to credit
are important considerations that impact optimal risk retention for those with high human
wealth and few financial assets.
These results suggest that many are approaching the process of risk management by
focusing on identifiable losses without recognizing the tradeo of risk and return when
choosing optimal risk management strategies. Most households are spending too much to
prevent property and casualty risks while simultaneously retaining risk in their investment
portfolio. This is neither wealth nor utility maximizing. However, the authors recognize that
implementing this model will be dicult since many households are not prepared to retain
risk they are accustomed to transferring.
3. Overweighting Losses
While the expected utility framework is rational in that it assumes disutility from wealth
changes to be equal to the reduction in expected consumption, individuals appear to weigh
gains and losses from risky choices dierently. In fact, the persistent popularity of insurance
products that protect against small losses suggests that individuals are willing to pay
dearly to protect against minor losses while simultaneously paying insucient attention
to much larger risks Kunreuther and Pauly 5. Using results from experimental data,
Kahneman and Tversky 1 point out three major limitations of expected utility theory which
include the consistent overestimation of low probability events and underestimation of high
probabilities. This would suggest that agents generally overinsure against rare events and
underinsure against more common events, according to expected utility theory. The second
finding is that utility functions commonly referred to as value functions in the Prospect
Theory literature are generally concave concerning gains and convex for losses Tversky
and Kahneman 6. Convexity for losses implies a large amount of disutility for relatively
small losses and only a modest increase in disutility for larger losses. An individual with a
prospect theory value function will place greater emphasis on avoiding small losses and a
reduced emphasis on large losses than if they followed a conventional utility function.
The third problem with expected utility theory is that of absolute losses versus relative
losses. As indicated by Kahneman and Tversky 1, individuals in experimental settings are
shown to make decisions based on changes in their financial position rather than the impact
on their final wealth. Kunreuther and Pauly 5 show that if an asset is only monetary, then it
is rational to assess values based on absolute wealth changes. Instead, individuals appear to
consider their current wealth a reference point and any loss from that reference point induces
greater disutility than the dollar value of the loss would suggest.
Modeling a risky decision using prospect theory involves a reference point R from
which gains and losses are assessed. The most relevant reference point is initial wealth R
W0 . There is a small net loss associated with paying insurance premiums; however, there
is an even larger net loss in the case of a fire under no insurance. If a fire occurs F, there
is a small net loss with and a large net loss without insurance; however, there is a no loss if
insurance is not purchased and no fire occurs NF. We define Xi Wi R, where i {F, NF},
as the net gains/losses. Because the utility function is assumed to be asymmetric around R,
we then define the two parts of the value function which can be written as

0 if Xi 0,

vXi d if Xi < 0 with probability , 3.1

if Xi < 0 with probability 1 ,
where is the loss aversion parameter and is greater than 1. The value function under the
purchase of insurance VI becomes

VI g d g 1 3.2
such that g is a weighting function that accounts for overvaluing small probabilities and
undervaluing large probabilities. The same reference point is used when we consider the case
where no insurance is purchased. In this situation a fire results in a net loss and no fire results
in a no loss. In this scenario we obtain

0 if Xi 0,
vXi 3.3
H if Xi < 0,
which then converts the value function VNI to be

VNI g H . 3.4
To apply this function to the fire example above we assume 0.88 and 2.25 as suggested
by Tversky and Kahneman 7. To assess subjective probability biases we use the function
derived by al-Nowaihi and Dhami 8 who derive their function from a more general form
from Prelec 9, which is essentially

g e ln , 3.5
where we assume 0.80. Using this function allows us to inflate our probability of 1%
to 5% and deflate the probability of 99% to 98%. Given these adjusted probabilities and the
downside risk parameter, we now look for the premium amount that solves for VI VNI .
At some point the premium will be high enough to outweigh the ability of the individual to
manage the downside risk of a fire. For this scenario 3, 255. Since both value functions
are not functionally related to W, the optimal risk retention limit is unaected by W0 . This
diers from expected utility theory which assumes increasing risk tolerance with wealth from
a potential loss of a given magnitude.
While prospect theory is useful as a means of understanding how individuals actually
behave when faced with a decision to retain or transfer risk, it has little use as a normative
tool to improve risk management practices. Estimates of optimal insurance using a prospect
theory value function are high for small risks and low for more catastrophic risks. This
is consistent with the current market for consumer insurance with its broad overuse of
low deductibles, extended warranties, and protection of low-value tangible goods and
simultaneous underuse of insurance products protecting more catastrophic risks of liability
and loss of earnings. However, spending heavily to avoid small property and casualty risks
while maintaining an optimal investment portfolio that requires acceptance of market risk
involving potential losses of a much greater magnitude results in a wealth loss from framing
these decisions separately. Prospect theory thus lends itself to modeling positive behavior but
fails to guide practitioners or individuals in risk management.
Another possible shortcoming of Prospect Theory is that an insurance premium is

itself a small-stakes loss. Payment of a premium assumes that an individual is willing to
accept a small loss to avoid a possible larger loss; however, if the utility function is convex in
losses individuals are risk seeking when presented a choice that involves small and larger
losses, then loss aversion becomes an even less plausible explanation unless the premium is
not seen as a loss in itself. Such an individual may be induced into paying a relatively high
premium to avoid a small-stakes loss since the loss from the random event creates greater
disutility per dollar than the certain loss from the insurance premium. This appears to be
the case in decisions involving the choice of homeowners insurance deductibles. The choice
between a $250 and a $500 deductible implies the payment of an added premium say $25 to
avoid an unlikely loss of the $250 dierence between the two deductibles. For an individual
with a total wealth of $100,000 e.g., a destitute 80-year-old whose wealth consists of the
present value of social security, choosing a $250 deductible implies a coecient of relative
risk aversion of over 500 Sydnor 10. Mehra and Prescott 11 describe the equity risk
premium as a puzzle since it implies a coecient of relative risk aversion of around 30. If
the equity premium is a puzzle, then the risk retention premium is a mystery of inexplicable
magnitude.
The only reasonable explanation for observed property insurance behavior is that
the unexpected loss covered by insurance provides so much disutility that an individual is
willing to give up large amounts of expected wealth to avoid it. This would be plausible if
the losses caused painful regret when a peril caused a loss and if the premium itself was not
viewed as a loss. Braun and Muermann 12 incorporate the unhappiness that is caused by
regret into a utility function that can explain observed demand for insurance. Sydnor 10
modifies a utility function developed by Koszegi and Rabin 13 in which wealth losses are
amplified by a factor that represents the relative pain felt by a random loss relative to a gain in
wealth. The model suggests very strong amplification for low-stakes losses and more realistic
amplification for large-scale loss, which appears consistent with observed behavior.
4. Conclusions
Proper household management of property and casualty risk requires an assessment of
the dollar values of losses in various possible future states. When wealth in each state is
transformed based on an individuals risk tolerance, it is possible to estimate the level of
risk transfer through insurance. Using reasonable estimates of risk aversion, cost of property
insurance, and initial wealth, optimal risk retention can be as high as 20% of initial household
wealth. A risk retention limit of this magnitude would imply far higher deductibles in
insurance policies and the abandonment of many popular policies that protect small losses.
If human wealth is considered a component of total wealth, many young individuals would
avoid insuring against all but catastrophic losses.
Advances in portfolio management and dissemination of normative investment
science have led to broad acceptance of investment risk among households. For example,
the percentage of U.S. households owning stock increased from 32% in 1989 to 52%
in 2001 14. This increased acceptance of potential loss in investment portfolios has
resulted in a significant improvement in household welfare. Holding a portfolio that is
consistent with risk preferences implies an increase in expected utility relative to one that
is excessively conservative. Similarly, dissemination of normative risk management science
has the potential to improve welfare by illustrating the potential benefits of reducing costly
protection of small losses and increasing protection against catastrophic risks. This study
provides estimates of a large potential gain from increased acceptance of certain risks that are
costly to insure.
The evidence from household risk retention preference through home insurance
deductibles suggests that the market for property insurance reflects strong preferences for
loss aversion. Benartzi and Thaler 15 provide evidence that Prospect Theory may also
explain the observed high equity premium and an unwillingness to own risky assets.
However, a clear welfare loss results from simultaneous ownership of risky securities and
policy protection against small-scale risks. There is also evidence that individuals place
insucient weight on the utility loss from random large losses to total wealth by failing to
insure adequately for potentially large losses such as large-scale liability losses and losses
to human wealth e.g., through disability income insurance. In the case of insurance, where
there is little advice available including from financial professionals to maintain consistency
among risky financial decisions, behavior may not accurately reveal preferences. If this is the
case, then the application of a standard expected utility model may provide normative value
that can help guide individuals and advisors toward making better decisions Campbell
16.
References
1 D. Kahneman and A. Tversky, Prospect theory: an analysis of decision under risk, Econometrica, vol.
47, no. 2, pp. 263291, 1979.
2 R. G. Ibbotson, M. A. Milevsky, P. Chen, and K. X. Zhu, Lifetime financial advice: human capital,
asset allocation, and insurance, The Research Foundation of CFA Institute, 2007.
3 S. Hanna, Risk aversion and optimal insurance deductibles, in Proceedings of the 31th Annual
Conference on American Council on Consumer Interests (ACCI 89), pp. 141147, Columbia, Mo, USA,
1989.
4 L. Kaplow, The value of a statistical life and the coecient of relative risk aversion, Journal of Risk
and Uncertainty, vol. 31, no. 1, pp. 2334, 2005.
5 H. Kunreuther and M. Pauly, Insurance decision-making and market behavior, Foundations and
Trends in Microeconometrcs, vol. 1, no. 2, pp. 63127, 2005.
6 A. Tversky and D. Kahneman, Loss aversion in riskless choice: a reference dependent model,
Quarterly Journal of Economics, vol. 106, pp. 10391062, 1991.
8 A. al-Nowaihi and S. Dhami, A simple derivation of prelecs probability weighting func-
tion, Working Paper No. 05/20, University of Leicester, Department of Economics, 2005,
http://www.le.ac.uk/economics/research/RePEc/lec/leecon/dp05-20.pdf.
9 D. Prelec, The probability weighting function, Econometrica, vol. 66, no. 3, pp. 497527, 1998.
10 J. Sydnor, Overinsuring modest risks, 2009, http://wsomfaculty.case.edu/sydnor/deductibles
.pdf.
11 R. Mehra and E. C. Prescott, The equity premium: a puzzle, Journal of Monetary Economics, vol. 15,
no. 2, pp. 145161, 1985.
12 M. Braun and A. Muermann, The impact of regret on the demand for insurance, The Journal of Risk
and Insurance, vol. 71, no. 4, pp. 737767, 2004.
13 B. Koszegi and M. Rabin, Reference-dependent risk attitudes, American Economic Review, vol. 97,
no. 4, pp. 10471073, 2007.
14 E. Wol, Changes in Household Wealth in the 1980s and 1990s in the U.S., in International
Perspectives on Household Wealth, Elgar, 2004.
15 S. Benartzi and R. Thaler, Myopic loss aversion and the equity premium puzzle, The Quarterly
Journal of Economics, vol. 110, no. 1, pp. 7392, 1995.
16 J. Campbell, Household finance, Journal of Finance, vol. 61, no. 4, pp. 15531604, 2006.
doi:10.1155/2010/357321
Research Article
On Some Layer-Based Risk Measures with
Applications to Exponential Dispersion Models
Olga Furman1 and Edward Furman2

1
Actuarial Research Center, University of Haifa, Haifa 31905, Israel
2
Department of Mathematics and Statistics, York University, Toronto, ON, Canada M3J 1P3
Correspondence should be addressed to Edward Furman, [email protected]
Received 13 October 2009; Revised 21 March 2010; Accepted 10 April 2010
Academic Editor: Johanna Neslehova
Copyright q 2010 O. Furman and E. Furman. This is an open access article distributed under
the Creative Commons Attribution License, which permits unrestricted use, distribution, and
Layer-based counterparts of a number of well-known risk measures have been proposed and
studied. Namely, some motivations and elementary properties have been discussed, and the
analytic tractability has been demonstrated by developing closed-form expressions in the general
framework of exponential dispersion models.
1. Introduction
Denote by X the set of actuarial risks, and let 0 X X be a random variable rv with
cumulative distribution function cdf Fx, decumulative distribution function ddf Fx
1 Fx, and probability density function pdf fx. The functional H : X 0, is
then referred to as a risk measure, and it is interpreted as the measure of risk inherent in
X. Naturally, a quite significant number of risk measuring functionals have been proposed
and studied, starting with the arguably oldest Value-at-Risk or VaR cf. 1, and up to the
distorted cf. 25 and weighted cf. 6, 7 classes of risk measures.
More specifically, the Value-at-Risk risk measure is formulated, for every 0 < q < 1, as

VaRq X inf x : FX x q , 1.1
which thus refers to the well-studied notion of the qth quantile. Then the family of distorted
risk measures is defined with the help of an increasing and concave function g : 0, 1
0, 1, such that g0 0 and g1 1, as the following Choquet integral:

Hg X g Fx dx. 1.2
R
Last but not least, for an increasing nonnegative function w : 0, 0, and the so-
called weighted ddf F w x E1{X > x}wX/EwX the class of weighted risk measures
is given by

Hw X F w xdx. 1.3
R
Note that for at least once dierentiable distortion function, we have that the weighted class
contains the distorted one as a special case, that is, Hg X EXg FX is a weighted risk
measure with a dependent on F weight function.
Interestingly, probably in the view of the latter economic developments, the so-
called tail events have been drawing increasing attention of insurance and general finance
experts. Naturally therefore, tail-based risk measures have become quite popular, with the tail
conditional expectation TCE risk measure being a quite remarkable example. For 0 < q < 1
and thus FVaRq X / 0, the TCE risk measure is formulated as

1
TCEq X x dFx. 1.4
F VaRq X VaRq X
Importantly, the TCE belongs in the class of distorted risk measures with the distortion
function
x
gx 1 x <1q 1 x 1q , 1.5
1q
where 1 denotes the indicator function cf., e.g., 8, as well as in the class of weighted risk
measures with the weight function

wx 1 x VaRq X 1.6
cf., e.g., 6, 7. The TCE risk measure is often referred to as the expected shortfall ES and
the conditional Value-at-Risk CVaR when the pdf of X is continuous cf., e.g., 9.
Functional 1.4 can be considered a tail-based extension of the net premium HX
EX. Furman and Landsman 10 introduced and studied a tail-based counterpart of the
standard deviation premium calculation principle, which, for 0 < q < 1, the tail variance

TVq X Var X | X > VaRq X , 1.7
and a constant 0, is defined as
TSDq X TCEq X TV1/2

q X. 1.8
For a discussion of various properties of the TSD risk measure, we refer to Furman and
Landsman 10. We note in passing that for q 0, we have that TSDq X SDX
EX Var1/2 X.
The rest of the paper is organized as follows. In the next section we introduce and
motivate layer-based extensions of functionals 1.4 and 1.8. Then in Sections 3 and 4 we
analyze the aforementioned layer-based risk measures as well as their limiting cases in the
general context of the exponential dispersion models EDMs, that are to this end briefly
reviewed in the appendix. Section 5 concludes the paper.
2. The Limited TCE and TSD Risk Measures

Let 0 < q < p < 1 and let X X have a continuous and strictly increasing cdf. In many
practical situations the degree of riskiness of the layer VaRq X, VaRp X of an insurance
contract is to be measured certainly the layer width VaRp X VaRq X q,p > 0.
Indeed, the number of deductibles in a policy is often more than one, and/or there can
be several reinsurance companies covering the same insured object. Also, there is the so-
called limited capacity within the insurance industry to absorb losses resulting from, for
example, terrorist attacks and catastrophes. In the context of the aforementioned events, the
unpredictable nature of the threat and the size of the losses make it unlikely that the insurance
industry can add enough capacity to cover them. In these and other cases neither 1.4 nor
1.8 can be applied since 1 both TCE and TSD are defined for one threshold, only, and
2 the aforementioned pair of risk measures is useless when, say, the expectations of the
underlying risks are infinite, which can definitely be the case in the situations mentioned
above.
Note 1. As noticed by a referee, the risk measure H : X 0, is often used to price
insurance contracts. Naturally therefore, the limited TCE and TSD proposed and studied
herein can serve as pricing functionals for policies with coverage modifications, such as, for
example, policies with deductibles, retention levels, and so forth cf., 11, Chapter 8.
Next, we formally define the risk measures of interest.
Definition 2.1. Let xq VaRq X and xp VaRp X, for 0 < q < p < 1. Then the limited TCE
and TSD risk measures are formulated as
LTCEq,p X E X | xq < X xp , 2.1
and
LTSDq,p X E X | xq < X xp Var1/2 X | xq < X xp , 2.2
respectively.
Clearly, the TCE and TSD are particular cases of their limited counterparts. We note in
passing that the former pair of risk measures need not be finite for heavy tailed distributions,
and they are thus not applicable. In this respect, limited variants 2.1 and 2.2 can provide
a partial resolution. Indeed, for k 1, 2, . . . , we have that

F xp E X k | X xp F xq E X k | X xq
E X | xq < X xp
k
< , 2.3
F xp F xq
regardless of the distribution of X.

We further enumerate some properties of the LTSD risk measure, which is our main
object of study.
1 Translation Invariance. For any constant c 0, we have that
LTSDq,p X c LTSDq,p X c. 2.4
2 Positive Homogeneity. For any constant d > 0, we have that
LTSDq,p d X d LTSDq,p X. 2.5
3 Layer Parity. We call X X and Y X layer equivalent if for some 0 < q < p < 1,
such that xq yq , xp yp , and for every pair {t1 , t2 : q < t1 < t2 < p}, it holds that
Pxt1 < X xt2 Pyt1 < Y yt2 . In such a case, we have that
LTSDt1 ,t2 X LTSDt1 ,t2 Y . 2.6
Literally, this property states that the LTSD risk measure for an arbitrary layer is
only dependent on the cdf of that layer. Parity of the ddfs implies equality of LTSDs.
Although looking for original ways to assess the degree of actuarial riskiness is a
very important task, subsequent applications of various theoretical approaches to a real-
world data are not less essential. A significant number of papers have been devoted to
deriving explicit formulas for some tail-based risk measures in the context of various loss
distributions. The incomplete list of works discussing the TCE risk measure consists of, for
example, Hurlimann 12 and Furman and Landsman 13, gamma distributions; Panjer 14,
normal family; Landsman and Valdez 15, elliptical distributions; Landsman and Valdez
16, and Furman and Landsman 17, exponential dispersion models; and Chiragiev and
Landsman 18, Vernic 19, Asimit et al. 20, Pareto distributions of the second kind.
As we have already noticed, the unlimited tail standard deviation risk measure
has been studied in the framework of the elliptical distributions by Furman and Landsman
10. Unfortunately, all members of the elliptical class are symmetric, while insurance risks
are generally modeled by nonnegative and positively skewed random variables. These
peculiarities can be fairly well addressed employing an alternative class of distribution
laws. The exponential dispersion models include many well-known distributions such as
normal, gamma, and inverse Gaussian, which, except for the normal, are nonsymmetric,
have nonnegative supports, and can serve as adequate models for describing insurance risks
behavior. In this paper we therefore find it appropriate to apply both TSD and LTSD to EDM
distributed risks.
3. The Limited Tail Standard Deviation Risk Measure for

Exponential Dispersion Models
An early development of the exponential dispersion models is often attributed to Tweedie
21, however a more substantial and systematic investigation of this class of distributions
was documented by Jrgensen 22, 23. In his Theory of dispersion models, Jrgensen 24 writes
that the main raison detre for the dispersion models is to serve as error distributions for
generalized linear models, introduced by Nelder and Wedderburn 25. Nowadays, EDMs
play a prominent role in actuarial science and financial mathematics. This can be explained
by the high level of generality that they enable in the context of statistical inference for
widely popular distribution functions, such as normal, gamma, inverse Gaussian, stable, and
many others. The specificity characterizing statistical modeling of actuarial subjects is that
the underlying distributions mostly have nonnegative support, and many EDM members
possess this important phenomenon, for a formal definition of the EDMs, as well as for a
brief review of some technical facts used in the sequel, cf., the appendix.
We are now in a position to evaluate the limited TSD risk measure in the framework
of the EDMs. Recall that, for 0 < q < p < 1, we denote by xq , xp an arbitrary layer having
attachment point xq and width q,p . Also, let

h xq , xp ; , log F xp ; , F xq ; , 3.1

denote the generalized layer-based hazard function, such that

h xq , x1 ; , log F xq ; , h xq ; , ,

3.2

h x0 , xp ; , log F xp ; , h xp ; , ,

and thus

F xq ; ,
h xq , xp ; , h xq ; ,
F xq ; , F xp ; ,
3.3

F xp ; ,
h xp ; , .
F xq ; , F xp ; ,
The next theorem derives expressions for the limited TCE risk measure, which is a
natural precursor to deriving the limited TSD.
Theorem 3.1. Assume that the natural exponential family (NEF which generates EDM is regular
or at least steep cf. 24, page 48. Then the limited TCE risk measure
i for the reproductive EDM Y ED, 2 is given by

LTCEq,p Y 2 h xq , xp ; , 3.4
and
ii for the additive EDM X ED , is given by

LTCEq,p X h xq , xp ; , . 3.5
Proof. We prove the reproductive case only, since the additive case follows in a similar
fashion. By the definition of the limited TCE, we have that

F yq E Y | Y > yq F yp E Y | Y > yp
LTCEq,p Y . 3.6
F yp F yq
Further, following Landsman and Valdez 16, it can be shown that for every 0 < q < 1, we
have that

E Y | Y > yq 2 h yq ; , , 3.7
which then, employing 3.1 and 3.3, yields

F yq ; , 2 h yq ; , F yp ; , 2 h yp ; ,
LTCEq,p Y
F yq ; , F yp ; , 3.8

2 h yq , yp ; ,
and hence completes the proof.
In the sequel, we sometimes write LTCEq,p Y ; , in order to emphasize the

dependence on and .
Note 2. To obtain the results of Landsman and Valdez 16, we put p 1, and then, for
instance, in the reproductive case, we end up with

lim LTCEq,p Y 2 h yq ; , TCEq Y , 3.9
p1
subject to the existence of the limit.
Next theorem provides explicit expressions for the limited TSD risk measure for both
reproductive and additive EDMs.
Theorem 3.2. Assume that the NEF which generates EDM is regular or at least steep. Then the
limited TSD risk measure
i for the reproductive EDM Y ED, 2 is given by

3.10
LTSDq,p Y LTCEq,p Y 2 LTCEq,p Y ; ,

and
ii for the additive EDM X ED , is given by

3.11
LTSDq,p X LTCEq,p X LTCEq,p X; , .

Proof. We again prove the reproductive case, only. Note that it has been assumed that is
a dierentiable function, and thus we can dierentiate the following probability integral in
under the integral sign cf., the appendix:
yp

P yq < Y yp ey d y , 3.12
yq
and hence, using Definition 2.1, we have that

LTCEq,p Y ; , F yp ; , F yq ; ,

yp

yey d y
yq
yp 3.13

y2 ey y ey d y
yq

2 E Y 2 | 1 yq < Y yp E Y | 1 yq < Y yp ,
with the last line following from the appendix. Further, by simple rearrangement and
straightforward calculations, we obtain that
yp
y2 ey d y
yq
E Y 2 | yq < Y yp
F yp ; , F yq ; ,

/LTCEq,p Y ; , F yp ; , F yq ; ,
LTCEq,p Y 2

F yp ; , F yq ; , 3.14

2 LTCEq,p Y ; , LTCEq,p Y 2 log F yp ; , F yq ; ,

2
2 LTCEq,p Y ; , LTCEq,p Y ; , ,

which along with the definition of the limited TSD risk measure completes the proof.
We further consider two examples to elaborate on Theorem 3.2. We start with the
normal distribution, which occupies a central role in statistical theory, and its position in
statistical analysis of insurance problems is very dicult to underestimate, for example, due
to the law of large numbers.
Example 3.3. Let Y N, 2 be a normal random variable with mean and variance 2 ,
then we can write the pdf of Y as

1 1 y 2
f y exp
2 2
3.15

1 1 1 1 2
exp 2 y2 exp y , y R.
2 2 2 2
If we take and 1/ 2 , we see that the normal distribution is a reproductive EDM with
cumulant function 2 /2. Denote by and the pdf and the cdf, respectively, of
the standardized normal random variable. Then using Theorem 3.1, we obtain the following
expression for the limited TCE risk measure for the risk Y :

1 yq 1 yp
LTCEq,p Y 1 . 3.16
yp 1 yq
If we put p 1, then the latter equation reduces to the result of Landsman and Valdez 16.
Namely, we have that

1 yq
lim LTCEq,p Y TCEq Y . 3.17
p1 1 1 yq
Further, let zq yq / and zp yp /. Then
2
z z z z z zp
2 LTCEq,p Y ; , 2 1 .
q q p p q
3.18
zp zq zp zq
Consequently, the limited TSD risk measure is as follows:
LTSDq,p Y

2
zq zp 2 zq zq zp zp zq zp
1 .
zp zq zp zq zp zq
3.19
We proceed with the gamma distributions, which have been widely applied in various
fields of actuarial science. It should be noted that these distribution functions possess
positive support and positive skewness, which is important for modeling insurance losses. In
addition, gamma rvs have been well-studied, and they share many tractable mathematical
properties which facilitate their use. There are numerous examples of applying gamma
distributions for modeling insurance portfolios cf., e.g., 12, 13, 26, 27.
Example 3.4. Let X Ga, be a gamma rv with shape and rate parameters equal and ,
correspondingly. The pdf of X is
1 1
fx ex x1 x1 exp x log , x > 0. 3.20

Hence the gamma rv can be represented as an additive EDM with the following pdf:
1 1
fx x exp x log , 3.21

where x > 0 and < 0. The mean and variance of X are EX / and VarX /2 .
Also, , , and log. According to Theorem 3.1, the limited tail
conditional expectation is

F xp ; , 1 F xq ; , 1
LTCEq,p X . 3.22
F xp ; , F xq ; ,
Putting p 1 we obtain that

F xp ; , 1 F xq ; , 1 F xq ; , 1
lim TCEq X, 3.23
p1 F xp ; , F xq ; , F xq ; ,
which coincides with 13, page 643. To derive an expression for the limited TSD risk measure,
we use Theorem 3.2, that is,

F xp ; , 1 F xq ; , 1
LTCEq,p X; ,
F xp ; , F xq ; ,

F xp ; , 1 F xq ; , 1
2 3.24
F xp ; , F xq ; ,

F xp ; , 1 F xq ; , 1
.
F xp ; , F xq ; ,
Further, since for n 1, 2, . . . ,

F xp ; , n F xq ; , n

xp
1
xn1 exp x n log dx
xq n
xp 3.25
n
fx; , n x dx
xq
x xp
n p
fx; , n 1dx fx; , ndx ,

xq xq
the limited TSD risk measure for gamma is given by
LTSDq,p X

F xp ; , 1 F xq ; , 1

F xp ; , F xq ; ,

2
F xp ; , 2 F xq ; , 2 Fxp ; , 1 Fxq ; , 1
2 1 .
F xp ; , F xq ; , Fxp ; , Fxq ; ,
3.26
In the sequel, we consider gamma and normal risks with equal means and variances,
and we explore them on the interval t, 350, with 50 < t < 350. Figure 1 depicts the results.
Note that both LTCE and LTSD imply that the normal distribution is riskier than gamma for
lower attachment points and vice-versa, that is quite natural bearing in mind the tail behavior
of the two.
Although the EDMs are of pivotal importance in actuarial mathematics, they fail to
appropriately describe heavy-tailed insurance losses. To elucidate on the applicability of
the layer-based risk measures in the context of the probability distributions possessing heavy
tails, we conclude this section with a simple example.
350 360
350
300
340
Limited TSD
Limited TCE
330
250
320
310
200
300
150 290
50 100 150 200 250 300 350 50 100 150 200 250 300 350
Attachment point Attachment point
Gamma Gamma
Normal Normal
a b
Figure 1: LTCE and LTSD for normal and gamma risks with means 150 and standard deviations 100, alpha
2.
Example 3.5. Let X Pa, be a Pareto rv with the pdf

fx , x > > 0, 3.27
x1
and > 0. Certainly, the Pareto rv is not a member of the EDMs, though it belongs to the
log-exponential family LEF of distributions cf. 7. The LEF is defined by the dierential
equation

Fdx; , exp logx dx, 3.28

where is a parameter, is a measure, and log 0 x dx is a normalizing constant
the parameters should not be confused with the ones used in the context of the EDMs. Then
X is easily seen to belong in LEF with the help of the reparameterization dx x1 dx, and
.
In this context, it is straightforward to see that EX is infinite for 1, which thus
implies infiniteness of the TCE risk measure. We can however readily obtain the limited
variant as follows:

xp 1 1
1 xp xq xp xq
LTCEq,p X
dx , 3.29
P xq < X xp xq x 1
xp xq
1 1
that is finite for any > 0. Also, since, for example, for < 1, we have that xp xq < 0, the
limited TCE risk measure is positive, as expected. The same is true for 1.
We note in passing that, for > 1 and p 1 and thus xp , we have that

1 1
xp xq xp xq
TCEq X lim xq , 3.30
p1 1 1

xp xq
which confirms the corresponding expression in Furman and Landsman 8.
Except for the Pareto distribution, the LEF consists of, for example, the log-normal and
inverse-gamma distributions, for which expressions similar to 3.29 can be developed in the
context of the limited TCE and limited TSD risk measures, thus providing a partial solution
to the heavy-tailness phenomenon.
4. The Tail Standard Deviation Risk Measure for

Exponential Dispersion Models
The tail standard deviation risk measure was proposed in 10 as a possible quantifier of
the so-called tail riskiness of the loss distribution. The above-mentioned authors applied
this risk measure to elliptical class of distributions, which consists of such well-known pdfs
as normal and student-t. Although the elliptical family is very useful in finance, insurance
industry imposes its own restrictions. More specifically, insurance claims are always positive
and mostly positively skewed. In this section we apply the TSD risk measure to EDMs.
The following corollary develops formulas for the TSD risk measure both in the
reproductive and additive EDMs cases. Recall that we denote the ddf of say X by F; , to
emphasize the parameters and , and we assume that
lim LTSDq,p X < . 4.1

p1
The proof of the next corollary is left to the reader.
Corollary 4.1. Under the conditions in Theorem 3.1, the tail standard deviation risk measure is

4.2
TSDq Y TCEq Y 2 TCEq Y ; ,

in the context of the reproductive EDMs, and

4.3
TSDq X TCEq X TCEq X; ,

in the context of the additive EDMs.
We further explore the TSD risk measure in some particular cases of EDMs, which
seem to be of practical importance.
Example 4.2. Let Y N, 2 be again some normal rv with mean and variance 2 . Then
we easily evaluate the TSD risk measure using Corollary 4.1 and Example 3.3 as follows:

2
zq 2 zq zq
TSDq X 1 zq , 4.4
1 zq 1 zq 1 zq
which coincides with 10.
Example 4.3. Let X Ga, be a gamma rv with shape and scale parameters equal and ,
correspondingly. Taking into account Example 3.4 and Corollary 4.1 leads to
TSDq X

2
F xq ; , 1 F xq ; , 2 Fxq ; , 1

2 1
F xq ; , F xq ; , Fxq ; , 4.5

2
F xq ; 1, F xq ; 2, Fxq ; 1,
,
2 1
F xq ; , F xq ; , Fxq ; ,
where the latter equation follows because of the reparameterization and .
We further discuss the inverse Gaussian distribution, which possesses heavier tails
than, say, gamma distribution, and therefore it is somewhat more tolerant to large losses.
Example 4.4. Let Y IG, be an inverse Gaussian rv. We then can write its pdf as

y 1 1
f y exp , y 0, , 4.6
2y3 22 2y
cf. 24, whichmeans that Y belongs to the reproductive EDMs, with 1/22 and
1/ 2. Further, due to Corollary 4.1 we need to calculate

2
2
/F Y yq ; ,
TCEq Y ; , log F Y yq ; , .
F Y yq ; ,
4.7
To this end, note that the ddf of Y is

yq yq
F yq ; , 1 e 2/
1 4.8
yq yq
cf., e.g., 28, where is the ddf of the standardized normal random variable. Hence, by
simple dierentiation and noticing that
23/2 3 , 4.9
we obtain that

F yq ; ,

yq yq
yq 1 e 2/
yq 1
yq yq

yq
2e 2/
1 .
yq
4.10
Notably,

yq yq
yq 1 e 2/
yq 1 , 4.11
yq yq
and therefore 4.10 results in

yq
F yq ; , 2e 2/
1 . 4.12
yq
Consequently, the expression for the TCE risk measure, obtained by Landsman and Valdez
16, is simplified to

2 yq
TCEq Y ; , e
2/
1 . 4.13
F yq ; , yq
In order to derive the TSD risk measure we need to dierentiate again, that is,

2 yq
TCEq Y ; , e
2/
1
F yq ; , yq
4.14
2e 2/
/y y / 1
q q

3 1 ,
F yq ; ,
where we use 3 . Further, we have that

2e 2/
/y q y q / 1

F yq ; ,

3 e2/ yq 1 2/ yq / yq
2 4.15

F yq ; ,
2
2e2/ yq
2 ,
F yq ; ,

where yq /yq yq / 1. Therefore

yq
TSDq Y 1 2e
2/
F yq ; ,

3
2/ 2
e2/ yq 1 2/ yq / yq e yq

1 2
F yq ; , F yq ; ,
4.16
subject to VarY 3 /.
5. Concluding Comments
In this work we have considered certain layer-based risk measuring functionals in the context
of the exponential dispersion models. Although we have made an accent on the absolutely
continuous EDMs, similar results can be developed for the discrete members of the class.
Indeed, distributions with discrete supports often serve as frequency models in actuarial
mathematics. Primarily in expository purposes, we further consider a very simple frequency
distribution, and we evaluate the TSD risk measure for it. More encompassing formulas can
however be developed with some eort for other EDM members of, say, the a, b, 0 class cf.,
11, Chapter 6 as well as for limited TCE/TSD risk measures.
Example 5.1. Let X Poisson be a Poisson rv with the mean parameter . Then the
probability mass function of X is written as
1 x 1
px e exp x log , x 0, 1, . . . , 5.1
x! x!
which belongs to the additive EDMs in view of the reparametrization log, 1, and
e .
Motivated by Corollary 4.1, we dierentiate cf. 16, for the formula for the TCE risk
measure

p xq ; , 1
TCEq X; , e
1
F xq ; , 1
2

p xq ; , 1 p xq ; , 1 pxq ; , 1
e 1 xq e e

F xq ; , 1 F xq ; , 1 Fxq ; , 1
2
F x 1; , 1 p x ; , 1 px ; , 1
e ,
q q q
xq e e

F xq ; , 1 F xq ; , 1 Fxq ; , 1
5.2
where the latter equation follows because

F xq ; , 1 p xq ; , 1 F xq 1; , 1 . 5.3
The formula for the TSD risk measure is then
TSDq X

2
p xq ; , 1 F xq 1; , 1 p xq ; , 1 pxq ; , 1
,
e 1 e zq e
F xq ; , 1 F xq ; , 1 F xq ; , 1 Fxq ; , 1
5.4
where EX VarX e and zq xq e .
Appendix
A. Exponential Dispersion Models
Consider a -finite measure on R and assume that is nondegenerate. Next definition is
based on 24.
Definition A.1. The family of distributions of X ED , for , is called

the additive exponential dispersion model generated by . The corresponding family of
distributions of Y X/ ED, 2 , where and 2 1/ are the mean value
and the dispersion parameters, respectively, is called the reproductive exponential dispersion
model generated by . Moreover, given some measure the representation of X ED ,
is as follows:
expx dx. A.1

If in addition the measure has density c x; with respect to some fixed measure
typically Lebesgue measure or counting measure, the density for the additive model is
f x; , c x; expx , x R. A.2
Similarly, we obtain the following representation of Y ED, 2 as

exp y dy , A.3
where denotes transformed by the duality transformation X Y/ 2 . Again if the

measure has density cy; with respect to a fixed measure, the reproductive model has
the following pdf:

f y; , c y; exp y , y R. A.4
Note that and are called canonical and index parameters, { R : < }
for some function called the cumulant, and is the index set. Throughout the paper, we
use X ED , 2 and X ED, for the additive form with parameters and 2 and the
reproductive form with parameters and , correspondingly, depending on which notation
is more convenient.
We further briefly review some properties of the EDMs related to this work. Consider
the reproductive form first, that is, Y ED, 2 , then
i the cumulant generating function cgf of Y is, for t/,

t
Kt log E etY log exp y d y
R

t

log exp exp y d y A.5
R

t
,

ii the moment generating function mgf of Y is given by

t
Mt exp , A.6

iii the expectation of Y is

Kt
EY , A.7
t t0
iv the variance of Y is

2 Kt
VarY 2 2 . A.8
t2 t0
Consider next an rv X following an additive EDM, that is, X ED , . Then, in a

similar fashion,
i the cgf of X is
Kt t , A.9
ii the mgf of X is
Mt exp t , A.10
iii the expectation of X is
EX , A.11
iv the variance of X is
VarX 2 . A.12
For valuable examples of various distributions belonging in the EDMs we refer to

Jrgensen 24.
Acknowledgments
This is a concluding part of the authors research supported by the Zimmerman Foundation
of Banking and Finance, Haifa, Israel. In addition, Edward Furman acknowledges the support
of his research by the Natural Sciences and Engineering Research Council NSERC of
Canada. Also, the authors are grateful to two anonymous referees and the editor, Professor
Johanna Neslehova, for constructive criticism and suggestions that helped them to revise the
paper.
References
1 D. H. Leavens, Diversification of investments, Trusts and Estates, vol. 80, no. 5, pp. 469473, 1945.
2 D. Denneberg, Non-Additive Measure and Integral, vol. 27 of Theory and Decision Library. Series B:
Mathematical and Statistical Methods, Kluwer Academic Publishers, Dordrecht, The Netherlands, 1994.
3 S. Wang, Insurance pricing and increased limits ratemaking by proportional hazards transforms,
Insurance: Mathematics & Economics, vol. 17, no. 1, pp. 4354, 1995.
4 S. S. Wang, Premium calculation by transforming the layer premium density, ASTIN Bulletin, vol.
26, pp. 7192, 1996.
5 S. S. Wang, V. R. Young, and H. H. Panjer, Axiomatic characterization of insurance prices, Insurance:
Mathematics & Economics, vol. 21, no. 2, pp. 173183, 1997.
6 E. Furman and R. Zitikis, Weighted premium calculation principles, Insurance: Mathematics &
7 E. Furman and R. Zitikis, Weighted pricing functionals with applications to insurance: an overview,
North American Actuarial Journal, vol. 13, no. 4, pp. 483496, 2009.
8 E. Furman and Z. Landsman, On some risk-adjusted tail-based premium calculation principles,
Journal of Actuarial Practice, vol. 13, pp. 175191, 2006.
9 W. Hurlimann, Conditional value-at-risk bounds for compound Poisson risks and a normal
approximation, Journal of Applied Mathematics, no. 3, pp. 141153, 2003.
10 E. Furman and Z. Landsman, Tail variance premium with applications for elliptical portfolio of
risks, Astin Bulletin, vol. 36, no. 2, pp. 433462, 2006.
11 S. A. Klugman, H. H. Panjer, and G. E. Willmot, Loss Models from Data to Decisions, Wiley Series in
Probability and Statistics, John Wiley & Sons, Hoboken, NJ, USA, 3rd edition, 2008.
12 W. Hurlimann, Analytical evaluation of economic risk capital for portfolios of gamma risks, Astin
Bulletin, vol. 31, no. 1, pp. 107122, 2001.
13 E. Furman and Z. Landsman, Risk capital decomposition for a multivariate dependent gamma
portfolio, Insurance: Mathematics & Economics, vol. 37, no. 3, pp. 635649, 2005.
14 H. H. Panjer, Measurement of risk, solvency requirements, and allocation of capital within financial
conglomerates, Research Report 01-15, Institute of Insurance and Pension Research, University of
Waterloo, 2002.
15 Z. M. Landsman and E. A. Valdez, Tail conditional expectations for elliptical distributions, North
American Actuarial Journal, vol. 7, no. 4, pp. 5571, 2003.
16 Z. Landsman and E. A. Valdez, Tail conditional expectations for exponential dispersion models,
Astin Bulletin, vol. 35, no. 1, pp. 189209, 2005.
17 E. Furman and Z. Landsman, Multivariate Tweedie distributions and some related capital-at-risk
analyses, Insurance: Mathematics and Economics, vol. 46, no. 2, pp. 351361, 2010.
18 A. Chiragiev and Z. Landsman, Multivariate Pareto portfolios: TCE-based capital allocation and
divided dierences, Scandinavian Actuarial Journal, vol. 2007, no. 4, pp. 261280, 2007.
19 R. Vernic, Tail conditional expectation for the multivariate Pareto distribution of the second kind:
another approach, Methodology and Computing in Applied Probability. In press.
20 A. V. Asimit, E. Furman, and R. Vernic, On a multivariate Pareto distribution, Insurance: Mathematics
and Economics, vol. 46, no. 2, pp. 308316, 2010.
21 M. C. K. Tweedie, Functions of a statistical variate with given means, with special reference to
Laplacian distributions, Proceedings of the Cambridge Philosophical Society, vol. 49, pp. 4149, 1947.
22 B. Jrgensen, Some properties of exponential dispersion models, Scandinavian Journal of Statistics,
vol. 13, no. 3, pp. 187197, 1986.
23 B. Jrgensen, Exponential dispersion models, Journal of the Royal Statistical Society. Series B, vol. 49,
no. 2, pp. 127162, 1987.
24 B. Jrgensen, The Theory of Dispersion Models, vol. 76 of Monographs on Statistics and Applied Probability,
Chapman & Hall, London, UK, 1997.
25 J. A. Nelder and R. W. M. Wedderburn, Generalized linear models, Journal of the Royal Statistical
Society, Series A, vol. 135, pp. 370384, 1972.
26 J. Rioux and S. Klugman, Toward a unified approach to fitting loss models, 2004, http://www
.iowaactuariesclub.org/library/lossmodels.ppt.
27 E. Furman, On a multivariate gamma distribution, Statistics & Probability Letters, vol. 78, no. 15, pp.
23532360, 2008.
28 R. S. Chhikara and J. L. Folks, Estimation of the inverse Gaussian distribution function, Journal of
the American Statistical Association, vol. 69, pp. 250254, 1974.
doi:10.1155/2010/726389
Research Article
Pricing Equity-Indexed Annuities under
Stochastic Interest Rates Using Copulas
Patrice Gaillardetz
Department of Mathematics and Statistics, Concordia University, Montreal, QC, Canada H3G 1M8
Correspondence should be addressed to Patrice Gaillardetz, [email protected]
Received 1 October 2009; Accepted 18 February 2010
Academic Editor: Johanna Neslehova
Copyright q 2010 Patrice Gaillardetz. This is an open access article distributed under the Creative
We develop a consistent evaluation approach for equity-linked insurance products under

stochastic interest rates. This pricing approach requires that the premium information of standard
insurance products is given exogenously. In order to evaluate equity-linked products, we derive
three martingale probability measures that reproduce the information from standard insurance
products, interest rates, and equity index. These risk adjusted martingale probability measures are
determined using copula theory and evolve with the stochastic interest rate process. A detailed
numerical analysis is performed for existing equity-indexed annuities in the North American
market.
1. Introduction
An equity-indexed annuity is an insurance product whose benefits are linked to the
performance of an equity market. It provides limited participation in the performance of
an equity index e.g., S&P 500 while guaranteeing a minimum rate of return. Introduced by
Keyport Life Insurance Co. in 1995, equity-indexed annuities have been the most innovative
insurance product introduced in recent years. They have become increasingly popular since
their debut and sales have broken the $20 billion barrier $23.1 billion in 2004, reaching $27.3
billion in 2005. Equity-indexed annuities have also reached a critical mass with a total asset
of $93 billion in 2005 2006 Annuity Fact Book Tables 7-8 from the National Association
for Variable Annuities NAVA. See the monograph by Hardy 1 for comprehensive
discussions on these products.
The traditional actuarial pricing approach evaluates the premiums of standard life
insurance products as the expected present value of its benefits with respect to a mortality
law plus a security loading. Since equity-linked products are embedded with various types of
financial guarantees, the actuarial approach is dicult to extend to these products and often
produces premiums inconsistent with the insurance and financial markets. Many attempts
have been made to provide consistent pricing approaches for equity-linked products using
financial and economical approaches. For instance, Brennan and Schwartz 2 and Boyle and
Schwartz 3 use option pricing techniques to evaluate life insurance products embedded
with some financial guarantees. Bacinello and Ortu 4, 5 consider the case where the interest
rate is stochastic. More recently, Mller 6 employs the risk-minimization method to evaluate
equity-linked life insurances. Young and Zariphopoulou 7 evaluate these products using
utility theory1 . Particularly for equity-indexed annuities, Tiong 8 and Lee 9 obtain closed-
form formulas for several annuities under the Black-Scholes-Merton framework. Moore 10
evaluates equity-indexed annuities based on utility theory. Lin and Tan 11 and Kijima and
Wong 12 consider more general models for equity-indexed annuities, in which the external
equity index and the interest rates are general stochastic dierential equations.
The liabilities and premiums of standard insurance products are influenced by the
insurer financial performance. Indeed, insurance companies adjust their premiums according
to the realized return from their fixed income and other financial instruments as well
as market pressure. Therefore, mortality security loadings underlying insurance pricing
approach evolve with the financial market. With the current financial crisis, a flexible
approach for equity-linked products that allows interdependency between risks should be
used. Hence, we generalize the approach of Gaillardetz and Lin 13 to stochastic interest
rates. Similarly to Wuthrich et al. 14, they introduce a market consistent valuation method
for equity-linked products by combining probability measures using copulas. Indeed, the
deterministic interest rate assumption may be adequate for short-term derivative products;
however, it is undesirable to extrapolate for longer maturities as for the financial guarantees
embedded in equity-linked products. Therefore, we use the approach of Gaillardetz 15
to model standard insurance products under stochastic interest rates. It supposes the
conditional independence between the insurance and interest rate risks. Here, this approach
is generalized to models that are based on copulas.
Similarly to Gaillardetz and Lin 13, we assume that the premium information of term
life insurances, pure endowment insurances, and endowment insurances at all maturities
is obtainable. We obtain martingale measures for each standard insurance product under
stochastic interest rates. To this end, it is required to assume that the volatilities for standard
insurance prices are given exogenously. Gaillardetz 15 provides additional structure to
find an implicit volatilities for the standard insurance and annuity products. Then, the
martingale probability measures for the insurance and interest rate risks are combined with
the martingale measure from the equity index. These extend martingale measures are used to
evaluate equity-linked insurance contracts and equity-indexed annuities in particular.
This paper is organized as follows. The next section presents financial models for
the interest rates and equity index as well as insurance model. We then derive martingale
measures for those standard insurance products under stochastic interest rates in Section 3. In
Section 4, we derive the martingale measures for equity-linked products. Section 5 focuses on
recursive pricing formulas for equity-linked contracts. Finally, we examine the implications
of the proposed approaches on the EIAs by conducting a detailed numerical analysis in
Section 6.
2. Underlying Financial Models

In this section, we present a multiperiod discrete model that describes the dynamic of a
stock index and the interest rate. These lattice models have been intensively used to model
stocks, stock indices, interest rates, and other financial securities due to their flexibility and
tractability; see Panjer et al. 16 and Lin 17, for example. Moreover, as it often happens
when working in a continuous framework, it becomes necessary to resort to simulation
methods in order to obtain a solution to the problems we are considering. Moreover, the
premiums obtained from discrete models converge rapidly to the premiums obtained with
the corresponding continuous models when considering equity-indexed annuities.
2.1. Interest Rate Model

Similarly to Gaillardetz 15, it is assumed that the short-term ratefollows that of Black et al.
18 BDT, which means that the short-term rate follows a lattice model that is recombining
and Markovian. Particularly, the short-term rate can take exactly t 1 distinct values at year t
denoted by rt, 0, rt, 1, . . . , rt, t. Indeed, rt, l represents the short-term rate between time
t and t 1 that has made l up moves. The short-term rate today, r0, is equal to r0, 0, and
in the case where rt rt, l, the short-term rate at time t1, rt1 can only take two values,
either rt 1, l decrease or rt 1, l 1 increase. We consider the short-term rate process
under the martingale measure Q and hence, the discounted value process Lt, T /Bt is a
martingale. Lt, T represents the price at time t of a default-free, zero-coupon bond paying
one monetary unit at time T and Bt, the money market account, represents one monetary
unit B0 1 accumulated at the short-term rate

t1
Bt 1 ri. 2.1
i 0
Let qt, l be the probability under Q that the short-term rate increases at time t 1
given rt rt, l. That is
qt, l Qrt 1 rt 1, l 1 | rt rt, l, 2.2
for 0 l t, which is set to be 0.5 under the BDT model. Figure 1 describes the dynamic of
the short-term rate process.
The BDT model also assumes that short-term rate process matches an array of yields
volatilities r 1, r 2, . . ., which is assumed to be observable from the financial market.
This vector is deterministic, specified at time 0, and each element is defined by
r t2 Varln rt | rt 1 rt 1, l

2 2.3
rt, l 1
0.5 ln ,
rt, l
for l 0, 1, . . . , t 1 and t 1, . . . . Hence, rt, l 1 is larger than rt, l thus, 2.3 may be
rewritten as follows:

rt, l 1
r t 0.5 ln . 2.4
rt, l
r3, 3
q2, 2
q1, 1 r2, 2
1 q2, 2
r1, 1 r3, 2
q0, 0 1 q1, 1 q2, 1
r0 r2, 1
1 q0, 0 q1, 0 1 q2, 1
r1, 0 r3, 1
1 q1, 0 q2, 0
r2, 0
1 q2, 0
r3, 0
Figure 1: The probability tree of the BDT model process over 3 years.
Equation 2.4 holds for l {0, 1, . . . , t 1} and leads to
rt, l rt, 01l/t rt, tl/t , 2.5
for l 0, 1, . . . , t. Equations 2.4 and 2.5 lead to

1 rt, t
r t ln . 2.6
2t rt, 0
By matching the market prices and the model prices, we have

1
L0, t E
Bt
2.7
2t1 1 l1 1
lt2 1
t1
1 rm, lm 1 ,
1 r0 l 0 l l l l m 1
1 2 1 t1 t2
where E represents the expectation with respect to Q. Replacing rt, l, l 1, 2, . . . , t 1, in
2.7 using 2.5 leads to a system of two equations 2.6 and 2.7 with two unknowns rt, t
and rt, 0, which can be solved for all t.
2.2. Index Model

Similar to Gaillardetz and Lin 13, we suppose that each year is divided into N trading
subperiods of equal length N 1 , which means that the set of trading dates for the index
is {0, , 2, . . .}. We also assume a lattice index model such that the index process Sk, k
, 2, . . ., has two possible outcomes Skdk and Skuk given Sk for the time
period k , k, where S0 is the initial level of the index. The index level Sk dk at
time k represents the index level when the index value goes down and Skuk represents
the index level when the index values goes up. Since the short-term rate is a yearly process,
we assume that the values dk and uk are constant for each year. Hence, we may write
dk dt and uk ut for k t 1 , t 1 2, . . . , t. Because of the number of trading
dates per year, the time-k value of the money-market account is given by

Bk 1 ri , 2.8
i {0,,...,k}
for k 0, , 2, . . . . Here, is the floor function.

A martingale measure for the financial model needs to be determined for the valuation
of equity-linked products. This martingale measures Q should be such that Lt, T /Bt, t
0, 1, . . . , T , and Sk/Bk, k 0, , . . . , , remain martingales. Note that the goal of this
section is to derive the conditional distribution of the index process. Hence, the constraints
imposed by the martingale discounted value process Lt, T /Bt are to be discussed later. Let
k, l be the conditional probability that the index value goes up during the period k , k
given rt 1 rt 1, l, that is,
QSk Sk ut | Sk , rt 1 rt 1, l k, l,

2.9
QSk Sk dtSk , rt 1 rt 1, l 1 k, l,
for k t 1 , t 1 2, . . . , t and t 1, 2, . . . . Supposing that the discounted value process

Sk/Bk is a martingale implies
1 rt 1, l dt

k, l , 2.10
ut dt
for k t1, t12, . . . , t and t 1, 2 . . . . From 2.10 it is obvious that k, l is constant
over each year, that is, k, l t, l for k t 1 , t 1 2, . . . , t. The no-arbitrage thus
requires
dt < 1 rt 1, 0 , 1 rt 1, t 1 < ut, 2.11
for t 1, 2, . . .. The previous conditions may not be respected for the BDT model when long
maturity or high volatility are considered. In this case, the bounded trinomial model from
Hull and White 19 would be more suitable.
Under this model, the ratio St/St 1 takes N 1 possible values denoted t, i,
i 0, 1, . . . , N, which are defined by
t, i uti dtNi , i 0, 1, . . . , N. 2.12
Their corresponding conditional martingale probabilities are

N
St
Q t, i | rt 1 rt 1, l t, li 1 t, lNi , 2.13
St 1 i
for i 0, . . . , N.
t 1, l Stut 12
Stut 1
t 1, l 1 t 1, l
St Stut 1dt 1

1 t 1, l t 1, l
Stdt 1
1 t 1, l
Stdt 12
Figure 2: The probability tree of the index under stochastic interest rates over t and t 1 when N 2 given
rt rt, l.
The model assumes the usual frictionless market: no tax, no transaction costs, and so
forth. Furthermore, for practical implementation purposes, one may also use current forward
rates for r.
Figure 2 presents the conditional index process tree under stochastic interest rates
when N 2 for the time period t, t 1.
For notational convenience, let
it {i0 , . . . , it }, 2.14
which represents the indexs realization up to time t with

t
St, it S0 l, il , 2.15
l 0
for t 0, 1, . . ., where 0, i0 1.
2.3. Insurance Models

In this subsection, we introduce lattice models for the standard insurance products under
stochastic interest rates. We will use the standard actuarial notation which can be found in
Bowers et al. 20. Let T x be the future lifetime of insured x of age x and the curtate-
future-lifetime
Kx T x 2.16
the number of future complete years lived by the insured x prior to death. For notational
purposes, let
lt {l0 , l1 , . . . , lt } 2.17
represent the realization of the short-term rate process up to time t with ri ri, li , i
0, 1, . . . , t, where l0 0.
For integers t and n t n, let V j x, t, n, lt } denote, respectively, the time-t prices for
the n-year term life insurance j 1, n-year pure endowment insurance j 2, and n-year
endowment insurance j 3 given that the short-term rate followed the path lt .
The value process W 1 x, t, n, lt1 of n-year term life insurance is defined by

Bt

, Kx < t,
BKx 1
1
W x, t, n, lt1 V 1 x, t, n, {lt1 , lt1 1}, Kx t, rt rt, lt1 1, 2.18

V 1 x, t, n, {l , l }, Kx t, rt rt, lt1 ,
t1 t1
with V 1 x, n, n, ln 0. Note that lt represents the interest rate information known by the
process, but does not stand as an indexing parameter.
Similarly, define W 2 x, t, n, lt1 to be the value process of the n-year pure endowment
insurance and it is given by

0, Kx < t,

W 2 x, t, n, lt1 V 2 x, t, n, {lt1 , lt1 1}, Kx t, rt rt, lt1 1, 2.19

2
V x, t, n, {lt1 , lt1 }, Kx t, rt rt, lt1 ,
with V 2 x, n, n, ln 1.

Finally, let W 3 x, t, n, lt1 denote the value process generated by the n-year
endowment insurance

Bt

, Kx < t,
BKx 1
3
W x, t, n, lt1 V 3 x, t, n, {lt1 , lt1 1}, Kx t, rt rt, lt1 1, 2.20

V 3 x, t, n, {l , l }, Kx t, rt rt, lt1 ,
t1 t1
with V 3 x, n, n, ln 1

The processes W j x, t, n, j 1, 2, 3, represent the intrinsic values of the standard
insurance products and are presented in Figure 3.
3. Martingale Measures for Insurance Models

In this section, we employ a method similar to the approach of Gaillardetz 15 to
derive a martingale probability measure for each of the value processes introduced in
the last section. Gaillardetz 15 derives these martingale measures under conditional
independence assumptions. Here, we relax this assumption by using copulas to describe
1{j{1,3}} and rt 1, lt 1
j
bx t, 3, lt
j
bx t, 2, lt 1{j{1,3}} and rt 1, lt
j
V x, t, n, lt and rt, lt j
bx t, 1, lt
V j x, t 1, n, {lt , lt 1}

j and rt 1, lt 1
bx t, 0, lt
V j x, t 1, n, {lt , lt } and rt 1, lt
Figure 3: The probability tree of the combined insurance product j 1, 2, 3 and short-term rate processes
between t and t 1 given Kx t and r0, r1, . . . , rt.
possible dependence structures between interest rates and insurance products. It is important
to point out that these probabilities are age-dependent and include an adjustment for the
mortality risk since we use the information from the insurance market.
j
The martingale measures Qx , j 1, 2, 3, are defined such that W j x, t, n/Bt
and Lt, T /Bt are martingales. As mentioned in Section 2, we assume that the time-0
premiums V j x, 0, n, j 1, 2, 3, of the term life insurance, pure endowment insurance,
and endowment insurance are given exogenously. The annual short-term rate process rt is
governed by the BDT model with qt, l 0.5 and volatilities r t are given exogenously for
l 0, 1, . . . , t and t 1, 2, . . . . The conditional martingale probability of each possible outcome
is defined by
j j
bx t, 0, lt Qx Kx > t, rt 1 rt 1, lt | Kx t, lt ,
j j
bx t, 1, lt Qx Kx > t, rt 1 rt 1, lt 1 | Kx t, lt ,
3.1
j j
bx t, 2, lt Qx Kx t, rt 1 rt 1, lt | Kx t, lt ,
j j
bx t, 3, lt Qx Kx t, rt 1 rt 1, lt 1 | Kx t, lt ,
for j 1, 2, 3. These martingale probabilities are presented above each branch in Figure 3.
j
The main objective of this section is to determine bx s that will be used to evaluate
equity-indexed annuities in later sections.
To ensure that the discounted value process Lt, T /Bt is a martingale, we must have
j j 1 j j 1

bx t, 0, lt bx t, 2, lt , bx t, 1, lt bx t, 3, lt , 3.2
2 2
for j 1, 2, 3, t 0, 1, . . ., and all lt . Note that the martingale mortality and survival
probabilities are given, respectively, by
j j j j

qx t, lt Qx Kx t | Kx t, lt bx t, 2, lt bx t, 3, lt ,
3.3
j j j j
px t, lt Qx Kx > t | Kx t, lt bx t, 0, lt bx t, 1, lt .
As in the short-term rate model, additional structure is needed to set the time-t
premiums. Similar to Black et al. 18, we suppose that the volatilities of insurance liabilities
j
x t, lt1 are defined at time t by

1 1
x t, lt1 2 Varx ln W 1 x, t, n, lt1 | Kx t, lt1 , 3.4
j j

x t, lt1 2 Varx ln 1 W j x, t, n, lt1 | Kx t, lt1 , 3.5
j
for j 2, 3, t 1, 2, . . . , and all lt1 . Here, Varx represents the conditional variance with
j
respect to Qx . We assume that the volatilities are deterministic but vary over time and are
given exogenously. Gaillardetz 15 uses the natural logarithm function to ensure that each
process remains strictly positive. Since W 1 is close to 0, it directly uses lnW 1 to ensure
that the process remains strictly greater than 0. On the other hand, it uses ln1 W j for
j 2, 3 to ensure that the processes are strictly smaller than 1 since W j s are closer to 1.
j
In order to identify the martingale probabilities bx , Gaillardetz 15 assumes the
independence or the conditional independence between the interest rate process and the
insurers life. Here, the additional structure is provided by the choice of copulas. Indeed,
the dependence structure between the interest rates and the premiums of insurance products
is modeled using a copula. The main advantage of using copulas is that they separate a joint
distribution function in two parts: the dependence structure and the marginal distribution
functions. We use them because of their mathematical tractability and, based on the Sklars
Theorem, they can express all multivariate distributions. A comprehensive introduction may
be found in Joe 21 or Nelsen 22. Frees and Valdez 23, Wang 24, and Venter 25 have
given an overview of copulas and their applications to actuarial science. Cherubini et al. 26
present the applications of copulas in finance.
There exists a wide range of copulas that may define a joint cumulative distribution
function. The simplest one is the independent copula

CI FY1 y1 , FY2 y2 FY1 y1 FY2 y2 , 3.6
where FY1 and FY2 are marginal cumulative distribution functions. Extreme copulas are
defined using the upper and lower Frechet-Hoeding bounds, which are given by

CU FY1 y1 , FY2 y2 min FY1 y1 , FY2 y2 , 3.7

CD FY1 y1 , FY2 y2 max FY1 y1 FY2 y2 1, 0 . 3.8
One of the most important families of copulas is the archimedean copulas. Among them, the
Cook-Johnson Clayton copula is widely used in actuarial science because of its desirable
properties and simplicity. The Cook-Johnson copula with parameter > 0 is given by
CJ 1/
C FY1 y1 , FY2 y2 FY1 y1 FY2 y2 1 . 3.9
The Gaussian 1 1 copula, which is often used in finance, is defined as

CG FY1 y1 , FY2 y2 1 FY1 y1 , 1 FY2 y2 , 3.10
where is the bivariate standard normal cumulative distribution function with correlation
coecient and 1 is the inverse of the standard normal cumulative distribution function.
Hence, the parameter in formulas 3.9 and 3.10 indicates the level of dependence between
the insurance products and interest rates.
The joint cumulative distribution of W j and rt is obtained using a copula Ct , that
is,
j

Qx W j x, t 1, n, lt y1 , rt 1 y2 | Kx t, lt
j 3.11

Ct Qx W j x, t 1, n, lt y1 | Kx t, lt , Q rt 1 y2 | lt ,
for j 1, 2, 3, where the copula may be defined by either 3.6, 3.7, 3.8, 3.9, or 3.10.
The martingale probabilities have the following constraints:
j
bx t, 2, lt
j

Qx W j x, t 1, n, lt 1, rt 1 rt 1, lt | Kx t, lt
j

Qx W j x, t 1, n, lt 1, rt 1 rt 1, lt | Kx t, lt
j

Qx W j x, t 1, n, lt V j x, t 1, n, {lt , lt }, rt 1 rt 1, lt | Kx t, lt .
3.12
It follows from 3.11 that
j
bx t, 2, lt
j
Qx rt 1 rt 1, lt | Kx t, lt
j 3.13
Ct Qx W j x, t 1, n, lt V j x, t 1, n, {lt , lt } | Kx t, lt ,

Qrt 1 rt 1, lt | lt .
Using the following inequality
V j x, t 1, n, {lt , lt } V j x, t 1, n, {lt , lt 1} 3.14
in 3.13 leads to
j
j j

bx t, 2, lt 0.5 Ct bx t, 0, lt bx t, 1, lt , 0.5
j 3.15
0.5 Ct px t, lt , 0.5 ,
for j 1, 3, and we have for j 2

2 2
bx t, 2, lt Qx W 2 x, t 1, n, lt 0, rt 1 rt 1, lt | Kx t, lt
3.16
2
Qx W 2 x, t 1, n, lt 0, rt 1 rt 1, lt | Kx t, lt .
It follows from 3.11 that

2 2
bx t, 2, lt Ct Qx W 2 x, t 1, n, lt 0 | Kx t, lt , Qrt 1 rt 1, lt | lt

2
Ct qx t, lt , 0.5 .
3.17
3.1. Term Life Insurance

Proposition 3.1. For given V 1 x, 0, n, l0 (n 1, 2, . . . , ), copulas Ct , and volatilities
1
x t, lt1 (t 1, 2, . . . , and all lt1 ), the age-dependent, mortality risk-adjusted martingale
probabilities are given by

1
bx t, 2, lt 0.5 Ct 1 V 1 x, t, t 1, lt 1 rt, lt , 0.5 , 3.18
1 1
bx t, 3, lt V 1 x, t, t 1, lt 1 rt, lt bx t, 2, lt , 3.19
1 1 1
bx t, i, lt bx t, i 2, lt , 3.20
2
for i 0, 1, where the price at time t is defined recursively using
V 1 x, t 1, n, {lt , lt }

1 1
V 1 x, t, n, lt 1 rt, lt bx t, 2, lt bx t, 3, lt 3.21
0.5
,
1 1 1 1 1 1 1
bx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt
bx t, 0, lt bx t, 1, lt e
V 1 x, t 1, n, {lt , lt 1}

1 1
V 1 x, t, n, lt 1 rt, lt bx t, 2, lt bx t, 3, lt 3.22
0.5
,
1 1 1 1 1 1 1
bx t, 1, lt bx t, 0, lt ebx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt
for t 0, 2, . . . , n 2 and all lt .
Proof. The proof is similar to the proof of Proposition 3.1 of Gaillardetz 15 and can be found
in Gaillardetz 27.
With the martingale structure identified, the n-year term life insurance premiums may
be reproduced as the expected discounted payo of the insurance

1 1{Kx<n}
V 1 x, 0, n Ex . 3.23
BKx 1
3.2. Pure Endowment Insurance

Proposition 3.2. For given V 2 x, 0, n, l0 (n 1, 2, . . . , ), copulas Ct , and volatilities
2
x t, lt1 (t 1, 2, . . . , and all lt1 ), the age-dependent, mortality risk-adjusted martingale
probabilities are given by

2
bx t, 2, lt Ct 1 V 2 x, t, t 1, lt 1 rt, lt , 0.5 , 3.24
2 2
bx t, 3, lt 1 V 2 x, t, t 1, lt 1 rt, lt bx t, 2, lt , 3.25
2 1 2
bx t, i, lt bx t, i 2, lt , 3.26
2
V 2 x, t 1, n, {lt , lt }

0.5

2 2 2 2 2 2
2 bx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt
V x, t, n, lt 1 rt, lt bx t, 1, lt 1e
0.5
,
2 2 2 2 2 2 2
V 2 x, t 1, n, {lt , lt 1}

0.5

2 2 2 2 2 2
V 2 x, t, n, lt 1rt, lt bx t, 0, lt 1 ebx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt
0.5
,
2 2 2 2 2 2 2
3.27
for t 0, 2, . . . , n 2 and all lt .
Proof. The proof is similar to the proof of Proposition 3.2 of Gaillardetz 15 and can be found
in Gaillardetz 27.
With the martingale structure identified, the n-year pure endowment insurance
premiums may be reproduced as the expected discounted payo of the insurance

2 1{Kxn}
V 2 x, 0, n Ex . 3.28
Bn
3.3. Endowment Insurance

There is no general solution for the endowment insurance products since the n-year
endowment insurance price at time n 2 may not be expressed using only either mortality
or survival probabilities. For the n-year term-life insurance, the time-n 1 price is
determined based on the death martingale probabilities and the n-year pure endowment
price may be obtained using the survival probabilities at time n 1. Therefore, once
you combine both products to form an endowment insurance, there is no way to solve
explicitly for the martingale probabilities. However, closed-from solutions may be derived
for the independent, upper and lower copulas. Numerical methods need to be used for
the Cook-Johnson and Gaussian copulas. Furthermore, the width of the participation rate
bands for the unified approach is narrow under deterministic interest see Gaillardetz and
Lin 13. For these reasons, we are focusing on the independent and Frechet-Hoeding
bounds.
3
Proposition 3.3. For given V 3 x, 0, n, l0 (n 1, 2, . . . , ), copulas C, and volatilities x t, lt1
(t 1, 2, . . . , and all lt1 ), the age-dependent, mortality risk-adjusted martingale probabilities are
given by

V 3 x, t, t 2, lt 1 rt, lt

1 1

0.5

1 rt 1, lt 1 rt 1, lt 1

1 1

, Independent,

2
1 rt 1, lt 1 rt 1, lt 1

3
bx t, 2, lt 0, Upper, 3.29

V 3 x, t, t 2, lt 1 rt, lt

0.5
1

1

1 rt 1, lt 1 rt 1, lt 1

1

1 , Lower,
1 rt 1, lt

3

bx t, 2, lt , Independent,

V 3 x, t, t 2, lt 1 rt, lt

0.5 1 1
3
bx t, 3, lt 1 rt 1, lt 1 rt 1, lt 1 3.30

1

1 , Upper,

1 rt 1, lt 1

0, Lower,
3 3
bx t, i, lt 0.5 bx t, i 2, lt , 3.31
V 3 x, t 1, n, {lt , lt }

3 3
V 3 x, t, n, lt 1 rt, lt bx t, 2, lt bx t, 3, lt
0.5

3 3 3 3 3
3 bx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt
bx t, 1, lt 1e
0.5

3 3 3 3 3
3 3
bx t, 0, lt bx t, 1, lt ebx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt ,
V 3 x, t 1, n, {lt , lt }

3 3
V 3 x, t, n, lt 1 rt, lt bx t, 2, lt bx t, 3, lt
0.5

3 3 3 3 3
3
bx t, 0, lt 1 ebx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt
0.5

3 3 3 3 3
3 3
bx t, 1, lt bx t, 0, lt ebx t,0,lt bx t,1,lt /{bx t,0,lt bx t,1,lt } x t1,lt ,
3.32
for t 0, 2, . . . , n 3 and all lt .
Proof. The proof can be found in Gaillardetz 27.
Since we suppose that the time-0 insurance prices, the insurance volatilities, the zero-
coupon bond prices, and the interest rate volatility are given exogenously, it is possible to
extract the stochastic structure of each insurance products using Propositions 3.1, 3.2, and
3.3. There are constraints on the parameters because the martingale probabilities should be
strictly positive. However, there is no closed-form solution for the stochastic interest models.
Theoretically, there exists a natural hedging between the insurance and annuity
products. However, Gaillardetz and Lin 13 argue that it is reasonable to evaluate insurances
and annuities separatelysince in practice due to certain regulatory and accounting constraints
and issues such as moral hazard and anti-selection.
3.4. Determination of Insurance Volatility Structure

For implementation purposes, we now relax the assumption of exogenous insurance
j
volatilities. In Subsections 3.1, 3.2, and 3.3, the volatilities of insurance liabilities x t, lt1
defined by either 3.4 or 3.5 were supposed to be known. However, identifying these
volatilities is extremely challenging due to the lack of empirical data and studies. Similar
to Gaillardetz 15, we extract an implied volatility from the insurance market under certain
assumptions.
There are three dierent sources that define the insurance volatilities: the interest rates,
the insurance prices, and the martingale probabilities. The implied insurance volatilities is
obtained assuming that the short-term rate has no impact on the martingale probabilities.
Thus, we extract the insurance volatility such that the martingale probabilities in the case
of an up move from the interest rate process are equal to the martingale probabilities in the
j
case of a down move. Let x t, lt1 j 1, 2, 3, t 1, 2, . . ., and all lt1 denote the implied
volatilities defined by 3.4 for j 1 and 3.5 for j 2, 3, under the following constraint:
j j
qx t 1, {lt , lt } qx t 1, {lt , lt 1}, 3.33
for j 1, 2, 3. In other words, insurance companies that do not react to the interest rate
j
change should have an insurance volatility close to x . Gaillardetz 13, 27 explain that
behavior of insurance companies facing the interest rate shifts could be understood through
these volatilities. They also describe recursive formulas to obtain numerically the implied
volatilities. In the following examples, equity-indexed annuity contracts are evaluated using
the implied volatilities, which are obtained from 3.33.
4. Martingale Measures for Equity-Linked Products

Due to their unique designs, equity-linked products involve mortality and financial risks
since these type of contracts provide both death and accumulation/survival benefits.
Moreover, the level of these benefits are linked to the financial market performance and
an equity index in particular. Hence, it is natural to assume that equity-linked products
belong to a combined insurance and financial markets since they are simultaneously subject
to the interest rate, equity, and mortality risks. Similar to Section 3, we evaluate these types
of products by evaluating the death benefits and survival benefits separately. Under this
approach, two martingale measures again need to be generated: one for death benefits and
another for survival benefits. Furthermore, these martingale measures should be such that
they reproduce the index values in Section 2 and the premiums of insurance products under
stochastic interest rates in Section 3. In other words, the marginal probabilities derived in
j
the previous sections should be preserved, and the martingale measures Qx , j 1, 2, 3 are
such that {W j x, t, n/Bt, t 0, 1, . . .}, {Lt, T /Bt, t 0, 1, . . . , T and T 1, . . .}, and
j
{Sk/Bk, k 0, , . . .} will remain martingales. Let ex t, i, it , lt denote the martingale
j
probability under Qx such that x survives and

St 1
rt 1 rt 1, lt ,

St
t 1, i, i 0, . . . , N,

St 1

rt 1 rt 1, lt 1, t 1, i N 1, i N 1, . . . , 2N 1
St
4.1
or the martingale probability such that x dies and

St 1
rt 1 rt 1, lt ,

St
t, i 2N 2, i 2N 2, . . . , 3N 2,

St 1

rt 1 rt 1, lt 1, t, i 3N 3, i 3N 3, . . . , 4N 3,
St
4.2
between t and t 1 given St, Kx t, and lt as illustrated in Figure 4. The function is
given explicitly by 2.12.
j
What remains is to determine the probabilities ex s for all it and lt . We introduce
the dependency between the index process, the short-term rate, and the premiums of
j
insurance products using copulas. Let Gx , j 1,2,3, denote this joint conditional cumulative
distribution function over time t and t 1. That is
j j

Gx y1 , y2 , y3 ; it , lt Qx St 1 y1 , W j x, t 1, n, lt y2 ,
4.3
rt 1 y3 | Kx t, it , lt .
As explained, the marginal cumulative distribution functions of the insurance products and
the index are preserved under the extended measures, that is,
j
Gx , y2 , y3 ; it , lt
j

Qx W j x, t 1, n, lt y2 , rt 1 y3 | it , lt , 4.4
j
Gx y1 , , ; it , lt Q St 1 y1 | it , lt ,
which are determined using 3.18, 3.19, and 3.20 for j 1, 3.24, 45, and 3.26 for
j 2, 3.29, 3.30, as well as 3.31 for j 3, and 2.13 for the index. Let Ct be the choice
j
of copula, then the cumulative distribution function Gx is defined by
j j j

Gx y1 , y2 , y3 ; it , lt Ct Gx y1 , , ; it , lt , Gx , y2 , y3 ; it , lt , 4.5
where t represents the free parameter between t and t 1 that indicate the level of
dependence between the insurance product, interest rate, and the index processes. Here,
the copula Ct could be defined using either 3.6, 3.7, 3.8, 3.9, or 3.10. Note that in
j
some cases, for example, the lower copula 3.8, the function Gx would not be a cumulative
j
distribution function. We also remark that Gx s are functions of Kx t, but for notational
simplicity we suppress Kx.
The martingale probabilities can be obtained from the cumulative distribution
function and are given by
j j

ex t, i, it , lt Gx Stt 1, i, V j x, t 1, n, {lt , lt }, rt 1, lt ; it , lt
4.6
j
Gx Stt 1, i 1, V j x, t 1, n, {lt , lt }, rt 1, lt ; it , lt ,
for i 0, . . . , N,
j j

ex t, i, it , lt Gx Stt 1, i N 1, V j x, t 1, n, {lt , lt 1}, rt 1, lt 1; it , lt
j

Gx Stt 1, i N 2, V j x, t 1, n, {lt , lt 1}, rt 1, lt 1; it , lt ,
4.7
1{j {1,3}} ; rt 1, lt 1; Stt 1, N

.
.
.
1{j {1,3}} ; rt 1, lt 1; Stt 1, 0
V j x, t, n, lt 1{j {1,3}} ; rt 1, lt ; Stt 1, N

rt, lt .
.
St .
V j x, t 1, n, {lt , lt 1}; rt 1, lt 1; Stt 1, 0
V j x, t 1, n, {lt , lt }; rt 1, lt ; Stt 1, N
.
.
.
V j x, t 1, n, {lt , lt }; rt 1, lt ; Stt 1, 0
Figure 4: The probability tree of the combined insurance product j 1, 2, 3, short-term rate, and index
processes between t and t 1 given that Kx t, lt , and it .
for i N 1, . . . , 2N 1,
j j
ex t, i, it , lt Gx Stt 1, i 2N 2, 1, rt 1, lt ; it , lt
j
Gx Stt 1, i 2N 3, 1, rt 1, lt ; it , lt
4.8
j
Gx Stt 1, i 2N 2, V j x, t 1, n, {lt , lt }, rt 1, lt ; it , lt
j

Gx Stt 1, i 2N 3, V j x, t 1, n, {lt , lt }, rt 1, lt ; it , lt ,
for i 2N 2, . . . , 3N 2, and
j j
ex t, i, it , lt Gx Stt 1, i 3N 3, 1, rt 1, lt 1; it , lt
j
Gx Stt 1, i 3N 4, 1, rt 1, lt 1; it , lt
j

Gx Stt 1, i 3N 3, V j x, t 1, n, {lt , lt }, rt 1, lt 1; it , lt
j
j

Gx Stt 1, i 3N 4, V j x, t 1, n, {lt , lt }, rt 1, lt 1; it , lt
j
j

Gx Stt 1, i 3N 3, V j x, t 1, n, {lt , lt }, rt 1, lt ; it , lt
j

Gx Stt 1, i 3N 4, V j x, t 1, n, {lt , lt }, rt 1, lt ; it , lt , 4.9
j j
for i 3N 3, . . . , 4N 3 and j 1, 3, where Gx Stt1, 1, . . . ; it , lt 0 and Gx . . . ; it , lt
is obtained using 4.5. Similarly, for j 2,

2 2
ex t, i, it , lt Gx Stt 1, i, V 2 x, t 1, n, {lt , lt }, rt 1, lt ; it , lt

2
Gx Stt 1, i 1, V 2 x, t 1, n, {lt , lt }, rt 1, lt ; it , lt
4.10
2
Gx Stt 1, i, V 2 x, t 1, n, {lt , lt 1}, rt 1, lt ; it , lt

2
Gx Stt 1, i 1, V 2 x, t 1, n, {lt , lt 1}, rt 1, lt ; it , lt ,
for i 0, . . . , N,

2 2
ex t, i, it , lt Gx Stt 1, i N 1, V 2 x, t 1, n, {lt , lt 1}, rt 1, lt 1; it , lt

2
Gx Stt 1, i N 2, V 2 x, t 1, n, {lt , lt 1}, rt 1, lt 1; it , lt
2
Gx Stt 1, i N 1, 0, rt 1, lt 1; it , lt

2
Gx Stt 1, i N 1, V 2 x, t 1, n, {lt , lt 1}, rt 1, lt ; it , lt
2
Gx Stt 1, i N 2, 0, rt 1, lt 1; it , lt

2
Gx Stt 1, i N 2, V 2 x, t 1, n, {lt , lt 1}, rt 1, lt ; it , lt
2
Gx Stt 1, i N 1, 0, rt 1, lt ; it , lt
2
Gx Stt 1, i N 2, 0, rt 1, lt ; it , lt ,
4.11
for i N 1, . . . , 2N 1,
2 2
ex t, i, it , lt Gx Stt 1, i 2N 2, 0, rt 1, lt ; it , lt
4.12
2
Gx Stt 1, i 2N 3, 0, rt 1, lt ; it , lt ,
for i 2N 2, . . . , 3N 2, and
2 2
ex t, i, it , lt Gx Stt 1, i 3N 3, 0, rt 1, lt 1; it , lt
2
Gx Stt 1, i 3N 4, 0, rt 1, lt 1; it , lt
4.13
2
2
Gx Stt 1, i 3N 4, 0, rt 1, lt ; it , lt ,
for i 3N 3, . . . , 4N 3.
Consider now an equity-linked product that pays

DKx 1 if Kx 0, 1, . . . , n 1,
4.14
Dn if Kx n.
For notational convenience, we sometimes use Dt, it to specify the indexs realization.
Let P 1 x, t, n, it , lt denote the premium at time t of the equity-linked contract death
benefit given that x is still alive and the index and short-term rate processes have taken the
path it and lt , respectively. With the martingale structure identified by 4.6, 4.7, 4.8, and
4.9, P 1 x, t, n, it , lt may be obtained as the expected discounted payos

DKx 11{Kx<n}
Btit , lt , Kx t ,
1
P 1 x, t, n, it , lt Ex 4.15
BKx 1
1 1
where Ex represents the expectation with respect to Qx .
On the other hand, let P 2 x, t, n, it , lt denote the premium at time t of the equity-
linked product accumulation benefit given that x is still alive and the index process has
taken the path it . With the martingale structure identified by 4.10, 4.11, 4.12, and 4.13,
P 2 x, t, n, it , lt may be obtained as the expected discounted payos

Dn1{Kxn}
Btit , lt , Kx t .
2
P 2 x, t, n, it , lt Ex 4.16
Bn
Let P x, t, n, it , lt denote the premium at time t of an n-year equity-linked product

issue to x with its payo defined by 4.14. In particular, P x, 0, n, i0 , l0 is the amount
invested by an insured or, from insurers point of view, the premium paid by the
policyholder. We assume that P x, t, n, it , lt may be decomposed in two dierent premiums;
P 1 x, t, n, it , lt is the premium to cover the death benefit and P 2 x, t, n, it , lt is the premium
to cover the accumulation benefit. That is, we assume that
P x, t, n, it , lt P 1 x, t, n, it , lt P 2 x, t, n, it , lt , 4.17
where P 1 x, t, n, it , lt and P 2 x, t, n, it , lt are obtained using 4.15 and 4.16, respectively.
An alternative approach is to evaluate equity-linked products containing death

and accumulation benefits in a unified manner, using the pricing information from the
endowment insurance products. In this case, the n-year equity-linked product premium
P x, t, n, it , lt may be obtained as expected discounted payos

DKx 11{Kx<n1} Dn1{Kxn1}
P x, t, n, it , lt
3
Ex
Btit , lt , Kx t .
BKx 1 Bn
4.18
Figure 4 presents the dynamic of the equity-linked premiums for time period t, t 1.
Bear in mind that the first approach presented in this section evaluates equity-linked
products by loading the death and survival probabilities separately. The second approach
evaluates the equity-linked product using unified loaded probabilities.
5. Evaluation of Equity-Linked Products

In this section, we evaluate equity-linked contracts using recursive algorithms. It follows from
4.15 and 4.16 that
P 1 x, t, n, it , lt

N
1 1 1
ex t, v 2N 2, it , lt ex t, v 3N 3, it , lt Dt 1, {it , v}
1 rt, lt v 0
N
1
ex t, v, it , lt P 1 x, t 1, n, {it , v}, {lt , lt }
v 0

1 1
ex t, v N 1, it , lt P x, t 1, n, {it , v}, {lt , lt 1} ,
5.1
with P 1 x, n, n, in , ln 0 and

N
2 1 2
P x, t, n, it , lt ex t, v, it , lt P 2 x, t 1, n, {it , v}, {lt , lt }
1 rt, lt v 0

2 2
5.2
with P 2 x, n, n, in , ln Dn, in .

Table 1: Point-to-point with term-end design for various interest rate volatilities.
Decomposed approach Unified approach

CJ CJ
r t CI CU CD C0.5 C2 G
C0.1 G
C0.3 CI CU CD
3% Minimum guarantee on 90% premium
0% 62.17 56.27 68.58 59.94 58.38 63.08 59.59 70.28 69.98 70.05
20% 4% 62.17 55.85 69.17 59.79 58.07 63.12 59.46 70.27 69.40 70.64
8% 62.15 55.44 69.75 59.63 57.75 63.15 59.31 70.25 68.83 71.23
30% 0% 48.95 43.20 55.15 47.02 45.63 49.79 46.54 55.46 55.14 55.25
4% 48.95 42.93 55.54 46.93 45.43 49.82 46.45 55.45 54.74 55.68
8% 48.94 42.66 55.93 46.83 45.23 49.85 46.37 55.45 54.35 56.11
0% 44.58 41.32 47.87 43.45 42.65 45.06 43.20 53.25 53.12 52.88
20% 4% 44.57 40.93 48.38 43.32 42.36 45.09 43.07 53.25 52.56 53.46
8% 44.55 40.53 48.89 43.18 42.07 45.12 42.94 53.24 52.00 54.03
0% 32.20 29.12 35.38 31.25 30.56 32.63 30.96 39.27 39.13 38.98
30% 4% 32.20 28.89 35.69 31.18 30.40 32.65 30.89 39.27 38.77 39.36
8% 32.19 28.65 36.00 31.10 30.23 32.67 30.81 39.27 38.41 39.75
3
A recursive formula to evaluate P x, t, n, it , lt under Qx is determined using 4.18,
that is,
P x, t, n, it , lt

N

1 3 3
ex t, v 2N 2, it , lt ex t, v 3N 3, it , lt Dt 1, {it , v}
1 rt, lt v 0
N
5.3
3
ex t, v, it , lt P 3 x, t 1, n, {it , v}, {lt , lt }
v 0

3
for t 0, . . . , n 2, where

N
n, ln1 v 1 n, ln1 Nv
P x, n 1, n, in1 , ln1 Dn, {in1 , v}. 5.4
v 0
1 rn 1, ln1
Note that the surrender options for equity-linked products under stochastic interest
rates are evaluated in Gaillardetz 27.
6. Valuation of Equity-Indexed Annuities: Numerical Examples

This section implements numerically the methods we developed previously by considering
two types of equity-indexed annuities. They appeal to investors because they oer the
same protection as conventional annuities by limiting the financial risk, but also provide
participation in the equity market. From Lin and Tan 11 and Tiong 8, EIA designs may be
generally grouped in two broad classes: Annual Reset and Point-to-Point. The index growth
on an EIA with the former is measured and locked in each year. Particularly, the index growth
with a term-end point design is calculated using the index value at the beginning and the end
of each year. In the latter, the index growth is based on the growth between two time points
over the entire term of the annuity. In the case of the term-end point design, the growth is
evaluated using the beginning and ending values of the index. The cost of the EIA contract is
reflected through the participation rate. Hence, the participation rate is expected to be lower
for expensive designs.
Our examples involve five-year EIAs issued to a male-aged 55 with minimum interest
rate guarantee of either 3% on 100% of the premium or 3% on 90% of the premium. For
illustration purposes, we assume that the insurance product values, V j x, 0, n j 1, 2, 3
and n 1, 2, 3, 4, 5, are determined using the standard deviation premium principle see
Bowers et al. 20 with a loading factor of 5.00% based on the 1980 US CSO table see
http://www.soa.org/. We also assume that the short-term rate rt follows the BDT where
the volatility is either 0%, 4%, or 8%. The observed price of the zero-coupon bond L0, T
is assumed to be equal to 1.05T for T 1, 2, . . . , 5. Hence, the interest rate model may be
calibrated using 2.6 and 2.7. For simplification purposes, the index will be governed by
the Cox et al. 28 model where S0 1 and the number of trading dates N is 3. In this
recombining model, the index at time k Sk has two possible outcomes at time k : it is
either increasing to Sk USk or decreasing to Sk dSk. The increasing and
decreasing factors u and d are supposed to be constant and are obtained from the volatility of
the index

. This volatility is assumed to be constant and is either 20% or 30%. In other words,
u e / N
0.2, 0.3 and d u1 . The index conditional martingale probability structure is
obtained using 2.10. The conditional joint distribution of the interest rates and the insurance
products are obtained using Propositions 3.1, 3.2, and 3.2. Here, these martingale probabilities
are determined based on the implied insurance volatilities, which are derived numerically
under the constraint given in 3.33.
The analysis is performed using the point-to-point and reset EIA classes with term-end
point design.
6.1. Point-to-Point
We first consider one of the simplest classes of EIAs, known as the point-to-point. Their
payos in year t can be represented by
t
Dt max min 1 Rt, 1 t , 1 g , 6.1
where represents the participation rate and the gain Rt need to be defined depending
on the design. It also provides a protection against the loss from a down market 1 gt .
The cap rate 1 t reduces the cost of such contract since it imposes an upper bound on the
maximum return.
As explained in Lin and Tan 11, an EIA is evaluated through its participation rate
. Without loss of generality, we suppose that the initial value of EIA contracts is one
monetary unit. The present value of the EIA is a function of the participation rate through
the payo function D. We then solve numerically for , the critical participation rate, such
Table 2: Point-to-point with term-end design and 15% cap rate for various interest rate volatilities.

CJ CJ
C0.1 G
C0.3 CI CU CD
0% 65.65 59.00 72.98 62.95 61.16 66.69 62.73 75.95 76.61 75.43
20% 4% 65.65 58.61 73.40 62.82 60.88 66.72 62.62 75.95 76.10 75.85
8% 65.64 58.23 73.85 62.68 60.61 66.75 62.50 75.94 75.64 76.29
0% 58.94 51.26 67.42 55.97 54.00 60.11 55.62 69.87 70.94 69.33
30% 4% 58.94 51.10 67.54 55.91 53.89 60.12 55.59 69.88 70.66 69.52
8% 58.94 50.96 67.70 55.86 53.79 60.13 55.56 69.90 70.45 69.76
0% 45.20 41.85 48.58 44.00 43.18 45.69 43.78 54.60 54.93 54.10
20% 4% 45.19 41.44 49.07 43.87 42.89 45.72 43.66 54.60 54.35 54.63
8% 45.17 41.03 49.56 43.73 42.59 45.74 43.52 54.59 53.80 55.16
0% 34.25 30.86 37.78 33.12 32.34 34.73 32.87 43.20 43.99 42.60
30% 4% 34.25 30.63 38.04 33.06 32.19 34.75 32.81 43.20 43.46 42.96
8% 34.25 30.40 38.31 32.99 32.04 34.77 32.74 43.21 42.97 43.45
that P x, 0, n, i0 , l0 1, where P x, 0, n, i0 , l0 is obtained using 4.17 for the first approach
or using 4.18 for the second approach by holding all other parameter values constant.
6.1.1. Term-End Point

In practice, various designs for Rt have been proposed. The term-end point design is the
simplest crediting method. It measures the index growth from the start to the end of a term.
The index on the day the contract is issued is taken as the starting index, and the index on the
day the policy matures or the time of death is taken as the ending index. Hence, the gain
provided by the point-to-point EIA with term-end point may expressed as
St
Rt 1. 6.2
S0
The EIA payo given in 6.1 is defined by

t
t t
Dt, it max min 1 l, il 1 , 1 , 1 g . 6.3
l 0
Tables 1 and 2 give the critical participation rates based on 5.1 and 5.2 for the
decomposed approach as well as 5.3 for the unified approach over dierent short-term rate
volatilities 0%, 4%, and 8%. The index volatility is set to either 20% or 30%. We present the
participation rates of 5-year EIA contracts with the term-end design without cap rate
in Table 1 and 15% cap rate in Table 2. We consider two types of minimum guarantees:
90% and 100% and both with g 3%.
The participation rates obtained for r t 0% are consistent with the corresponding
participation rates under deterministic interest rates presented in Gaillardetz and Lin 13.
Table 3: Annual reset with term-end point design for various interest rate volatilities.

CJ CJ
C0.1 G
C0.3 CI CU CD
0% 38.91 37.19 40.51 38.42 38.02 39.16 38.19 43.26 43.25 43.27
20% 4% 38.91 37.27 40.41 38.44 38.07 39.15 38.21 43.26 43.31 43.21
8% 38.91 37.35 40.32 38.46 38.12 39.14 38.23 43.27 43.37 43.17
0% 28.23 26.72 29.59 27.86 27.54 28.42 27.65 31.39 31.38 31.39
30% 4% 28.23 26.79 29.50 27.88 27.59 28.41 27.67 31.39 31.43 31.34
8% 28.23 26.87 29.41 27.89 27.63 28.41 27.69 31.39 31.50 31.30
0% 35.42 33.93 36.74 35.04 34.71 35.62 34.82 40.59 40.57 40.55
20% 4% 35.42 33.92 36.76 35.03 34.70 35.62 34.82 40.59 40.52 40.60
8% 35.41 33.90 36.77 35.02 34.69 35.61 34.81 40.59 40.49 40.66
0% 25.19 23.87 26.36 24.91 24.65 25.35 24.71 28.89 28.86 28.87
30% 4% 25.19 23.88 26.34 24.91 24.66 25.35 24.72 28.89 28.85 28.89
8% 25.19 23.89 26.32 24.91 24.67 25.34 24.72 28.90 28.85 28.91
As expected, the participation rates for the independent copulas decrease as the interest rate
volatility increases; however, this eect is negligible for 5-year contracts.
The independent copula may be obtained by letting 0 in the Cook-Johnson
copula. Similarly, the Frechet-Hoeding upper bound is obtained by letting . This
explains that the participation rates with t 0.5 are closer to the independent one than
the participation rates obtained using t 2, which are closer to the upper copula. Setting
t 0 in the Gaussian copula also leads to the independent copula. The participation rates
are between the independent copula and the lower copula when t 0.1. On the other
hand, when t 0.3 the participation rates are between the independent copula and the
upper copula.
The width of participation rate bands for the decomposed approach increases as the
short-term rate volatility increases. Here, the participation rate band represents the dierence
between the participation rates obtained from the lower copula and the upper copula. Indeed,
the participation rate for the upper copula 90% and 20% decreases from 56.27%
r t 0% to 55.44% r t 8%, meanwhile under the lower copula the participation rate
passes from 68.58% r t 0% to 69.75% r t 0%. This is due to the fact that increasing
r introduces more uncertainty in the model.
As we increase the volatility of the index, the participation rate decreases since a higher
volatility leads to more valuable embedded financial options. As expected, the participation
rates for 100% are lower than the corresponding values with 90%.
The dependence eects for the unified approach are negligible since there is a natural
hedging between the death and accumulation benefits. The introduction of stochastic
interest rates has more impact when 90% than when 100% because the participation
rates are higher. Although the participation rates are higher when 20%, the dependence
has relatively more impact if 30% because the model is more risky.
As expected, imposing a ceiling on the equity return that can be credited increases
the participation rates. Furthermore, the magnitude of the increments is more significant
Table 4: Annual reset with term-end point design and 15% cap rate for various interest rate volatilities.
Decomposed Approach Unified Approach

CJ CJ
C0.1 G
C0.3 CI CU CD
3% Minimum Guarantee on 90% Premium
0% 42.49 38.58 45.86 41.35 40.44 43.04 40.86 52.63 52.62 52.65
20% 4% 42.49 38.78 45.65 41.39 40.56 43.02 40.91 52.64 52.76 52.53
8% 42.49 38.98 45.44 41.44 40.69 43.00 40.97 52.66 52.91 52.44
0% 35.80 33.10 38.12 35.02 34.40 36.18 34.69 42.82 42.80 42.83
30% 4% 35.81 33.28 37.93 35.06 34.50 36.17 34.73 42.83 42.94 42.72
8% 35.81 33.47 37.75 35.09 34.60 36.16 34.78 42.84 43.09 42.62
3% Minimum Guarantee on 100% Premium
0% 35.42 33.93 37.69 35.04 34.71 35.62 34.82 47.46 47.45 47.35
20% 4% 35.42 33.92 37.73 35.03 34.70 35.62 34.82 47.46 47.36 47.46
8% 35.41 33.90 37.77 35.02 34.69 35.61 34.81 47.47 47.28 47.60
0% 30.40 27.60 32.72 29.73 29.17 30.75 29.36 39.41 39.45 39.34
30% 4% 30.40 27.63 32.66 29.74 29.19 30.75 29.37 39.42 39.45 39.36
8% 30.40 27.67 32.62 29.75 29.22 30.74 29.38 39.43 39.48 39.40
in a high volatility market. This is because the eect of the volatility diminishes as the cap
rate decreases and hence the behavior of the EIA payo is similar for dierent ranges of
volatilities. This is particularly observable when 90%.
6.2. Annual Reset

We now consider the most popular class of EIAs, known as the annual reset. They appeal to
investors because they oer similar features as the point-to-point class; however, the interest
credited to a annual reset EIA contract cannot be lost. This lock-in feature protects the
investor against a poor performance of the index over a particular year. The payo of this
type of EIA contracts is defined by
t
t
Dt max maxmin1 Rl , 1 , 1, 1 g , 6.4
l 1
where Rl represents the realized gain in year l, which varies from product to
product.
The cases where is set to 0 are known as annual reset EIAs and the cases where
> 0 are known as annual yield spreads. Furthermore, the participation levels in those cases
> 0 are typically 100%. As mentioned previously, in the case of annual reset, we fix 0
and determine the critical participation rate while fixing g, , and . In the traditional yield
spread needs to be determined such that the cost of EIA embedded options is covered by
the initial premium while fixing 100%, g, , and .
6.2.1. Term-End Point

In the case of annual reset EIA with term-end point, the index return is calculated each
year by comparing the indices at the beginning and ending policy anniversaries. Hence, the
participation rate may be expressed as
St
Rt 1. 6.5
St 1
In this case, the EIA payo given in 6.4 is defined by

t
t
Dt, it max max min 1 l, il 1 v, 1 , 1 , 1 g . 6.6
l 1
Tables 3 and 4 consider an annual reset EIA with term-end point design for various
cap rates and 15%. In this numerical illustration, we consider the same set of
parameters; particularly, the short-term rate volatility is either equal to 0%, 4%, or 8% and
the index volatility is set to either 20% or 30% with N 3. We find such that 4.17 for the
decomposed approach and 4.18 for the unified approach are equal to 1 by setting 0%.
The annual reset with term-end point design is more expensive than the point-to-
point with the term-end point design. The participation rates from the upper copula increase
while the ones from the lower copula decrease as r t increases for both approaches when
90%. This leads to narrower participation rate bands for the decomposed approach.
These behaviors are inverted for 100%. In that case, it leads to wider participation
rate bands for the decomposed approach. The imposition of a 15% cap rate may increase
the participation rate as much as 10%. However, there is no impact on the participation rate
under the decomposed approach when 20% and 100%.
7. Conclusions
The purpose of this paper is to generalize the approach presented by Gaillardetz and Lin
13 under stochastic interest rates. To this end, martingale probability measures for each of
the term life, pure endowment, and endowment insurances are introduced under stochastic
interest rates. Using the insurance market information, we obtain equity-linked martingale
measures that combined the insurance, interest rates, and index information. Although the
choice of copulas is somewhat arbitrary, with additional premium information from certain
equity-linked products, we would be able to narrow down the choices. We present two
dierent pricing approaches for equity-linked products. The first approach evaluates death
benefits and accumulation/survival benefits separately. In the second approach, we evaluate
the death benefits and the survival benefits in a unified manner by using the endowment
insurance products to define the martingale measure. A detailed numerical analysis is then
performed for existing EIAs in the North American market.
Our methodology may be used to evaluate variable annuities segregated fund
contracts in Canada because of the similarity in payo structure between EIAs and VAs.
Furthermore, our approach may also be used to evaluate Universal Life insurances, variable
Universal Life insurances, and others equity-linked products.
Acknowledgments
This research was supported by the Natural Sciences and Engineering Research Council of
Canada and a Ph. D. grant from the Casualty Actuarial Society and the Society of Actuaries.
The author is very grateful to Professor X. Sheldon Lin for his valuable comments and
suggestions.
Endnotes
1. Utility theory is also used to price standard life insurance products.
References
1 M. R. Hardy, Investment Guarantees: Modeling and Risk Management for Equity- Linked Life Insurance,
John Wiley & Sons, Hoboken, NJ, USA, 2003.
2 M. J. Brennan and E. S. Schwartz, The pricing of equity-linked life insurance policies with an asset
value guarantee, Journal of Financial Economics, vol. 3, pp. 195213, 1976.
3 P. Boyle and E. Schwartz, Equilibrium prices of equity linked insurance policies with an asset value
guarantee, Journal of Risk and Insurance, vol. 44, pp. 639660, 1977.
4 A. R. Bacinello and F. Ortu, Pricing guaranteed securities-linked life insurance under interest-rate
risk, in Proceedings of the 3rd AFIR International Colloquium on Actuarial Approach For Financial Risks,
pp. 3555, Rome, Italy, April 1993.
5 A. R. Bacinello and F. Ortu, Single and periodic premiums for guaranteed equitylinked life insurance
under interest-rate risk: the lognormal Vasicek case, in Financial Modelling, pp. 125, Physica,
Berlin, Germany, 1994.
6 T. Mller, Risk-minimizing hedging strategies for unit-linked life insurance contracts, Astin Bulletin,
vol. 28, no. 1, pp. 1747, 1998.
7 V. R. Young and T. Zariphopoulou, Pricing dynamic insurance risks using the principle of equivalent
utility, Scandinavian Actuarial Journal, no. 4, pp. 246279, 2002.
8 S. Tiong, Valuing equity-indexed annuities, North American Actuarial Journal, vol. 4, no. 4, pp. 149
163, 2000.
9 H. Lee, Pricing equity-indexed annuities with path-dependent options, Insurance: Mathematics &
10 K. S. Moore, Optimal surrender strategies for equity-indexed annuity investors, Insurance:
11 X. S. Lin and K. S. Tan, Valuation of equity-Indexed annuities under stochastic interest rate, North
12 M. Kijima and T. Wong, Pricing of ratchet equity-indexed annuities under stochastic interest rates,
13 P. Gaillardetz and X. S. Lin, Valuation of equity-linked insurance and annuity products with
binomial models, North American Actuarial Journal, vol. 10, no. 4, pp. 117144, 2006.
14 M.V. Wuthrich, H. Buhlmann, and H. Furrer, Market-Consistent Actuarial Valuation, EAA Lecture
Notes, Springer, Berlin, Germany, 2008.
15 P. Gaillardetz, Valuation of life insurance products under stochastic interest rates, Insurance:
16 H. H. Panjer, et al., Financial Economics with Applications to Investment, Insurance and Pensions, The
Actuarial Foundation, Schaumburg, Ill, USA, 1998.
17 X. S. Lin, Introductory Stochastic Analysis for Finance and Insurance, Wiley Series in Probability and
Statistics, John Wiley & Sons, Hoboken, NJ, USA, 2006.
18 F. Black, E. Derman, and W. Toy, A one-factor model of interest rates and its application to treasury
bond options, Financial Analysts Journal, vol. 46, pp. 3339, 1990.
19 J. Hull and A. White, Numerical procedures for implementing term structure models I: single-factor
models, Journal of Derivatives, vol. 2, no. 1, p. 716.
20 N. L. Bowers Jr., U. G. Hans, J. C. Hickman, D. A. Jones, and Nesbitt, Actuarial Mathematics, Society of
Actuaries, Schaumburg, Ill, USA, 2nd edition, 1997.
21 H. Joe, Multivariate Models and Dependence Concepts, vol. 73 of Monographs on Statistics and Applied
Probability, Chapman and Hall, London, UK, 1997.
22 R. B. Nelsen, An Introduction to Copulas, vol. 139 of Lecture Notes in Statistics, Springer, New York, NY,
USA, 1999.
23 E. W. Frees and E. A. Valdez, Understanding relationships using copulas, North American Actuarial
Journal, vol. 2, no. 1, pp. 125, 1998.
24 S. Wang, Aggregation of correlated risk portfolios: models and algorithms, in Proceedings of Casualty
Actuarial Society, pp. 848939, Arlington, Va, USA, 1998.
25 G. G. Venter, Tails of copulas, in Proceedings of Casualty Actuarial Society, pp. 68113, 2000.
26 G. Cherubini, E. Luciano, and W. Vecchiato, Copula Methods in Finance, Wiley Finance Series, John
Wiley & Sons, Chichester, UK, 2004.
27 P. Gaillardetz, Equity-linked annuities and insurances, Ph.D. thesis, University of Toronto, Toronto,
Ontario, Canada, 2006.
28 J. C. Cox, S. A. Ross, and M. Rubinstein, Option pricing: a simplified approach, Journal of Financial
doi:10.1155/2010/754851
Research Article
Local Likelihood Density Estimation and
Value-at-Risk
Christian Gourieroux1 and Joann Jasiak2

1
CREST and University of Toronto, Canada
2
York University, Canada
Correspondence should be addressed to Joann Jasiak, [email protected]
Received 5 October 2009; Accepted 9 March 2010
Copyright q 2010 C. Gourieroux and J. Jasiak. This is an open access article distributed under
This paper presents a new nonparametric method for computing the conditional Value-at-Risk,
based on a local approximation of the conditional density function in a neighborhood of a
predetermined extreme value for univariate and multivariate series of portfolio returns. For
illustration, the method is applied to intraday VaR estimation on portfolios of two stocks traded on
the Toronto Stock Exchange. The performance of the new VaR computation method is compared
to the historical simulation, variance-covariance, and J. P. Morgan methods.
1. Introduction
The Value-at-Risk VaR is a measure of market risk exposure for portfolios of assets. It has
been introduced by the Basle Committee on Banking Supervision BCBS and implemented
in the financial sector worldwide in the late nineties. By definition, the VaR equals the Dollar
loss on a portfolio that will not be exceeded by the end of a holding time with a given
probability. Initially, the BCBS has recommended a 10-day holding time and allowed for
computing the VaR at horizon 10 days by rescaling the VaR at a shorter horizon and loss
probability 1%; see, 1, page 3, Banks use the VaR to determine the required capital to
be put aside for coverage of potential losses. The required capital reserve is defined as

RCt MaxVaRt , M m1/60 60 h1 VaRth , see, 1, page 14 and 2, page 2, where M is
a multiplier set equal to 3, and m takes a value between 0 and 1 depending on the predictive
quality of the internal model used by the bank. The VaR is also used in portfolio management
and internal risk control. Therefore, some banks compute intradaily VaRs, at horizons of one
or two hours, and risk levels of 0.5%, or less.
Formally, the conditional Value-at-Risk is the lower-tail conditional quantile and

satisfies the following expression:
Pt xt1 < VaRt , 1.1
where xt is the portfolio return between t 1 and t, denotes the loss probability, and
Pt represents the conditional distribution of xt1 given the information available at time t.
Usually, the information set contains the lagged values xt , xt1 , . . . of portfolio returns. It can
also contain lagged returns on individual assets, or on the market portfolio.
While the definition of the VaR as a market risk measure is common to all banks,
the VaR computation method is not. In practice, there exist a variety of parametric,
semiparametric, and nonparametric methods, which dier with respect to the assumptions
on the dynamics of portfolio returns. They can be summarized as follows see, e.g., 3.
(a) Marginal VaR Estimation

The approach relies on the assumption of i.i.d. returns and comprises the following methods.
(1) Gaussian Approach

The VaR is the -quantile, obtained by inverting the Gaussian cumulative distribution
function
VaR Ext1 1 V xt1 1/2 , 1.2
where Ext1 is the expected return on a portfolio, V xt1 is the variance of portfolio returns,
and 1 is the -quantile of the standard normal distribution. This method assumes the
normality of returns and generally underestimates the VaR. The reason is that the tails of the
normal distribution are much thinner than the tails of an empirical marginal distribution of
portfolio returns.
(2) Historical Simulation (see [1])

VaR is approximated from a sample quantile at probability , obtained from historical
data collected over an observation period not shorter than one year. The advantage of this
method is that it relaxes the normality assumption. Its major drawback is that it provides poor
approximation of small quantiles at s such as 1%, for example, as extreme values are very
infrequent. Therefore, a very large sample is required to collect enough information about
the true shape of the tail. According to the asymptotic properties of the empirical quantile,
at least 200300 observations, that is, one year, approximately, are needed for 5% and at
least 1000, that is, 4 years are needed for 1%, both for a Gaussian tail. For fatter tails, even
more observations can be required see, e.g., the discussion in 3.
(3) Tail Model Building

The marginal quantile at a small risk level is computed from a parametric model of the
tail and from the sample quantiles at a larger . For example, McKinsey Inc. suggests to
infer the 99th quantile from the 95th quantile by multiplying the latter one by 1.5, which is
the weight based on a zero-mean Gaussian model of the tail. This method is improved by
considering two tail quantiles. If a Gaussian model with mean and variance is assumed

to fit the tail for < 10%, then the VaR, for any < 10%, can be calculated as follows, Let

VaR10% and VaR5% denote the sample quantiles at risk levels 5% and 10%. From 1.2,
the estimated mean and variance in the tail arise as the solutions of the system

VaR10% 1 10%
m ,
1.3

VaR5% 1 5%
m .
The marginal VaR at any loss probability less than 10% is calculated as

VaR 1
m , 1.4
are solutions of the above system. Equivalently, we get

where m,

VaR
VaR10% 1 1 10%
1 . 1.5

VaR5%
VaR10% 5% 1 10%

Thus, VaR
is a linear combination of sample quantiles VaR10%
and VaR5% with the
weights determined by the Gaussian model of the tail.
This method is parametric as far as the tail is concerned and nonparametric for the
central part of the distribution, which is left unspecified.
The marginal VaR estimation methods discussed so far do not account for serial
dependence in financial returns, evidenced in the literature. These methods are often applied
by rolling, that is, by averaging observations over a window of fixed length, which implicitly
assumes independent returns, with time dependent distributions.
(b) Conditional VaR Estimation

These methods accommodate serial dependence in financial returns.
(1) J. P. Morgan
The VaR at 5% is computed by inverting a Gaussian distribution with conditional mean zero
and variance equal to an estimated conditional variance of returns. The conditional variance
is estimated from a conditionally Gaussian IGARCH-type model of volatility t2 , called the
Exponentially Weighted Moving Average, where t2 t1 2
1 xt1
2
, and parameter is
arbitrarily fixed at 0.94 for any portfolio 4.
(2) CaViar [5]

The CaViar model is an autoregressive specification of the conditional quantile. The model is
estimated independently for each value of , and is nonparametric in that respect.
Table 1: Computation of the VaR.
Parametric Semi-parametric Nonparametric

i.i.d. Gaussian tail model building historical simulation
approach advisory firms regulators
Serial IGARCH by J. P. Morgan
CaViar
Dependence DAQ
(3) Dynamic Additive Quantile (DAQ) [6]

This is a parametric, dynamic factor model of the conditional quantile function.
Table 1 summarizes all the aforementioned methods.
This paper is intended to fill in the empty cell in Table 1 by extending the tail model
building method to the conditional Value-at-Risk. To do that, we introduce a parametric
pseudomodel of the conditional portfolio return distribution that is assumed valid in a
neighbourhood of the VaR of interest. Next, we estimate locally the pseudodensity, and use
this result for calculating the conditional VaRs in the tail.
The local nonparametric approach appears preferable to the fully parametric
approaches for two reasons. First, the nonparametric methods are too sensitive to
specification errors. Second, even if the theoretical rate of convergence appears to be smaller
than that of a fully parametric method under the assumption of no specification error in
the latter one, the estimator proposed in this paper is based on a local approximation of
the density in a neighborhood where more observations are available than at the quantile of
interest.
The paper is organized as follows. Section 2 presents the local estimation of a
probability density function from a misspecified parametric model. By applying this
technique to a Gaussian pseudomodel, we derive the local drift and local volatility, which
can be used as inputs in expression 1.2. In Section 3, the new method is used to compute
the intraday conditional Value-at-Risk for portfolios of two stocks traded on the Toronto Stock
Exchange. Next, the performance of the new method of VaR computation is compared to
other methods, such as the historical simulation, Gaussian variance-covariance method, J. P.
Morgan IGARCH, and ARCH-based VaR estimation in Monte Carlo experiments. Section 4
discusses the asymptotic properties of the new nonparametric estimator of the log-density
derivatives. Section 5 concludes the paper. The proofs are gathered in Appendices.
2. Local Analysis of the Marginal Density Function

The local analysis of a marginal density function is based on a family of pseudodensities.
Among these, we define the pseudodensity, which is locally the closest to the true density.
Next, we define the estimators of the local pseudodensity, and show the specific results
obtained for a Gaussian family of pseudodensities.
2.1. Local Pseudodensity

Let us consider a univariate or multivariate random variable Y , with unknown density f0 ,
and a parametric multivariate family of densities F {fy, , }, called the family of
pseudodensities where the parameter set Rp . This family is generally misspecified. Our
method consists in finding the pseudodensity fy; 0 , which is locally the closest to the true
density. To do that we look for the local pseudo-true value of parameter .
In the first step, let us assume that variable Y is univariate and consider an
approximation on an interval A c h, c h, centered at some value c of variable Y .
The pseudodensity is derived by optimizing the Kullback-Leibler criterion evaluated from
the pseudo and true densities truncated over A. The pseudo-true value of is

c,h Argmax E0 1ch<Y <ch log fY ;

E0 1ch<Y <ch log 1ch<y<ch f y; dy

2.1
1
Argmax E0 1ch<Y <ch log fY ;
2h

1 1
E0 1ch<Y <ch log 1ch<y<ch f y; dy,

2h 2h
where E0 denotes the expectation taken with respect to the true probability density function
pdf, henceforth f0 . The pseudo-true value depends on the pseudofamily, the true pdf, the
bandwidth, and the location c. The above formula can be equivalently rewritten in terms of
the uniform kernel Ku 1/211,1 u. This leads to the following extended definition of
the pseudo-true value of the parameter which is valid for vector Y of any dimension d, kernel
K, bandwidth h, and location c:

1 Y c
c,h Argmax E0 d K log fY ;
h h
2.2
1 Y c 1 yc
E0 d K log K f y; dy.
h h hd h
Let us examine the behavior of the pseudo-true value when the bandwidth tends to zero.
Definition 2.1. i The local parameter function (l.p.f.) is the limit of c,h when h tends to zero,
given by
c, f0 lim c,h , 2.3

h0
when this limit exists.
f0 .
ii The local pseudodensity is fy; c,
The local parameter function provides the set of local pseudo-true values indexed by
c, while the local pseudodensity approximates the true pdf in a neighborhood of c. Let us
now discuss some properties of the l.p.f.
Proposition 2.2. Let one assume the following:

A.1 There exists a unique solution to the objective function maximized in 2.2 for any h, and
f0 exists.
the limit c,

A.2 The
kernel K is continuous on Rd , of order 2, such that Kudu 1, uKudu
0, uu Kudu 2 , positive definite.
A.3 The density functions fy, and f0 y are positive and third-order dierentiable with
respect to y.
A.4 dim p d, and, for any c in the support of f0 ,

log fc;
, Rd . 2.4
y

A.5 For
h small and any c, the following integrals exist: Ku log fc uh; f0 c uhdu,
Kuf0 c uhdu, Kufc uh; du, and are twice dierentiable under the integral
sign with respect to h.
Then, the local parameter function is a solution of the following system of equations:

log f c; c; f0 log f0 c
, c. 2.5
y y
Proof . See Appendix A.
The first-order conditions in Proposition 2.2 show that functions fy, c, f0 and
f0 c have the same derivatives at c. When p is strictly larger than d, the first-order conditions
are not sucient to characterize the l.p.f.
Assumption A.1 is a local identification condition of parameter . As shown in
the application given later in the text, it is verified to hold for standard pseudofamilies of
densities such as the Gaussian, where c, f0 has a closed form. The existence of a limit
f0 is assumed for expository purpose. However, the main result concerning the first-
c,
order conditions is easily extended to the case when c,h exists, with a compact parameter
set . The proof in Appendix A shows that, even if the limh 0 c, f0 does not exist, we get

limh 0 log fc, c,h /y log f0 c/y, c. This condition would be sucient to define
a local approximation to the log-derivative of the density.
It is known that a distribution is characterized by the log-derivative of its density due
to the unit mass restriction. This implies the following corollary.
Corollary 2.3. The local parameter function characterizes the true distribution.
2.2. Estimation of the Local Parameter Function and

of the Log-Density Derivative
Suppose that y1 , . . . , yT are observations on a strictly stationary process Yt of dimension d.
Let us denote by f0 the true marginal density of Yt and by {fy : , } a misspecified
pseudoparametric family used to approximate f0 . We now consider the l.p.f. characterization

of f0 , and introduce nonparametric estimators of the l.p.f. and of the marginal density.
The estimator of the l.p.f. is obtained from formula 2.2, where the theoretical
expectations are replaced by their empirical counterparts:

T
1 yt c
T c Argmax d
K log f yt ;
t1 h
h
2.6

T
1 yt c 1 yc
d
K log K f y; dy .
t1 h
h hd h
The above estimator depends on the selected kernel and bandwidth. This estimator allows us
to derive from Proposition 2.2 a new nonparametric consistent estimator of the log-density
derivative defined as

log f c; T c

log fT c 2.7
.
y y
The asymptotic properties of the estimators of the l.p.f. and log-density derivatives are
discussed in Section 4, for the exactly identified case p d. In that case, T c is characterized
by the system of first-order conditions 2.7.
The quantity fc, T c is generally a nonconsistent estimator of the density f0 c
at c see, e.g., 7 for a discussion of such a bias in an analogous framework . However,
a consistent estimator of the log-density and thus of the density itself, obtained as the
exponential function of the log-density is derived directly by integrating the estimated log-
density derivatives under the unit mass restriction. This oers a correction for the bias, and
is an alternative to including additional terms in the objective function see, e.g., 7, 8.
2.3. Gaussian Pseudomodel

A Gaussian family is a natural choice of pseudomodel for local analysis, as the true density
is locally characterized by a local mean and a local variance-covariance matrix. Below, we
provide an interpretation of the Gaussian local density approximation. Next, we consider a
Gaussian pseudomodel parametrized by the mean only, and show the relationship between
the l.p.f. estimator and two well-known nonparametric estimators of regression and density,
respectively.
(i) Interpretation
For a Gaussian pseudomodel indexed by mean m and variance , we have

log f y; m,
1 y m . 2.8
y
Thus, the approximation associated with a Gaussian pseudofamily is the standard

one, where the partial derivatives of the log-density are replaced by a family of hyperplanes
parallel to the tangent hyperplanes. These tangent hyperplanes are not independently
defined, due to the Schwartz equality

2 log f y 2 log f y
; i
/ j. 2.9
yi yj yj yi
The Schwartz equalities are automatically satisfied by the approximated densities

because of the symmetry of matrix 1 .
(ii) Gaussian Pseudomodel Parametrized by the Mean and Gaussian Kernel

Let us consider a Gaussian kernel: K of dimension d, where denotes the pdf of the
standard Normal N0, Id.
Proposition 2.4. The l.p.f. estimator for a Gaussian pseudomodel parametrized by the mean and with
a Gaussian kernel can be written as
1 h2
2 log
T c c m T c c c 1 h fT c, 2.10
h2 c
where

T

d
t1 1/h yt c /h yt
T c
m

2.11
T
t1 1/h yt c /h
d
is the Nadaraya-Watson estimator of the conditional mean mc EY | Y c c, and

1 T
1 yt c
fT c 2.12
T t1 hd h
is the Gaussian kernel estimator of the unknown value of the true marginal pdf at c.
Proof. See Appendix B.
In this special case, the asymptotic properties of T c follow directly from the
asymptotic properties of fT c and fT c/c 9. In particular, T c converges to c
log f0 c/y, when T and h tend to infinity and zero, respectively, with T hd2 0.
Alternatively, the asymptotic behavior can be inferred from the Nadaraya-Watson
estimator 10, 11 in the degenerate case when the regressor and the regressand are identical.
Section 5 will show that similar relationships are asymptotically valid for non-Gaussian
pseudofamilies.
2.4. Pseudodensity over a Tail Interval

Instead of using the local parameter function and calibrating the pseudodensity locally
about a value, one could calibrate the pseudodensity over an interval in the tail. We thank
an anonymous referee for this suggestion. More precisely, we could define a pseudo-true
parameter value

f y;
c, f0 Argmax E0 1Y >c log , 2.13
Sc,
where S denotes the survival function, and consider an approximation of the true distribution
over a tail interval fy; c, f0 , for y > c. From a theoretical point of view, this approach
can be criticized as it provides dierent approximations of f0 y depending on the selected
value of c, c < y.
3. From Marginal to Conditional Analysis

Section 2 described the local approach to marginal density estimation. Let us now show
the passage from the marginal to conditional density analysis and the application to the
conditional VaR.
3.1. General Approach to VaR Computation

The VaR analysis concerns the future return on a given portfolio. Let xt denote the return
on that portfolio at date t. In practice, the prediction of xt is based on a few summary
statistics computed from past observations, such as a lagged portfolio return, realized market
volatility, or realized idiosyncratic volatility in a previous period. The application of our

method consists in approximating locally the joint density of series yt y1t , y2t , whose
component y1t is xt , and component y2t contains the summary statistics, denoted by zt1 .
Next, from the marginal density of yt , that is, the joint density of y1t and y2t , we derive the
conditional density of y1t given y2t , and the conditional VaR.
The joint density is approximated locally about c which is a vector of two components,
c c1 , c2 . The first component c1 is a tail value of portfolio returns, such as the 5% quantile
of the historical distribution of portfolio returns, for example, if the conditional VaR at < 5%
needs to be found. The second component c2 is the value of the conditioning set, which is
fixed, for example, at the last observed value of the summary statistics in y2t zt1 . Due to
the dierence in interpretation, the bandwidths for c1 and c2 need to be dierent.
The approach above does not suer from the curse of dimensionality. Indeed, in
practice, y1 is univariate, and the number of summary statistics is small often less than 3,
while the number of observations is suciently large 250 per year for a daily VaR.
3.2. Gaussian Pseudofamily

When the pseudofamily is Gaussian, the local approximation of the density of yt is

characterized by the local mean and variance-covariance matrix. For yt y1t , y2t , these
moments are decomposed by blocks as follows:

1 c 11 c 12 c
c , c . 3.1
2 c 21 c 22 c
The local conditional first and second-order moments are functions of these joint moments:
1|2 c 1 c 12 c1

22 c2 c, 3.2
1|2 c 11 c 12 c1

22 c21 c. 3.3
When y1t xt is univariate, these local conditional moments can be used as inputs in the
basic Gaussian VaR formula 1.2.
The method is convenient for practitioners, as it suggests them to keep using the
misspecified Gaussian VaR formula. The only modifications are the inputs, which become
the local conditional mean and variance in the tail that are easy to calculate given the closed-
form expressions given above.
Even though the theoretical approach is nonparametric, its practical implementation
is semi-parametric. This is because, once an appropriate location c has been selected, the local
pseudodensity estimated at c is used to calculate any VaR in the tail. Therefore, the procedure
can be viewed as a model building method, in which the two benchmark loss probabilities are
arbitrarily close. As compared with other model building approaches, it allows for choosing
a location c with more data-points in its neighborhood than the quantile of interest.
4. Application to Value-at-Risk
The nonparametric feature of our localized approach requires the availability of a sucient
number of observations in a neighborhood of the selected c. This requirement is easily
satisfied when high-frequency data are used and an intraday VaR is computed. We first
consider an application of this type. It is followed by a Monte-Carlo study, which provides
information on the properties of the estimator when the number of observations is about 200,
which is the sample size used in practice for computing the daily VaR.
4.1. Comparative Study of Portfolios

We apply the local conditional mean and variance approach to intraday data on financial
returns and calculate the intraday Value-at-Risk. The financial motivation for intraday risk
analysis is that internal control of the trading desks and portfolio management is carried
out continuously by banks, due to the use of algorithmic trading that implements automatic
portfolio management, based on high-frequency data. Also, the BCBS in 2, page 3, suggests
that a weakness of the current daily risk measure is that it is based on the end-of-day
positions, and disregards the intraday trading risk. It is known that intraday stock price
variation can be often as high as the variation of the market closure prices over 5 to 6
consecutive days.
Our analysis concerns two stocks traded on the Toronto Stock Exchange: the Bank
of Montreal BMO and the Royal Bank ROY from October 1st to October 31, 1998, and
all portfolios with nonnegative allocations in these two stocks. This approach under the no-
short-sell constraint will suce to show that allocations of the least risky portfolios dier,
depending on the method of VaR computation.
From the tick-by-tick data, we select stock prices at a sampling interval of two minutes,
and compute the two minute returns xt x1t , x2t . The data contain a large proportion of
zero price movements, which are not deleted from the sample, because the current portfolio
values have to be computed from the most recent trading prices.
The BMO and ROY sample consists of 5220 observations on both returns from October
1 to October 31, 1998. The series have equal means of zero. The standard deviations are 0.0015
and 0.0012 for BMO and ROY, respectively. To detect the presence of fat tails, we calculate the
kurtosis, which is 5.98 for BMO and 3.91 for ROY, and total range, which is 0.0207 for BMO
and 0.0162 for ROY. The total range is approximately 50 for BMO and 20 for ROY times
greater than the interquartile range, equal to 0.0007 in both samples.
The objective is to compute the VaR for any portfolio that contains these two assets.
Therefore, yt y1t , y2t has two components; each of which is a bivariate vector. We are
interested in finding a local Gaussian approximation of the conditional distribution of y1t xt
given y2t xt1 in a neighborhood of values c1 c11 , c12 of xt and c2 c21 , c22 of
xt1 which does not mean that the conditional distribution itself is Gaussian . We fix c21
c22 0. Because a zero return is generally due to nontrading, by conditioning on zero past
returns, we investigate the occurrence of extreme price variations after a non-trading period.
As a significant proportion of returns is equal to zero, we eliminate smoothing with respect
to these conditioning values in our application.
The local conditional mean and variance estimators were computed from formulae
3.2-3.3 for c11 0.00188 and c12 0.00154, which are the 90% upper percentiles of the
sample distribution of each return on the dates preceded by zero returns. The bandwidth for
xt was fixed at h 0.001, proportionally to the dierence between the 10% and 1% quantiles.
The estimates are
1 6.54 103 , 2 0.48 103 ,

12 4.1
11 10.2 106 , 22 1.33 106 , 0.0034.
11 22
They can be compared to the global conditional moments of the returns, which are the
moments computed from the whole sample, Ext | xt1 0, V xt | xt1 0. Their
estimates are
1 2.057 105 , 2 1.359 104 ,
12 4.2
11 2.347 106 , 22 1.846 106 , 0.0976.
11 22
As the conditional distribution of xt given xt1 0 has a sharp peak at zero, it comes as no
surprise that the global conditional moments estimators based on the whole sample lead to
smaller Values-at-Risk than the localized ones. More precisely, for loss probability 5% and a
portfolio with allocations a, 1 a, 0 a 1, in the two assets, the Gaussian VaR is given by
1/2
VaR5%, a a, 1 a 1.64 a, 1 aa, 1 a , 4.3
and determines the required capital reserve for loss probability 5%. Figure 1 presents the
Values-at-Risk computed from the localized and unlocalized conditional moments, for any
admissible portfolios of nonnegative allocations. The proportion a invested in the BMO is
measured on the horizontal axis.
0.012
0.01
0.008
0.006
0.004
0.002
0 0.2 0.4 0.6 0.8 1
Unlocalized
Localized
Figure 1: Localized and Unlocalized VaRs.
As expected, the localized VaR lies far above the unlocalized one. This means that
the localized VaR implies a larger required capital reserve. We also note that, under the
unlocalized VaR, the least risky portfolio contains equal allocations in both assets. In contrast,
the localized measure suggests to invest the whole portfolio in a single asset to avoid extreme
risks under the no-short-sell constraint.
4.2. Monte-Carlo Study

The previous application was based on a quite large number of data more than 5000 on
trades in October 1998 and risk level of 5%. It is natural to assess the performance of the new
method in comparison to other methods of VaR computation, for smaller samples, such as
200 resp. 400 observations that correspond to one year resp., two years of daily returns
and for a smaller risk level of 1%.
A univariate series of 1000 simulated portfolio returns is generated from an ARCH1
model, with a double exponential Laplace error distribution. More precisely, the model is
1/2
xt 0.4 0.95xt1
2
ut , 4.4
where the errors ut are i.i.d. with pdf
1
gu exp|u|. 4.5
2
The error distribution has exponential tails that are slightly heavier than the tails of a
Gaussian distribution. The data generating process are assumed to be unknown to the person
who estimates the VaR. In practice, that person will apply a method based on a misspecified
model such as the i.i.d. Gaussian model of returns in the Gaussian variance-covariance
method or the IGARCH model of squared returns by J. P. Morgan with an ad-hoc fixed
parameter 0.94. Such a procedure leads to either biased, or inecient estimators of the VaR
level.
The following methods of VaR computation at risk level of 1% are compared. Methods
1 to 4 are based on standard routines used in banks, while method 5 is the one proposed in
this paper.
1 The historical simulation based on a rolling window of 200 observations. We will
see later Figure 2 that this approach results in heavy smoothing with respect to time. A
larger bandwidth would entail even more smoothing.
2 The Gaussian variance-covariance approach based on the same window.
3 The IGARCH-based method by J. P. Morgan:

t 1 1%0.06
VaR 0.94h xth
2
. 4.6
h0
4 Two conditional ARCH-based procedures that consist of the following steps. First,
we consider a subset of observations to estimate an ARCH1 model:
1/2
xt a0 a1 xt1
2
vt , 4.7
where vt are i.i.d. with an unknown distribution. First, the parameters a0 and a1 are estimated
by the quasi-maximum likelihood, and the residuals are computed. From the residuals we
infer the empirical 1% quantile q, say. The VaR is computed as VaR t 1/2
a0 a1 xt2 q.
We observe that the ARCH parameter estimators are very inaccurate, which is due to the
exponential tails of the error distribution. Two subsets of data were used to estimate the
ARCH parameters and the 1%-quantile. The estimator values based on a sample of 200
observations are a0 8.01, a1 0.17, and q 3.85. The estimator values based on a sample
of 800 observations are a0 4.12, a1 0.56, and q 2.78. We find that the ratios a1 /a0 are
quite far from the true value 0.95/0.4 used to generate the data, which is likely due to fat tails.
5 Localized VaR.
We use a Gaussian pseudofamily, a Gaussian kernel, and two dierent bandwidths for
the current and lagged value of returns, respectively. The bandwidths were set proportional
to the dierence between the 10% and 1% quantiles, and the bandwidth for the lagged return
is 4 times the bandwidth for the current return. Their values are 1.16 and 4.64, respectively.
We use a Gaussian kernel resp., a simple bandwidth instead of an optimal kernel resp.,
an optimal bandwidth for the sake of robustness. Indeed, an optimal approach may not be
suciently robust for fixing the required capital. Threshold c is set equal to the 3%-quantile
of the marginal empirical distribution. The localized VaRs are computed by rolling with a
window of 400 observations.
For each method, Figures 2, 3, 4, 5, 6 and 7 report the evolution of the true VaR
corresponding to the data generating model along with the evolution of the estimated VaR.
For clarity, only 200 data points are plotted.
The true VaR features persistence and admits extreme values. The rolling methods
such as the historical simulation and variance-covariance method produce stepwise patterns
of VaR, as already noted, for example, by Hull and White 12. These patterns result from the
i.i.d. assumption that underlies the computations. The J. P. Morgan IGARCH approach creates
spurious long memory in the estimated VaR and is not capable to recover the dynamics of
the true VaR series. The comparison of the two ARCH-based VaRs shows that the estimated
60
VaR
40
20
0
0 50 100 150 200
Time
Figure 2: True and Historical Simulation-Based VaRs.
60
VaR
40
20
0
0 50 100 150 200
Time
Figure 3: True and Variance-Covariance Based VaRs.
60
VaR
40
20
0
0 50 100 150 200
Time
Figure 4: True and IGARCH-Based VaR.

60
VaR
40
20
0
0 50 100 150 200
Time
Figure 5: True and ARCH-Based VaR 200 obs.
60
VaR
40
20
0
0 50 100 150 200
Time
Figure 6: True and ARCH-Based VaR 800 obs.
60
VaR
40
20
0
0 50 100 150 200
Time
Figure 7: True and Localized VaR.

paths strongly depend on the estimated ARCH coecients. When the estimators are based
on 200 observations, we observe excess smoothing. When the estimators are based on 800
observations, the model is able to recover the general pattern, but overestimates the VaR
when it is small and underestimates the VaR when it is large. The outcomes of the localized
VaR method are similar to the second ARCH model, with a weaker tendency to overestimate
the VaR when it is small.
The comparison of the dierent approaches shows the good mimicking properties of
the ARCH-based methods and of the localized VaR. However, these methods need also to be
compared with respect to their tractability. It is important to note that the ARCH parameters
were estimated only once and were kept fixed for future VaR computations. The approach
would become very time consuming if the ARCH model was reestimated at each point in
time. In contrast, it is very easy to regularly update the localized VaR.
5. Properties of the Estimator of the Local Parameter Function

5.1. Asymptotic Properties
In this section, we discuss the asymptotic properties of the local pseudomaximum likelihood
estimator under the following strict stationarity assumption.
Assumption 5.1. The process Y Yt is strictly stationary, with marginal pdf f0 .
Let us note that the strict stationarity assumption is compatible with nonlinear
dynamics, such as in the ARCH-GARCH models, stochastic volatility models, and so forth,
All proofs are gathered in Appendices.
The asymptotic properties of the local P. M. L. estimator of are derived along
the following lines. First, we find the asymptotic equivalents of the objective function and
estimator, that depend only on a limited number of kernel estimators. Next, we derive the
properties of the local P. M. L. estimator from the properties of these basic kernel estimators.
As the set of assumptions for the existence and asymptotic normality of the basic kernel
estimators for multivariates dependent observations can be found in the literature see the
study by Bosq in 13, we only list in detail the additional assumptions that are necessary
to satisfy the asymptotic equivalence. The results are derived under the assumption that
is exactly identified see Assumptions 5.2 and 5.3. In the overidentified case p > d,
the asymptotic analysis can be performed by considering the terms of order h3 , h4 in the
expansion of the objective function see Appendix A, which is out of the scope of this paper.
Let us introduce the additional assumptions.
Assumption 5.2. The parameter set is a compact set and p d.
f0 of the system of equations:

Assumption 5.3. i There exists a unique solution c;
log fc; log f0 c

, 5.1
y y
and this solution belongs to the interior of .

f0 /y is nonsingular.
ii The matrix 2 log fc, c,
Assumption 5.4. The following kernel estimators are strongly consistent:
T a.s.
i 1/T t1 1/h Kyt
d
c/h f0 c,
T
c/hyt c/hyt c/h 2 f0 c,
a.s.
ii 1/T t1 1/h Kyt
d
T a.s.
iii 1/T h t1 1/h Kyt
d
c/hyt c/h f0 c2 log f0 c/y.
Assumption 5.5. In any neighbourhood of , the third-order derivatives 3 log fy; /

yi yj yk , i, j, k varying, are dominated by a function ay such that y3 ay is integrable.
Proposition 5.6. The local pseudomaximum likelihood estimator T c exists and is strongly
f0 under Assumptions 5.15.5.
consistent for the local parameter function c;
Proof. See Appendix C.
It is possible to replace the set of Assumptions 5.4 by sucient assumptions

concerning directly the kernel, the true density function f0 , the bandwidth h, and the Y
process. In particular it is common to assume that the process Y is geometrically strong
mixing, and that h 0, T hd /log T 2 , when T tends to infinity see 1315.
Proposition 5.7. Under Assumptions 5.15.5 the local pseudomaximum likelihood estimator is

asymptotically equivalent to the solution T c of the equation:

log f c; T c 1 1 5.2
2 T c c,
m
y h2
where:
T

t1 K yt c /h yt
T c T
m

5.3
t1 K yt c /h
is the Nadaraya-Watson estimator of mc EY | Y c c based on the kernel K.
Proof. See Appendix D.
Therefore the asymptotic distribution of T c may be derived from the properties of
m T c c, which are the properties of the Nadaraya-Watson estimator in the degenerate case
when the regressand and the regressor are identical. Under standard regularity conditions
13, the numerator and denominator of 1/h2 m T c c have the following asymptotic
properties.
Assumption 5.8. If T , h 0, T hd2 , and T hd4 0, we have the limiting

distribution

1
d2
T
yt c
f0 c
Th K yt c 2

T hd2 t1 h y

T
yt c
d 1
Th K f0 c
T hd t1 h 5.4

uu K 2 udu uK 2 udu

d
N 0, f0 c .

2
u K udu K udu
2
The formulas of the first- and second-order asymptotic moments are easy to verify see
Appendix E. Assumption 5.8 is implied by sucient conditions concerning the kernel, the
process... see, 13. In particular it requires some conditions on the multivariate distribution
of the process such as supt1 <t2 ft1 ,t2 f f < , where ft1 ,t2 denotes the joint p.d.f. of Yt1 , Yt2
and f f the associated product of marginal distributions, and supt1 <t2 <t3 <t4 ft1 ,t2 ,t3 ,t4 < ,
where ft1 ,t2 ,t3 ,t4 denotes the joint p.d.f of Yt1 , Yt2 , Yt3 , Yt4 . Note that the rate of convergence
of the numerator is slower than the rate of convergence of the denominator since we study
a degenerate case, when the Nadaraya-Watson estimator is applied to a regression with the
regressor equal to the regressand.
We deduce that the asymptotic distribution of
*
1 log f0 c
T hd2 2
T c c 2
m 5.5
h y
is equal to the asymptotic distribution of

yt c
2 f0 c
T
1 1
T hd2 K yt c , 5.6
f0 c T hd2 t1 h y

which is N0, 1/f0 c uu K 2 udu.
By the -method we find the asymptotic distribution of the local pseudomaximum
likelihood estimator and the asymptotic distribution of the log-derivative of the true p.d.f..
Proposition 5.9. Under Assumptions 5.15.8 one has the following.

(i)

log f c; T c 2
1
log f0 c 1
T hd2 d 2
N 0, f c uu K udu
2
. 5.7
y y 0
(ii)

1 1
* 2
log f c; c; fo

2
T hd2 T c c; f0
d
N 0,
y f0 c
5.8

1
1 log f c; c; f0
2

uu K 2 udu 2
y
The first-order asymptotic properties of the estimator of the log-derivative of the

density function do not depend on the pseudofamily, whereas the value of the estimator
does. It is beyond the scope of this paper to discuss the eect of the pseudofamily when
dimension p is strictly larger than d. Nevertheless, by analogy to the literature on local
estimation of nonparametric regression and density functions see, e.g., the discussion in 7,
we expect that the finite sample bias in the associated estimator of the density will diminish
when the pseudofamily is enlarged, that is, when the dimension of the pseudoparameter
vector increases. For a univariate proces yt , the functional estimator of the log-derivative
log f0 c/y may be compared to the standard estimator
T

log f0 c t1 1/hK yt c /h

T

, 5.9
t1 1/hK yt c /h
y
where K is the derivative of the kernel of the standard estimator. The standard estimator has
a rate of convergence equal to that of the estimator introduced in this paper and the following
asymptotic distribution:
*
log f0 c log f0 c d 1 2
Th3
N 0, Ku du . 5.10
y y 4 f0 c
The asymptotic distributions of the two estimators of the log-derivative of the density
function are in general dierent, except when |dKu/du| |uKu|, which, in particular,
arises when the kernel is Gaussian. In such a case the asymptotic distributions of the
estimators are identical.
5.2. Asymptotic versus Finite Sample Properties

In kernel-based estimation methods, the asymptotic distributions of estimators do not
depend on serial dependence and are computed as if the data were i.i.d. However, serial
dependence aects the finite sample properties of estimators and the accuracy of the
theoretical approximation. Pritsker 16 see also work by Conley et al. in 17 illustrates
this point by considering the finite sample properties of Ait-Sahalias test of continuous time
model of the short-term interest rate 18 in an application to data generated by the Vasicek
model.
The impact of serial correlation depends on the parameter of interest, in particular on

whether this parameter characterizes the marginal or the conditional density. This problem
is not specific to the kernel-based approaches, but arises also in other methods such as the
OLS. To see that, consider a simple autoregressive model yt yt1
t , where
t is IIN0, 2 .
The expected value of yt is commonly estimated from the empirical mean m yT that has
asymptotic variance V m /T 1/11, where V yt /1 . In contrast, the
2 2 2 2
2
autoregressive coecient is estimated by t yt yt1 / t yt1 and has asymptotic variance
V 1 2 /T .
If serial dependence is disregarded, both estimators m and have similar asymptotic
eciencies that are 2 /T and 1/T , respectively. However, when tends to one while 2
remains fixed, the variance of m tends to infinity whereas the variance of tends to zero.
This simple example shows that omission of serial dependence does not have the same eect
on the marginal parameters as opposed to the conditional ones. Problems considered by
Conley et al. 17 or Pritsker 16 concern the marginal long run distribution of yt , while
our application is focused on a conditional parameter, which is the conditional VaR. This
parameter is derived from the analysis of the joint pdf fyt , yt1 as in the previous example
2
was derived from the bivariate vector 1/T t yt yt1 , 1/T t yt1 . Due to cointegration
between yt and yt1 in the case of extreme persistence, we can reasonably expect that the
estimator of the conditional VaR has good finite sample properties, even when the point
estimators fy t , yt1 do not. The example shows that in finite sample the properties of the
estimator of a conditional parameter can be even better than those derived under the i.i.d.
assumption.
6. Conclusions
This paper introduces a local likelihood method of VaR computation for univariate or
multivariate data on portfolio returns. Our approach relies on a local approximation of the
unknown density of returns by means of a misspecified model. The method allows us to
estimate locally the conditional density of returns, and to find the local conditional moments,
such as a tail mean and tail variance. For a Gaussian pseudofamily, these tail moments can
replace the global moments in the standard Gaussian formula used for computing the VaRs.
Therefore, our method based on the Gaussian pseudofamily is convenient for practitioners, as
it justifies computing the VaR from the standard Gaussian formula, although with a dierent
input, which accommodates both the thick tails and path dependence of financial returns.
The Monte-Carlo experiments indicate that tail-adjusted VaRs are more accurate than other
VaR approximations used in the industry.
Appendices
A. Proof of Proposition 2.2
Let us derive the expansion of the objective function

1 Y c
Ah E0 K log fY ;
hd h
A.1
1 Y c 1 yc
E0 d K log d
K f y; dy,
h h h h
when h approaches zero. By using the equivalence see Assumption A.1

1 yc
h2 2 gc 2
K g y dy gc Tr o h2 , A.2
hd h 2 yy
where Tr is the trace operator, we find that

h2 2 2
Ah f0 c log fc; Tr f0 c log fc;
2 yy

2 f0 c
2 2
h2 h2 2 fc;
f0 c Tr
log fc; Tr
o h2
2 yy 2 yy
h2 log f0 c 2 log f0 c

2 y y

h2 log fc; log f0 c 2 log fc; log f0 c
o h2 .
2 y y y y
A.3
The result follows.

The expansion above provides a local interpretation of the asymptotic objective
function at order h2 as a distance between the first-order derivatives of the logarithms of
the pseudo and true pdfs. In this respect the asymptotic objective function clearly diers
from the objective function proposed by Hjort and Jones 7, whose expansion defines an
l2 -distance between the true and pseudo pdfs.
B. Proof of Proposition 2.4

For a Gaussian kernel K of dimension d, we get

1 yc
1 cy
f y; dy y dy
hd h hd h
B.1

1 c
d/2
.
1 h2 1 h2
We have
+ +
T
1 yt c d +yt +2
T c Argmax d
log 2
t1 h
h 2 2

T
1 yt c d d c 2
log 2 log 1 h 2
t1 h
d h 2 2 21 h2
B.2
T

1 h2 t1 1/hd yt c /h yt 1
2 T

2c
t1 1/h yt c /h
h d h
1 h2 1
2
T c 2 c.
m
h h
Moreover we have
1 h2
T c c T c c
m
h2
T

1 h2 t11/hd yt c yt c /h
T

h2 t1 1/h
d y c /h
t
B.3

/c Tt1 1/hd yt c /h
1 h2

T d y c /h
t1 1/h t
log
1 h2 fT c.
c
C. Consistency
Let us consider the normalized objective function

1 T
1 yt c
AT,h 2 d
K log f yt ;
T h t1 h h
C.1

T
1 yt c 1 yc
d
K log K f y; dy .
t1 h
h hd h
It can be written as

T,h 1
T
1 yt c
A K
T h2 t1 hd h

log fc;
1 yt c 2 log fc; yt c
log fc; yt c
y 2 h yy h
C.2

+ +3 1 T
1 yt c
R1 yt ; +yt c+ K
T h2 t1 hd h

2
h2 1 2 fc;
log fc; Tr h R2 , h
3
2 fc; yy
where R1 yt ; , R2 ; h are the residual terms in the expansion. We deduce:

T,h 1
T
1 yt c log fc;
A 2 d
K
yt c
T h t1 h h y

1 T
1 yt c 1 yt c 2 log fc; yt c
K C.3
T t1 hd h 2 h yy h

1 1 T
1 yt c 1 2
2 fc;
K Tr residual terms.
2 T t1 hd h fc; yy
Under the assumptions of Proposition 5.7, the residual terms tend almost surely to zero,
uniformly on , while the main terms tend almost surely uniformly on to
log f0 c 2 log f0 c

A
y y
C.4
1 log fc; log f0 c 2 log fc; log f0 c
,
2 y y y y
which is identical to limh o Ah /h2 see Appendix A.

Then, by Jennrich theorem 19 and the identifiability condition, we conclude that the
estimator T c exists and is strongly consistent of c; f0 .
D. Asymptotic Equivalence
The main part of the objective function may also be written as

T,h 1
T
1 yt c log fc;
A 2 d
K
yt c
T h t1 h h y
D.1

1 T
1 yt c log fc; 2 log fc;
K .
T t1 hd h y y
We deduce that the local parameter function can be asymptotically replaced by the solution
T c of

log f c; T c 1 1/T h2 T

t1 1/h K yt c /h yt c
d
D.2
2

.
y 1/T Tt1 1/hd K yt c /h
E. The First- and Second-Order Asymptotic Moments

T
Let us restrict the analysis to the numerator term 1/T hd2 t1 KYt c/hYt c, which
implies the nonstandard rate of convergence.
(1) First-Order Moment

We get

1T
Yt c 1 Yt c
E K Y t c E K Y t c
T hd2 t1 h hd2 h

1 yc

d2 K y c f0 y dy
h h

1
Ku u f0 c uhdu
h

1 f0 c h2 2 f0 c
Kuu f0 c h u u uo h2
du
h y 2 yy

f0 c h 2 f0 c
2 Kuuu udu oh.
y 2 yy
E.1
(2) Asymptotic Variance

We have

1 T
Yt c 1 Yt c
V K Y t c V K Y t c
T hd2 t1 h T h2d4 h

1 2 Yt c
E K Y t cY t c
T h2d4 h

Yt c Yt c
E K Yt c E K Yt c
h h

1 d2 2 f0 c f0 c 2
K 2
uuu f0 c uhdu h
T hd2 y y

1 2 1
f0 c uu K udu o ,
T hd2 T hd2
E.2
1/2
which provides the rate of convergence T hd2 of the standard error. Moreover the
1/2
second term of the bias will be negligible if hT hd2 0 or T hd4 0.
(3) Asymptotic Covariance

Finally we have also to consider:

1 T
Yt c 1 T
Yt c
Cov K Yt c, K
T hd2 t1 h T h t1 h

1 Yt c Yt c
Cov K Y t c, K
T hd3 h h

1 2 Yt c Yt c Yt c E.3
E K Y t c E K Y t c EK
T hd3 h h h

1
d3
h2
K 2
uuf 0 c uhdu O h4
Th

1 1
f0 c uK udu o
2
.
T hd1 T hd1
Acknowledgment
The authors gratefully acknowledge financial support of NSERC Canada and of the Chair
AXA/Risk Foundation: Large Risks in Insurance.
References
1 Basle Committee on Banking Supervision, An Internal Model-Based Approach to Market Risk Capital
Requirements, Basle, Switzerland, 1995.
2 Basle Committee on Banking Supervision, Overview of the Amendment to the Capital Accord to Incorporate
Market Risk, Basle, Switzerland, 1996.
3 C. Gourieroux and J. Jasiak, Value-at-risk, in Handbook of Financial Econometrics, Y. Ait-Sahalia and
L. P. Hansen, Eds., pp. 553609, Elsevier, Amsterdam, The Netherlands, 2009.
4 J. P. Morgan, RiskMetrics Technical Manual, J.P. Morgan Bank, New York, NY, USA, 1995.
5 R. F. Engle and S. Manganelli, CAViaR: conditional autoregressive value at risk by regression
quantiles, Journal of Business & Economic Statistics, vol. 22, no. 4, pp. 367381, 2004.
6 C. Gourieroux and J. Jasiak, Dynamic quantile models, Journal of Econometrics, vol. 147, no. 1, pp.
198205, 2008.
7 N. L. Hjort and M. C. Jones, Locally parametric nonparametric density estimation, The Annals of
Statistics, vol. 24, no. 4, pp. 16191647, 1996.
8 C. R. Loader, Local likelihood density estimation, The Annals of Statistics, vol. 24, no. 4, pp. 1602
1618, 1996.
9 B. W. Silverman, Density Estimation for Statistics and Data Analysis, Monographs on Statistics and
Applied Probability, Chapman & Hall, London, UK, 1986.
10 E. Nadaraya, On estimating regression, Theory of Probability and Its Applications, vol. 10, pp. 186190,
1964.
11 G. S. Watson, Smooth regression analysis, Sankhya. Series A, vol. 26, pp. 359372, 1964.
12 J. Hull and A. White, Incorporating volatility updating into the historical simulation for VaR, The
Journal of Risk, vol. 1, pp. 519, 1998.
13 D. Bosq, Nonparametric Statistics for Stochastic Processes, vol. 110 of Lecture Notes in Statistics, Springer,
14 G. G. Roussas, Nonparametric estimation in mixing sequences of random variables, Journal of
Statistical Planning and Inference, vol. 18, no. 2, pp. 135149, 1988.
15 G. G. Roussas, Asymptotic normality of the kernel estimate under dependence conditions:
application to hazard rate, Journal of Statistical Planning and Inference, vol. 25, no. 1, pp. 81104, 1990.
16 M. Pritsker, Nonparametric density estimation and tests of continuous time interest rate models,
Review of Financial Studies, vol. 11, no. 3, pp. 449487, 1998.
17 T. G. Conley, L. P. Hansen, and W.-F. Liu, Bootstrapping the long run, Macroeconomic Dynamics, vol.
1, no. 2, pp. 279311, 1997.
18 Y. At-Sahalia, Testing continuous-time models of the spot interest rate, Review of Financial Studies,
vol. 9, no. 2, pp. 385426, 1996.
19 R. I. Jennrich, Asymptotic properties of non-linear least squares estimators, The Annals of
Mathematical Statistics, vol. 40, pp. 633643, 1969.
doi:10.1155/2010/718905
Research Article
Zengas New Index of Economic Inequality,
Its Estimation, and an Analysis of Incomes in Italy
Francesca Greselin,1 Leo Pasquazzi,1 and Ricardas Zitikis2

1
Dipartimento di Metodi Quantitativi per le Scienze Economiche e Aziendali,
Universita di Milano, Bicocca 20126, Milan, Italy
2
ON, Canada N6A 5B7
Academic Editor: Madan L. Puri
Copyright q 2010 Francesca Greselin et al. This is an open access article distributed under the
For at least a century academics and governmental researchers have been developing measures
that would aid them in understanding income distributions, their dierences with respect to
geographic regions, and changes over time periods. It is a fascinating area due to a number of
reasons, one of them being the fact that dierent measures, or indices, are needed to reveal dierent
features of income distributions. Keeping also in mind that the notions of poor and rich are relative
to each other, Zenga 2007 proposed a new index of economic inequality. The index is remarkably
insightful and useful, but deriving statistical inferential results has been a challenge. For example,
unlike many other indices, Zengas new index does not fall into the classes of L-, U-, and V -
statistics. In this paper we derive desired statistical inferential results, explore their performance
in a simulation study, and then use the results to analyze data from the Bank of Italy Survey on
Household Income and Wealth SHIW.
1. Introduction
Measuring and analyzing incomes, losses, risks, and other random outcomes, which we
denote by X, has been an active and fruitful research area, particularly in the fields of
econometrics and actuarial science. The Gini index is arguably the most popular measure
of inequality, with a number of extensions and generalizations available in the literature.
Keeping in mind that the notions of poor and rich are relative to each other, Zenga 1
constructed an index that reflects this relativity. We will next recall the definitions of the Gini
and Zenga indices.
Let Fx PX x denote the cumulative distribution function cdf of the random
variable X, which we assume to be nonnegative throughout the paper. Let F 1 p inf{x :
Fx p} denote the corresponding quantile function. The Lorenz curve LF p is given by
the formula see 2
p
1
LF p F 1 sds, 1.1
F 0
where F EX is the unknown true mean of X. Certainly, from the rigorous mathematical
point of view we should call LF p the Lorenz function, but this would deviate from the
widely accepted usage of the term Lorenz curve. Hence, curves and functions are viewed
as synonyms throughout this paper.
The classical Gini index GF can now be written as follows:
1
LF p
GF 1 p dp, 1.2
0 p
where p 2p. Note that p is a density function on 0, 1. Given the usual econometric
interpretation of the Lorenz curve 3, the function

LF p
GF p 1 , 1.3
p
which we call the Gini curve, is a relative measure of inequality see 4. Indeed, LF p/p is
the ratio between i the mean income of the poorest p 100% of the population and ii the
mean income of the entire population: the closer to each other these two means are, the lower
is the inequality.
Zengas 1 index ZF of inequality is defined by the formula
1

ZF ZF p dp, 1.4
0
where the Zenga curve ZF p is given by

LF p 1p
ZF p 1 . 1.5
p 1 LF p
The Zenga curve measures the inequality between i the poorest p 100% of the population
and ii the richer remaining 1 p 100% part of the population by comparing the mean
incomes of these two disjoint and exhaustive subpopulations. We will elaborate on this
interpretation later, in Section 5.
The Gini and Zenga indices GF and ZF are weighted averages of the Gini and
Zenga curves GF p and ZF p, respectively. However, while in the case of the Gini index the
weight function i.e., the density p 2p is employed, in the case of the Zenga index the
uniform weight function p 1 is used. As a consequence, the Gini index underestimates
comparisons between the very poor and the whole population, and emphasizes comparisons
which involve almost identical population subgroups. From this point of view, the Zenga
index is more impartial: it is based on all comparisons between complementary disjoint
population subgroups and gives the same weight to each comparison. Hence, the Zenga
index ZF detects, with the same sensibility, all deviations from equality in any part of the
distribution.
To illustrate the Gini curve GF p and its weighted version gF p GF pp,
and to also facilitate their comparisons with the Zenga curve ZF p, we choose the Pareto
distribution
x
0
Fx 1 , x x0 , 1.6
x
where x0 > 0 and > 0 are parameters. Later in this paper, we will use this distribution in a
simulation study, setting x0 1 and 2.06. Note that when > 2, then the second moment
of the distribution is finite. The heavy-tailed case 1 < < 2 is also of interest, especially
when modeling incomes of countries with very high economic inequality. We will provide
additional details on the case in Section 5.
Note 1. Pareto distribution 1.6 is perhaps the oldest model for income distributions. It dates
back to Pareto 5, and Pareto 6. Paretos original empirical research suggested him that the
number of tax payers with income x is roughly proportional to x1 , where is a parameter
that measures inequality. For historical details on the interpretation of this parameter in the
context of measuring economic inequality, we refer to Zenga 7. We can view the parameter
x0 > 0 as the lowest taxable income. In addition, besides being the greatest lower bound of
the distribution support, x0 is also the scale parameter of the distribution and thus does not
aect our inequality indices and curves, as we will see in formulas below.
Note 2. The Pareto distribution is positively supported, x x0 > 0. In real surveys, however,
in addition to many positive incomes we may also observe some zero and negative incomes.
This happens when evaluating net household incomes, which are the sums of payroll incomes
net wages, salaries, fringe benefits, pensions and net transfers pensions, arrears, financial
assistance, scholarships, alimony, gifts. Paid alimony and gifts are subtracted in forming
the incomes. However, negative incomes usually happen in the case of very few statistical
units. For example, in the 2006 Bank of Italy survey we observe only four households with
nonpositive incomes, out of the total of 7,766 households. Hence, it is natural to fit the
Pareto model to the positive incomes and keep in mind that we are actually dealing with
a conditional distribution. If, however, it is desired to deal with negative, null, and positive
incomes, then instead of the Pareto distribution we may switch to dierent ones, such as
Dagum distributions with three or four parameters 810.
Corresponding to Pareto distribution 1.6, the Lorenz curve is given by the formula
LF p 1 1 p11/ see 11, and thus the Gini curve becomes GF p 1 p11/ 1
p/p. In Figure 1a we have depicted the Gini and weighted Gini curves. The corresponding
Zenga curve is equal to ZF p 11p1/ /p and is depicted in Figure 1b, alongside the
Gini curve GF p for an easy comparison. Figure 1a allows us to appreciate how the Gini
weight function p 2p disguises the high inequality between the mean income of the very
poor and that of the whole population, and overemphasizes comparisons between almost
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

a b
Figure 1: The Gini curve GF pdashed; a and b, the weighted Gini curve gF psolid; a, and the
Zenga curve ZF psolid; b in the Pareto case with x0 1 and 2.06.
identical subgroups. The outcome is that the Gini index GF underestimates inequality. In
Figure 1b we see the dierence between the Gini and Zenga inequality curves. For example,
GF p for p 0.8 yields 0.296, which tells us that the mean income of the poorest 80% of
the population is 29.6% lower than the mean income of the whole population, while the
corresponding ordinate of the Zenga curve is ZF 0.8 0.678, which tells us that the mean
income of the poorest 80% of the population is 67.8% lower than the mean income of the
remaining richer part of the population.
The rest of this paper is organized as follows. In Section 2 we define two estimators of
the Zenga index ZF and develop statistical inferential results. In Section 3 we present results
of a simulation study, which explores the empirical performance of two Zenga estimators,
Z
n , including coverage accuracy and length of several types of confidence intervals.
n and Z
In Section 4 we present an analysis of the the Bank of Italy Survey on Household Income
and Wealth SHIW data. In Section 5 we further contribute to the understanding of the
Zenga index ZF by relating it to lower and upper conditional expectations, as well as to
the conditional tail expectation CTE, which has been widely used in insurance. In Section 6
we provide a theoretical background of the aforementioned two empirical Zenga estimators.
In Section 7 we justify the definitions of several variance estimators as well as their uses in
constructing confidence intervals. In Section 8 we prove Theorem 2.1 of Section 2, which is the
main technical result of the present paper. Technical lemmas and their proofs are relegated to
Section 9.
2. Estimators and Statistical Inference

Unless explicitly stated otherwise, our statistical inferential results are derived under the
assumption that data are outcomes of independent and identically distributed i.i.d. random
variables.
Hence, let X1 , . . . , Xn be independent copies of X. We use two nonparametric
estimators for the Zenga index ZF . The first one 12 is given by the formula

1n1
i1 ik1 Xk:n
Zn 1 , 2.1
n i1 n i1 nki1 Xk:n
where X1:n Xn:n are the order statistics of X1 , . . . , Xn . With X denoting the sample
mean of X1 , . . . , Xn , the second estimator of the Zenga index ZF is given by the formula
i1

n
k1 Xk:n i 1Xi:n i

n
Z log
i2
n
ki1 Xk:n iXi:n i1
i1 2.2

n1
X k1 Xk:n i 1Xi:n Xi:n
1 n log 1 n .
i1
Xi:n ki1 Xk:n iXi:n ki1 Xk:n
The two estimators Z n and Z

n are asymptotically equivalent. However, despite the fact that
the estimator Z
n is more complex, it will nevertheless be more convenient to work with when
establishing asymptotic results later in this paper.
Unless explicitly stated otherwise, we assume throughout that the cdf Fx of X is
a continuous function. We note that continuous cdfs are natural choices when modeling
income distributions, insurance risks, and losses see, e.g., 13.
Theorem 2.1. If the moment EX 2 is finite for some > 0, then one has the asymptotic
representation
n
n Z
n ZF 1 hXi oP 1, 2.3
n i1
where oP 1 denotes a random variable that converges to 0 in probability when n , and

hXi 1{Xi x} FxwF Fxdx 2.4
0
with the weight function
t
1 1 LF p 1 1 1 1
wF t 1 dp 1 dp. 2.5
F 0 p 1 LF p
2 F t p 1 LF p

n ZF is centered normal
In view of Theorem 2.1, the asymptotic distribution of n Z
with the variance F Eh X, which is finite see Theorem 7.1 and can be written as
2 2
follows:

F2 min Fx, F y FxF y wF FxwF F y dx dy. 2.6
0 0
Alternatively,
1 2
1 1 2.7
F2 twF tdF t 1 twF tdF t du.
0 0,u u,1
The latter expression of F2 is particularly convenient when working with distributions for
which the first derivative when it exists of the quantile F 1 t is a relatively simple function,
as is the case for a large class of distributions see, e.g., 14. However, irrespectively of what
expression for the variance F2 we use, the variance is unknown since the cdf Fx is unknown,
and thus F2 needs to be estimated empirically.
2.1. One Sample Case

Replacing the population cdf everywhere on the right-hand side of 2.6 by the empirical cdf

Fn x n1 ni1 1{Xi x}, where 1 denotes the indicator function, we obtain Theorem 7.2
the following estimator of the variance F2 :
n1
n1
min{k, l}k l
S2X,n
k1 l1
n nn
2.8

k l
wX,n wX,n Xk1:n Xk:n Xl1:n Xl:n ,
n n
where
k n
k
wX,n IX,n i JX,n i 2.9
n i1 ik1
with the following expressions for the summands IX,n i and JX,n i : first,
n
k2 Xk:n n 1X1:n 1 X1:n
IX,n 1 n n log 1 n . 2.10
k1 Xk:n k2 Xk:n
X1,n k2 Xk:n
Furthermore, for every i 2, . . . , n 1,
i1
Xk:n i 1Xi:n i
IX,n i n k1 2 log
n i1
ki1 Xk:n iXi:n
n n
ki1 Xk:n n iXi:n k1 Xk:n
n n n 2.11
ki1 Xk:n iXi:n ki1 Xk:n ki Xk:n
i1
1 k1 Xk:n i 1Xi:n Xi:n
n 2 log 1 n ,
Xi:n n
Xk:n iXi:n ki1 Xk:n
ki1

n i
JX,n i n log
ki1 Xk:n iXi:n i1
n 2.12
ki1 Xk:n n iXi:n Xi:n
log 1 n .
Xi:n nki1 Xk:n iXi:n ki1 Xk:n
Finally,

1 n
JX,n n log . 2.13
Xn,n n1
With the just defined estimator S2X,n of the variance F2 , we have the asymptotic result:

n Zn ZF
d N0, 1, 2.14
SX,n
where d denotes convergence in distribution.
2.2. Two Independent Samples

We now discuss a variant of statement 2.14 in the case of two populations when samples
are independent. Namely, let the random variables X1 , . . . , Xn F and Y1 , . . . , Ym H be
independent within and between the two samples. Just like in the case of the cdf Fx, here
we also assume that the cdf Hx is continuous and EY 2 < for some > 0. Furthermore,
we assume that the sample sizes n and m are comparable, which means that there exists
0, 1 such that
m
0, 1 2.15
nm
when both n and m tend to infinity. From statement 2.3 and its counterpart for Yi H we

then have that the quantity nm/n m Z
X,n Z

Y,m ZF ZH is asymptotically normal
with mean zero and the variance F2 1 H 2
. To estimate the variances F2 and H2
, we
2 2
use SX,n and SY,n , respectively, and obtain the following result:

X,n Z
Z
Y,m ZF ZH
d N0, 1. 2.16
1/nS2X,n 1/mS2Y,m
2.3. Paired Samples

Consider now the case when the two samples X1 , . . . , Xn F and Y1 , . . . , Ym H are
paired. Thus, we have that m n, and we also have that the pairs X1 , Y1 , . . . , Xn , Yn are
independent and identically distributed. Nothing is assumed about the joint distribution of
X, Y . As before, the cdfs Fx and Hy are continuous and both have finite moments
of order 2 , for some > 0. From statement 2.3 and its analog for Y we have that

Y,n ZF ZH is asymptotically normal with mean zero and the variance 2

X,n Z
n Z F,H
EhX hY 2 . The latter variance can of course be written as F2 2EhXhY H 2
.
2 2
Having already constructed estimators SX,n and SY,n , we are only left to construct an estimator
for EhXhY . Note that when X and Y are independent, then PX x, Y y
FxHy and thus the expectation EhXhY vanishes. To this end, we write the equation

EhXhY P X x, Y y FxH y wF FxwH H y dx dy. 2.17
0 0
Replacing the cdfs Fx and Hy everywhere on the right-hand side of the above equation
by their respective empirical estimators Fn x and Hn y, we have Theorem 7.3

n1
n1
1 k
k l
SX,Y,n 1 Yi,n Yl:n
k1 l1
n i1 nn
2.18

k l
wX,n wY,n Xk1:n Xk:n Yl1:n Yl:n ,
n n
where Y1,n , . . . , Yn,n are the induced by X1 , . . . , Xn order statistics of Y1 , . . . , Yn . Note that

when Y X, then Yi,n Yi:n and so the sum ki1 1{Yi,n Yl:n } is equal to min{k, l}; hence,
estimator 2.18 coincides with estimator 2.8, as expected. Consequently, S2X,n 2SX,Y,n S2Y,n
2
is an empirical estimator of F,H , and so we have that

Y,n ZF ZH
n ZX,n Z
d N0, 1. 2.19
S2X,n 2SX,Y,n S2Y,n
We conclude this section with a note that the above established asymptotic results
2.14, 2.16, and 2.19 are what we typically need when dealing with two populations,
or two time periods, but extensions to more populations and/or time periods would be a
worthwhile contribution. For hints and references on the topic, we refer to Jones et al. 15
and Brazauskas et al. 16.
3. A Simulation Study
Here we investigate the numerical performance of the estimators Z n and Z
n by simulating
data from Pareto distribution 1.6 with x0 1 and 2.06. These choices give the value
ZF 0.6, which is approximately seen in real income distributions. As to the artificial choice
x0 1, we note that since x0 is the scale parameter in the Pareto model, the inequality indices
and curves are invariant to it. Hence, all results to be reported in this section concerning the
coverage accuracy and size of confidence intervals will not be aected by the choice x0 1.
Following Davison and Hinkley 17, Chapter 5, we compute four types of confidence
intervals: normal, percentile, BCa, and t-bootstrap. For normal and studentized bootstrap
confidence intervals we estimate the variance using empirical influence values. For the
estimator Z
n , the influence values hXi are obtained from Theorem 2.1, and those for the

estimator Zn using numerical dierentiation as in Greselin and Pasquazzi 12.
In Table 1 we report coverage percentages of 10, 000 confidence intervals, for each
of the four types: normal, percentile, BCa, and t-bootstrap. Bootstrap-based approximations
Table 1: Coverage proportions of confidence intervals from the Pareto parent distribution with x0 1 and
2.06 ZF 0.6.
n
Z
n
Z
0.9000 0.9500 0.9750 0.9900 0.9000 0.9500 0.9750 0.9900
n Normal confidence intervals
200 0.7915 0.8560 0.8954 0.9281 0.7881 0.8527 0.8926 0.9266
400 0.8059 0.8705 0.9083 0.9409 0.8047 0.8693 0.9078 0.9396
800 0.8256 0.8889 0.9245 0.9514 0.8246 0.8882 0.9237 0.9503
n Percentile confidence intervals
200 0.7763 0.8326 0.8684 0.9002 0.7629 0.8190 0.8567 0.8892
400 0.8004 0.8543 0.8919 0.9218 0.7934 0.8487 0.8864 0.9179
800 0.8210 0.8777 0.9138 0.9415 0.8168 0.8751 0.9119 0.9393
n BCa confidence intervals
200 0.8082 0.8684 0.9077 0.9383 0.8054 0.867 0.9047 0.9374
400 0.8205 0.8863 0.9226 0.9531 0.8204 0.886 0.9212 0.9523
800 0.8343 0.8987 0.9331 0.9634 0.8338 0.8983 0.9323 0.9634
n t-bootstrap confidence intervals
200 0.8475 0.9041 0.9385 0.9658 0.8485 0.9049 0.9400 0.9675
400 0.8535 0.9124 0.9462 0.9708 0.8534 0.9120 0.9463 0.9709
800 0.8580 0.9168 0.9507 0.9758 0.8572 0.9169 0.9504 0.9754
have been obtained from 9, 999 resamples of the original samples. As suggested by Efron 18,
we have approximated the acceleration constant for the BCa confidence intervals by one-sixth
times the standardized third moment of the influence values. In Table 2 we report summary
statistics concerning the size of the 10, 000 confidence intervals. As expected, the confidence
intervals based on Z n and Z
n exhibit similar characteristics. We observe from Table 1 that
all confidence intervals suer from some undercoverage. For example, with sample size 800,
about 97.5% of the studentized bootstrap confidence intervals with 0.99 nominal confidence
level contain the true value of the Zenga index. It should be noted that the higher coverage
accuracy of the studentized bootstrap confidence intervals when compared to the other
ones comes at the cost of their larger sizes, as seen in Table 2. Some of the studentized
bootstrap confidence intervals extend beyond the range 0, 1 of the Zenga index ZF , but this
can easily be fixed by taking the minimum between the currently recorded upper bounds
and 1, which is the upper bound of the Zenga index ZF for every cdf F. We note that for the
BCa confidence intervals, the number of bootstrap replications of the original sample has to
be increased beyond 9, 999 if the nominal confidence level is high. Indeed, for samples of size
800, it turns out that the upper bound of 1, 598 out of 10, 000 of the BCa confidence intervals
based on Z n and with 0.99 nominal confidence level is given by the largest order statistics of
the bootstrap distribution. For the confidence intervals based on Z
n , the corresponding figure
is 1, 641.
4. An Analysis of Italian Income Data

In this section we use the Zenga index ZF to analyze data from the Bank of Italy Survey on
Household Income and Wealth SHIW. The sample of the 2006 wave of this survey contains
7, 768 households, with 3, 957 of them being panel households. For detailed information on
Table 2: Size of the 95% asymptotic confidence intervals from the Pareto parent distribution with x0 1
and 2.06 ZF 0.6.
n
Z
n
Z
min mean max min mean max
n Normal confidence intervals
200 0.0680 0.1493 0.7263 0.0674 0.1500 0.7300
400 0.0564 0.1164 0.7446 0.0563 0.1167 0.7465
800 0.0462 0.0899 0.6528 0.0462 0.0900 0.6535
n Percentile confidence intervals
200 0.0673 0.1456 0.4751 0.0667 0.1462 0.4782
400 0.0561 0.1140 0.4712 0.0561 0.1143 0.4721
800 0.0467 0.0883 0.4110 0.0468 0.0884 0.4117
n BCa confidence intervals
200 0.0668 0.1491 0.4632 0.0661 0.1497 0.4652
400 0.0561 0.1183 0.4625 0.0558 0.1186 0.4629
800 0.0465 0.0925 0.4083 0.0467 0.0927 0.4085
n t-bootstrap confidence intervals
200 0.0677 0.2068 2.4307 0.0680 0.2099 2.5148
400 0.0572 0.1550 2.0851 0.0573 0.1559 2.1009
800 0.0473 0.1159 2.2015 0.0474 0.1162 2.2051
the survey, we refer to the Bank of Italy 19 publication. In order to treat data correctly
in the case of dierent household sizes, we work with equivalent incomes, which we have
obtained by dividing the total household income by an equivalence coecient, which is the
sum of weights assigned to each household member. Following the modified Organization
for Economic Cooperation and Development OECD equivalence scale, we give weight 1 to
the household head, 0.5 to the other adult members of the household, and 0.3 to the members
under 14 years of age. It should be noted, however, thatas is the case in many surveys
concerning income analysishouseholds are selected using complex sampling designs. In
such cases, statistical inferential results are quite complex. To alleviate the diculties, in the
present paper we follow the commonly accepted practice and treat income data as if they
were i.i.d.
In Table 3 we report the values of Z n and Z
n according to the geographic area of the
households, and we also report confidence intervals for ZF based on the two estimators. We
note that two households in the sample had negative incomes in 2006, and so we have not
included them in our computations.
Note 3. Removing the negative incomes from our current analysis is important as otherwise
we would need to develop a much more complex methodology than the one oered in this
paper. To give a flavour of technical challenges, we note that the Gini index may overestimate
the economic inequality when negative, zero, and positive incomes are considered. In this
case the Gini index needs to be renormalized as demonstrated by, for example, Chen et al.
20. Another way to deal with the issue would be to analyze the negative incomes and their
concentration separately from the zero and positive incomes and their concentration.
Consequently, the point estimates of ZF are based on 7, 766 equivalent incomes with
n 0.6470 and Z
Z
n 0.6464. As pointed out by Maasoumi 21, however, good care is
Table 3: Confidence intervals for ZF in the 2006 Italian income distribution.

n estimator
Z
n estimator
Z
95% 99% 95% 99%
Lower Upper Lower Upper Lower Upper Lower Upper

Northwest: n 1988, Zn 0.5953, Zn 0.5948

Normal 0.5775 0.6144 0.5717 0.6202 0.5771 0.6138 0.5713 0.6196
Student 0.5786 0.6168 0.5737 0.6240 0.5791 0.6172 0.5748 0.6243
Percent 0.5763 0.6132 0.5710 0.6193 0.5758 0.6124 0.5706 0.6185
BCa 0.5789 0.6160 0.5741 0.6234 0.5785 0.6156 0.5738 0.6226
Northeast: n 1723, Z n 0.6108, Z
n 0.6108
Normal 0.5849 0.6393 0.5764 0.6478 0.5849 0.6393 0.5764 0.6479
Student 0.5874 0.6526 0.5796 0.6669 0.5897 0.6538 0.5836 0.6685
Percent 0.5840 0.6379 0.5773 0.6476 0.5839 0.6379 0.5772 0.6475
BCa 0.5894 0.6478 0.5841 0.6616 0.5894 0.6479 0.5842 0.6615
Center: n 1574, Z n 0.6316, Z
n 0.6316
Normal 0.5957 0.6708 0.5839 0.6826 0.5956 0.6708 0.5838 0.6827
Student 0.5991 0.6991 0.5897 0.7284 0.6036 0.7016 0.5977 0.7311
Percent 0.5948 0.6689 0.5864 0.6818 0.5948 0.6688 0.5863 0.6818
BCa 0.6024 0.6850 0.5963 0.7021 0.6024 0.6850 0.5963 0.7020
South: n 1620, Z n 0.6557, Z
n 0.6543
Normal 0.6358 0.6770 0.6293 0.6834 0.6346 0.6756 0.6282 0.6820
Student 0.6371 0.6805 0.6313 0.6902 0.6371 0.6796 0.6320 0.6900
Percent 0.6351 0.6757 0.6286 0.6828 0.6337 0.6742 0.6276 0.6812
BCa 0.6375 0.6793 0.6325 0.6888 0.6363 0.6778 0.6315 0.6873
n 0.6109, Z
Islands: n 861, Z
n 0.6095
Normal 0.5918 0.6317 0.5856 0.6380 0.5910 0.6302 0.5848 0.6364
Student 0.5927 0.6339 0.5864 0.6405 0.5928 0.6330 0.5874 0.6401
Percent 0.5897 0.6297 0.5839 0.6360 0.5885 0.6275 0.5831 0.6340
BCa 0.5923 0.6324 0.5868 0.6414 0.5914 0.6307 0.5860 0.6394
Italy entire population: n 7766, Z n 0.6470, Z
n 0.6464
Normal 0.6346 0.6596 0.6307 0.6636 0.6341 0.6591 0.6302 0.6630
Student 0.6359 0.6629 0.6327 0.6686 0.6358 0.6627 0.6331 0.6683
Percent 0.6348 0.6597 0.6314 0.6640 0.6343 0.6592 0.6309 0.6635
BCa 0.6363 0.6619 0.6334 0.6676 0.6358 0.6613 0.6330 0.6669
needed when comparing point estimates of inequality measures. Indeed, direct comparison
of the point estimates corresponding to the five geographic areas of Italy would lead us
to the conclusion that the inequality is higher in the central and southern areas when
compared to the northern area and the islands. But as we glean from pairwise comparisons
of the confidence intervals, only the dierences between the estimates corresponding to the
northwestern and southern areas and perhaps to the islands and the southern area may be
deemed statistically significant.
Moreover, we have used the paired samples of the 2004 and 2006 incomes of the 3,957
panel households in order to check whether during this time period there was a change
in inequality among households. In Table 4 we report the values of Z
n based on the panel
households for these two years, and the 95% confidence intervals for the dierence between
the values of the Zenga index for the years 2006 and 2004. These computations have been
Table 4: 95% confidence intervals for the dierence of the Zenga indices between 2006 and 2004 in the
Italian income distribution.
Northwest 926 pairs Northeast 841 pairs Center 831 pairs

n2006
Z 0.5797 Z
n2006 0.6199 Z
n2006 0.5921

n2004
Z 0.5955 Z
n2004 0.6474 Z
n2004 0.5766
Dierence 0.0158 Dierence 0.0275 Dierence 0.0155
Lower Upper Lower Upper Lower Upper
Normal 0.0426 0.0102 0.0573 0.0003 0.0183 0.0514
Student 0.0463 0.0103 0.0591 0.0017 0.0156 0.0644
Percent 0.0421 0.0108 0.0537 0.0040 0.0183 0.0505
BCa 0.0440 0.0087 0.0551 0.0022 0.0130 0.0593
South 843 pairs Islands 512 pairs Italy 3953 pairs

n2006
Z 0.6200 Z
n2006 0.6179 Z
n2006 0.6362

2004
2004
2004
Zn 0.6325 Zn 0.6239 Zn 0.6485
Dierence 0.0125 Dierence 0.0060 Dierence 0.0123
Lower Upper Lower Upper Lower Upper
Normal 0.0372 0.0129 0.0333 0.0213 0.0259 0.0007
Student 0.0365 0.0166 0.0351 0.0222 0.0264 0.0013
Percent 0.0372 0.0131 0.0333 0.0214 0.0253 0.0016
BCa 0.0351 0.0162 0.0331 0.0216 0.0255 0.0013
based on formula 2.19. Having removed the four households with at least one negative
income in the paired sample, we were left with a total of 3, 953 observations. We see that even
though we deal with large sample sizes, the point estimates alone are not reliable. Indeed, for
Italy as the whole and for all geographic areas except the center, the point estimates suggest
that the Zenga index decreased from the year 2004 to 2006. However, the 95% confidence
intervals in Table 4 suggest that this change is not significant.
5. An Alternative Look at the Zenga Index

In various contexts we have notions of rich and poor, large and small, risky and secure.
They divide the underlying populations into two parts, which we view as subpopulations.
The quantile F 1 p, for some p 0, 1, usually serves as a boundary separating the two
subpopulations. For example, we may define rich if X > F 1 p and poor if X F 1 p.
Calculating the mean value of the former subpopulation gives rise to the upper conditional
expectation EX | X > F 1 p, which is known in the actuarial risk theory as the conditional
tail expectation CTE. Calculating the mean value of the latter subpopulation gives rise to
the lower conditional expectation EX | X F 1 p, which is known in the econometric
literature as the absolute Bonferroni curve, as a function of p.
Clearly, the ratio

E X | X F 1 p
RF p 5.1
E X | X > F 1 p
of the lower and upper conditional expectations takes on values in the interval 0, 1. When X
is equal to any constant, which can be interpreted as the egalitarian case, then RF p is equal
to 1. The ratio RF p is equal to 0 for all p 0, 1 when the lower conditional expectation
is equal to 0 for all p 0, 1. This means extreme inequality in the sense that, loosely
speaking, there is only one individual who possesses the entire wealth. Our wish to associate
the egalitarian case with 0 and the extreme inequality with 1 leads to function 1RF p, which
coincides with the Zenga curve 1.5 when the cdf Fx is continuous. The area
1
E X | X F 1 p
1 dp
1 p
5.2
0 E X |X >F
beneath the function 1 RF p is always in the interval 0, 1. Quantity 5.2 is a measure
of inequality and coincides with the earlier defined Zenga index ZF when the cdf Fx is
continuous, which we assume throughout the paper.
Note that under the continuity of Fx, the lower and upper conditional expectations
are equal to the absolute Bonferroni curve p1 ALF p and the dual absolute Bonferroni curve
1 p1 F ALF p, respectively, where
p

ALF p F 1 tdt 5.3
0
is the absolute Lorenz curve. This leads us to the expression of the Zenga index ZF given by
1.4, which we now rewrite in terms of the absolute Lorenz curve as follows:
1
1 ALF p
ZF 1 1 dp. 5.4
0 p F ALF p
We will extensively use expression 5.4 in the proofs below. In particular, we will see in the
next section that the empirical Zenga index Z
n is equal to ZF with the population cdf Fx
replaced by the empirical cdf Fn x.
We are now in the position to provide additional details on the earlier noted Pareto
case 1 < < 2, when the Pareto distribution has finite EX but infinite EX 2 . The above
derived asymptotic results and thus the statistical inferential theory fail in this case. The
required adjustments are serious and rely on the use of the extreme value theory, instead
of the classical central limit theorem CLT. Specifically, the task can be achieved by first
expressing the absolute Lorenz curve ALF p in terms of the conditional tail expectation
CTE:
1
1
CTEF p F 1 tdt 5.5
1p p
using the equation ALF p F 1 pCTEF p. Hence, 5.4 becomes
1
1 CTEF 0
ZF 1 1 p dp, 5.6
0 p CTEF p
where CTEF 0 is of course the mean F . Note that replacing the population cdf Fx by its
empirical counterpart Fn x on the right-hand side of 5.6 would not lead to an estimator
that would work when EX 2 , and thus when the Pareto parameter 1 < < 2. A solution
to this problem is provided by Necir et al. 22, who have suggested a new estimator of the
conditional tail expectation CTEF p for heavy-tailed distributions. Plugging in that estimator
instead of the CTE on the right-hand side of 5.6 produces an estimator of the Zenga index
when EX 2 . Establishing asymptotic results for the new heavy-tailed Zenga estimator
would, however, be a complex technical task, well beyond the scope of the present paper, as
can be seen from the proofs of Necir et al. 22.
6. A Closer Look at the Two Zenga Estimators

Since samples are discrete populations, 5.2 and 5.4 lead to slightly dierent empirical
estimators of ZF . If we choose 5.2 and replace all population-related quantities by their
n , as seen from the proof of the
empirical counterparts, then we will arrive at the estimator Z
following theorem.
n is an empirical estimator of ZF .
Theorem 6.1. The empirical Zenga index Z
Proof. Let U be a uniform on 0, 1 random variable independent of X. The cdf of F 1 U is
F. Hence, we have the following equations:

EX X | X F 1 U
ZF 1 EU
EX X | X > F 1 U

1 Fx EX 1{X x}
1 dFx 6.1
0, Fx EX 1{X > x}

1 Fx 0,x y dF y
1 dFx.
0, Fx x, y dF y
Replacing every F on the right-hand side of 6.1 by Fn , we obtain
n
1n1
1 Fn Xi:n k1 Xk:n 1{Xk:n Xi:n }
1 n , 6.2
n i1 Fn Xi:n k1 Xk:n 1{Xk:n > Xi:n }
which simplifies to
i
1n1
1 i/n Xk:n
1 nk1 . 6.3
n i1 i/n ki1 Xk:n
n 12.
This is the estimator Z
If, on the other hand, we choose 5.4 as the starting point for constructing an empirical
estimator for ZF , then we first replace the quantile F 1 p by its empirical counterpart

Fn1 p inf x : Fn x p
6.4
i 1 i
Xi:n when p ,
n n
in the definition of ALF p, which leads to the empirical absolute Lorenz curve ALn p, and
then we replace each ALF p on the right-hand side of 5.4 by the just constructed ALn p.
Note that F ALF 1 ALn 1 X. These considerations produce the empirical Zenga
index Z
n , as seen from the proof of the following theorem.

n is an estimator of ZF .
Theorem 6.2. The empirical Zenga index Z

n is given by the equation:
Proof. By construction, the estimator Z
1
1 ALn p

n 1
Z 1 dp. 6.5
0 p X ALn p
Hence, the proof of the lemma reduces to verifying that the right-hand sides of 2.2 and 6.5
coincide. For this, we split the integral in 6.5 into the sum of integrals over the intervals
i 1, i/n for i 1, . . . , n. For every p i 1/n, i/n, we have ALn p Ci,n pXi:n ,
where
1i1
i1
Ci,n Xk:n Xi:n . 6.6
n k1 n

n n i,n , where
Hence, 6.5 can be rewritten as Z i1
i/n
1 1 i,n p
i,n 1 dp 6.7
n i1/n p i,n p
with
Ci,n X Ci,n
i,n , i,n . 6.8
Xi:n Xi:n
Consider first the case i 1. We have C1,n 0 and thus 1,n 0, which implies

X X1:n
1,n 1 log 1 n . 6.9
X1:n k2 Xk:n
Next, we consider the case i n. We have Cn,n X Xn:n and thus n,n 1, which implies

X n
n,n 1 log . 6.10
Xn:n n1
When 2 i n 1, then the integrand in the definition of i,n does not have any singularity,

since i,n > i/n due to nki1 Xk:n > 0 almost surely. Hence, after simple integration we have
that, for i 2, . . . , n 1,

i 1Xi:n i1k1 Xk:n i
i,n n log
ki1 Xk:n iXi:n i1
6.11
X i 1Xi:n i1k1 Xk:n Xi:n
1 n log 1 n .
Xi:n ki1 Xk:n iXi:n ki1 Xk:n
n
With the above formulas for i,n we easily check that the sum i1 i,n is equal to the right-
hand side of 2.2. This completes the proof of Theorem 6.2.
7. A Closer Look at the Variances

Following the formulation of Theorem 2.1 we claimed that the asymptotic distribution of

n ZF is centered normal with the finite variance 2 Eh2 X. The following theorem
n Z F
proves this claim.
n
Theorem 7.1. When EX 2 < for some > 0, then n1/2 i1 hXi converges in distribution
to the centered normal random variable

BFxwF Fxdx, 7.1
0
where Bp is the Brownian bridge on the interval 0, 1. The variance of is finite and equal to F2 .

Proof. Note that n1/2 ni1 hXi can be written as 0 en FxwF Fxdx, where en p
nEn p p is the empirical process based on the uniform on 0, 1 random variables Ui
FXi , i 1, . . . , n. We will next show that

en FxwF Fxdxd BFxwF Fxdx. 7.2
0 0
The proof is based on the well-known fact that, for every > 0, the following weak
convergence of stochastic processes takes place:

en p B p
1/2 , 0 p 1 1/2 , 0 p 1 . 7.3
p1/2 1 p p1/2 1 p
Hence, in order to prove statement 7.2, we only need to check that the integral

Fx1/2 1 Fx1/2 wF Fxdx 7.4
0
is finite. For this, by considering, for example, the two cases p 1/2 and p > 1/2 separately,
we first easily verify the bound |wF p| c c log1/p c log1/1 p. Hence, for every
> 0, there exists a constant c < such that, for all p 0, 1,
c
wF p . 7.5
p 1p

Bound 7.5 implies that integral 7.4 is finite when 0 1 Fx1/22 dx < , which is true
since the moment EX 2 is finite for some > 0 and the parameter > 0 can be chosen

as small as desired. Hence, n1/2 ni1 hXi d with denoting the integral on the right-
hand side of statement 7.2. The random variable is normal because the Brownian bridge
Bp is a Gaussian process. Furthermore, has mean zero because Bp has mean zero for
every p 0, 1. The variance of is equal to F2 because EBpBq min{p, q} pq for all
p, q 0, 1. We are left to show that E2 < . For this, we write the bound:

E 2 E BFxB F y wF FxwF F y dx dy
0 0
2 7.6
EB Fx wF Fxdx .
2
0
Since EB2 Fx Fx1 Fx, the finiteness of the integral on the right-hand side of
bound 7.6 follows from the earlier proved statement that integral 7.4 is finite. Hence,
E2 < as claimed, which concludes the proof of Theorem 7.1.
Theorem 7.2. The empirical variance S2X,n is an estimator of F2 .
Proof. We construct an empirical estimator for F2 by replacing every F on the right-hand side
of 2.6 by the empirical Fn . Consequently, we replace the function wF t by its empirical
version
t 1
1 ALn p 1 1
wX,n t 1 2 dp 1 dp. 7.7
p p X ALn p
0 X ALn p t
We denote the resulting estimator of F2 by S2X,n . The rest of the proof consists of verifying
that this estimator coincides with the one defined by 2.8. Note that min{Fn x, Fn y}
Fn xFn y 0 when x 0, X1:n Xn:n , and/or y 0, X1:n Xn:n , . Hence, the just
defined S2X,n is equal to
Xn:n Xn:n

min Fn x, Fn y Fn xFn y wX,n Fn xwX,n Fn y dx dy. 7.8
X1:n X1:n
Since Fn x k/n when x Xk:n , Xk1:n , we therefore have that
n1
n1
min{k, l} k l
S2X,n
k1 l1
n nn
7.9

k l
wX,n wX,n Xk1:n Xk:n Xl1:n Xl:n .
n n
Furthermore,
k/n 1
k 1 ALn p 1 1
wX,n 1
2 dp 1 dp
n p k/n p X ALn p
0 X ALn p
7.10

k
n
IX,n i JX,n i,
i1 ik1
where, using notations 6.6 and 6.8, the summands on the right-hand side of 7.10 are
i/n
1 1 i,n p
IX,n i 1 2 dp 7.11
Xi:n i1/n p i,n p
for all i 1, . . . , n 1, and
i/n
1 1 1
JX,n i 1 dp 7.12
Xi:n i1/n p i,n p
for all i 2, . . . , n. When i 1, then i,n 0. Hence, we immediately arrive at the expression
for IX,n 1 given by 2.10. When 2 i n 1, then

i,n i i,n i,n i,n 1
IX,n i log
Xi:n 2i,n i1 nXi:n i,n i,n i 1/ni,n i/n
7.13
1 i,n i,n i 1/n
1 log ,
Xi:n 2i,n i,n i/n
and, after some algebra, we arrive at the right-hand side of 2.11. When 2 i n 1, then
we have the expression

1 i 1 1 i,n i 1/n
JX,n i log 1 log , 7.14
Xi:n i,n i1 Xi:n i,n i,n i/n
which, after some algebra, becomes the expression recorded in 2.12. When i n, then i,n
1, and so we see that JX,n n is given by 2.13. This completes the proof of Theorem 7.2.
Theorem 7.3. The empirical mixed moment SX,Y,n is an estimator of EhXhY .
Proof. We proceed similarly to the proof of Theorem 7.2. We estimate the integrand PX
x, Y y FxHy using
1 n
1 n
1 n

1 Xi x, Yi y 1{Xi x} 1 Yi y . 7.15
n i1 n i1 n i1
After some rearrangement of terms, estimator 7.15 becomes
1 n
1 n
1 n

1 Xi:n x, Yi,n y 1{Xi:n x} 1 Yi:n y . 7.16
n i1 n i1 n i1
When x Xk:n , Xk1:n and y Yl:n , Yl1:n , then estimator 7.16 is equal to

n1 ki1 1{Yi,n Yl:n } k/nl/n, which leads us to the estimator SX,Y,n . This completes
the proof of Theorem 7.3.
8. Proof of Theorem 2.1

Throughout the proof we use the notation ALF p for the dual absolute Lorenz curve
1 1
p
F tdt, which is equal to F ALF p. Likewise, we use the notation ALn p for the
empirical dual absolute Lorenz curve.
Proof. Simple algebra gives the equations

1 1 ALn p ALF p
n Z
n ZF n 1 dp
0 p ALn p ALF p

1 1 ALn p ALF p
n 1 dp
0 p ALF p 8.1

1 1 ALF p
n 1 2 ALn p ALF p dp
0 p ALF p
OP rn,1 OP rn,2
with the remainder terms
1
1 1 1
rn,1 n 1 ALn p ALF p dp,
0 p ALn p ALF p
8.2

1 1 ALF p 1 1
rn,2 n 1 ALn p ALF p dp.
0 p ALF p ALn p ALF p
We will later show Lemmas 9.1 and 9.2 that the remainder terms rn,1 and rn,2 are of the order
oP 1. Hence, we now proceed with our analysis of the first two terms on the right-hand side
of 8.1, for which we use the general Vervaat process
p F 1 p

Vn p Fn1 t 1
F t dt Fn x Fxdx 8.3
0 0
and its dual version

1

Vn p Fn1 t 1
F t dt Fn x Fxdx. 8.4
p F 1 p
For mathematical and historical details on the Vervaat process, see Zitikis 23, Davydov and
1
Zitikis 24, Greselin et al. 25, and references therein. Since 0 Fn1 t F 1 tdt X F

and 0 Fn x Fxdx X F , adding the right-hand sides of 8.3 and 8.4 gives the
equation Vn p Vn p. Hence, whatever upper bound we have for |Vn p|, the same bound
holds for |Vn p|. In fact, the absolute value can be dropped from |Vn p| since Vn p is always
nonnegative. Furthermore, we know that Vn p does not exceed p Fn F 1 pFn1 p

F 1 p. Hence, with the notation en p nFn F 1 p p, which is the uniform on 0, 1
empirical process, we have that

n Vn p en p Fn1 p F 1 p . 8.5
Bound 8.5 implies the following asymptotic representation for the first term on the right-
hand side of 8.1:
1
1 ALn p ALF p
n 1 dp
0 p ALF p
F 1 p 8.6
1 1 1
n 1 Fn x Fxdx dp OP rn,3 ,
0 p ALF p 0
where
1
1 1
rn,3 1 en p Fn1 p F 1 p dp. 8.7
0 p ALF p
We will later show Lemma 9.3 that rn,3 oP 1. Furthermore, we have the following
asymptotic representation for the second term on the right-hand side of 8.1:
1
1 ALF p
n 1 2
ALn p ALF p dp
0 p ALF p
1 8.8
1 ALF p
n 1 Fn x Fxdx dp OP rn,4 ,
0 p AL2
F p F 1 p
where
1
ALF p 1
rn,4
1
1 en p Fn p F 1 p dp. 8.9
0 p AL2
F p
We will later show Lemma 9.4 that rn,4 oP 1. Hence, 8.1, 8.6 and 8.8 together with
the aforementioned statements that rn,1 , . . . , rn,4 are of the order oP 1 imply that
F 1 p
1 1
1
n Z
n ZF n 1 Fn x Fxdx dp
0 p ALF p 0

1 1 ALF p
n 1 Fn x Fxdx dp oP 1 8.10
0 p AL2
F p F 1 p
1 n
hXi oP 1.
n i1
This completes the proof of Theorem 2.1.
9. Negligibility of Remainder Terms

The following four lemmas establish the above noted statements that the remainder terms
rn,1 , . . . , rn,4 are of the order oP 1. In the proofs of the lemmas we will use a parameter
0, 1/2, possibly dierent from line to line but never depending on n. Furthermore, we will
frequently use the fact that
1 q
1
EX < implies
q
Fn t F 1 t dt oP 1. 9.1
0
Another technical result that we will frequently use is the fact that, for any > 0 as small as
desired,

n|Fn x Fx|
sup 1/2
OP 1 9.2
xR Fx 1 Fx1/2
when n .
Lemma 9.1. Under the conditions of Theorem 2.1, rn,1 oP 1.

1 1
Proof. We split the remainder term rn,1 n 0 dp into the sum of rn,1 n 0 dp
1
and rn,1 n 1 dp. The lemma follows if

1 for every > 0, the statement rn,1 oP 1 holds when n ,

2 rn,1 hOP 1 for a deterministic h 0 when 0, where OP 1 does not
depend on .
1
To prove part 1, we first note that when 0 < p < 1 , then ALF p 1 F 1 tdt,
1
which is positive, and ALn p 1 F 1 tdt oP 1 due to statement 9.1 with q 1. Hence,
we are left to show that, when n ,
1
1
n ALn p ALF p ALn p ALF p dp oP 1. 9.3
0 p
Since ALn p ALF p X F ALn p ALF p, statement 9.3 follows if
1 1
ALn p ALF p dp oP 1, 9.4
nX F
0 p
1
1 2
n ALn p ALF p dp oP 1. 9.5
0 p
1
We have n |X F | OP 1 and |ALn p ALF p| p 0 |Fn1 p F 1 p|2 dp1/2 . Since
1 1 1
0
|Fn p F 1 p|2 dp oP 1 and 0 p1 p dp < , we have statement 9.4. To prove
statement 9.5, we use bound 8.5 and reduce the proof to showing that
1 1 2
1 1 F p

n Fn x Fxdx dp oP 1, 9.6
n 0 p 0
1
1 1 2 2

en p Fn1 p F 1 p dp oP 1. 9.7
n 0 p
To prove statement 9.6, we use statement 9.2 and observe that
1 F 1 p 2 1
1 1 12
Fx 1/2
dx dp cF, p dp < . 9.8
0 p 0 0 p
To prove statement 9.7, we use the uniform on 0, 1 version of statement 9.2 and Holders
inequality, and in this way reduce the proof to showing that
1 1/a 1 1/b
2b
1 1 1
dp Fn p F 1 p dp oP 1 9.9
n 0 p2a 0
for some a, b > 1 such that a1 b1 1. We choose the parameters a and b as follows. First,
since EX 2 < , we set b 2 /2. Next, we choose > 0 on the left-hand side of
statement 9.9 so that 2a < 1, which holds when < /4 2 in view of the equation
a1 b1 1. Hence, statement 9.9 holds and thus statement 9.7 follows. This completes
the proof of part 1.

To establish part 2, we first estimate |rn,1 | from above using the bounds ALF p
1 pF 1 1/2 and ALn p 1 pFn1 1/2, which hold since 1/2. Hence, we have
1
reduced our task to verifying the statement n 1 |ALn p ALF p|dp hOP 1. Using
the Vervaat process Vn p and bound 8.5, we reduce the proof of the statement to showing
that the integrals
1 F 1 p

n |Fn x Fx|dx dp, 9.10
1 0
1
1
en p Fn p F 1 p dp 9.11
1
are of the order hOP 1 with possibly dierent h 0 in each case. In view of statement
9.2, we have the desired statement for integral 9.10 if the quantity
1 F 1 p
1/2
1 Fx dx dp 9.12
1 0
converges to 0 when 0, in which case we use it as h. The inner integral of 9.12 does not

exceed 0 1Fx1/2 dx, which is finite for all suciently small > 0 since EX 2 < for
some > 0. This completes the proof that quantity 9.10 is of the order hOP 1. To show
that quantity 9.11 is of a similar order, we use the uniform on 0, 1 version of statement
1
9.2 and reduce the task to showing that 1 |Fn1 p F 1 p|dp is of the order hOP 1.
By the Cauchy-Bunyakowski-Schwarz inequality, we have that
1 1 1/2
1 1

1 1
2

Fn p F p dp Fn p F p dp . 9.13
1 0
1
Since EX 2 < , we have 0 |Fn1 pF 1 p|2 dp oP 1, and so setting h establishes
the desired asymptotic result for integral 9.11. This also completes the proof of part 2, and
also of Lemma 9.1.

1
Proof. Like in the proof of Lemma 9.1, we split the remainder term rn,2 n 0 dp into the
1 1
sum of rn,2 n 0 dp and rn,2 n 1 dp. To prove the lemma, we need to
show the following.

1 For every > 0, the statement rn,2 oP 1 holds when n .

2 rn,2 hOP 1 for a deterministic h 0 when 0, where OP 1 does not
depend on .

To prove part 1, we first estimate |rn,2 | from above using the bounds p1 ALF p
1
F 1 1 < , ALF p 1 F 1 tdt > 0, and ALn p 1 F 1 tdt oP 1. This reduces
1
our task to showing that, for every > 0,

1

n ALn p AL p2 dp oP 1. 9.14
F
0

Since ALn p ALF p X F ALn p ALF p and n X F 2 oP 1, statement
9.14 follows from
1

n ALn p ALF p2 dp oP 1, 9.15
0
which is an elementary consequence of statement 9.5. This establishes part 1.

To prove part 2, we first estimate |rn,2 | from above using the bounds ALF p

1 pF 1 1/2 and ALn p 1 pFn1 1/2, and in this way reduce the task to showing
that
1
1
n ALn p ALF p dp hOP 1. 9.16
1 1p
Using the Vervaat process, statement 9.16 follows if
1
1
n |Fn x Fx|dx dp hOP 1, 9.17
1 1p F 1 p
1
1 1
en p Fn p F 1 p dp hOP 1 9.18
1 1p
with possibly dierent h 0 in each case. Using statement 9.2, we have that statement
9.17 holds with h defined as the integral
1
1 1/2
1 Fx dx dp, 9.19
1 1p F 1 p
which converges to 0 when 0 as the following argument shows. First, we write the
integrand as the product of 1 Fx and 1 Fx1/22 . Then we estimate the first factor

by 1 p . The integral 0 1 Fx1/22 dx is finite for all suciently small > 0 since
1
EX 2 < for some > 0. Since 1 1 p1 dp 0 when 0, integral 9.19 converges
to 0 when 0. The proof of statement 9.17 is finished.
We are left to prove statement 9.18. Using the uniform on 0, 1 version of statement
9.2, we reduce the task to showing that
1
1 1 1

1/2 Fn p F p dp hOP 1. 9.20
1 1p
In fact, we will see below that OP 1 can be replaced by oP 1. Using Holders inequality, we
have that the right-hand side of 9.20 does not exceed
1 1/a 1 1/b
b
1 1
1/2a dp Fn p F 1 p dp 9.21
1 1p 1
for some a, b > 1 such that a1 b1 1. We choose the parameters a and b as follows. Since
EX 2 < , we set b 2 , and so the right-most integral of 9.21 is of the order oP 1.
Furthermore, a 2 /1 < 2, which can be made arbitrarily close to 2 by choosing
suciently small > 0. Choosing > 0 so small that 1/2 a < 1, we have that the left-most
integral in 9.21 converges to 0 when 0. This establishes statement 9.18 and completes
the proof of Lemma 9.2.

1
1/2
Proof. We split the remainder term rn,3 0 dp into the sum of rn,3 0 dp and rn,3
1
dp. The lemma follows if the two summands are of the order oP 1.
1/2
1

To prove rn,3 oP 1, we use the bound ALF p 1/2 F 1 pdp and the uniform on
0, 1 version of statement 9.2, and in this way reduce our task to showing that
1/2
1 1 1

F n p F p dp oP 1. 9.22
p 1/2
0
This statement can be established following the proof of statement 9.20, with minor
modifications.
To prove rn,3
oP 1, we use the bound ALF p 1 pF 1 1/2, the fact that

supt |en t| OP 1, and statement 9.1 with q 1. The desired result for rn,3 follows, which
finishes the proof of Lemma 9.3.

1
1/2
1
Proof. We split rn,4 0 dp into the sum of rn,4 0 dp and rn,4 1/2 dp, and then
show that the two summands are of the order oP 1.

To prove rn,4 oP 1, we use the bounds p1 ALF p F 1 1/2 < and ALF p
1
F 1 pdp > 0 together with the uniform on 0, 1 version of statement 9.2. This reduces
1/2
1/2
our task to showing that 0 |Fn1 p F 1 p|dp oP 1, which holds due to statement 9.1
with q 1.

To prove rn,4 oP 1, we use the bound ALF p 1 pF 1 1/2 and the uniform on
0, 1 version of statement 9.2, and in this way reduce the proof to showing that
1
1 1 1

1/2 F n p F p dp oP 1. 9.23
1/2 1p
This statement can be established following the proof of statement 9.20. The proof of
Lemma 9.4 is finished.
Acknowledgments
The authors are indebted to two anonymous referees and the editor in charge of the
manuscript, Madan L. Puri, for their constructive criticism and suggestions that helped them
to improve the paper. The research has been partially supported by the 2009 F.A.R. Fondo
di Ateneo per la Ricerca at the University of Milan Bicocca, and the Natural Sciences and
Engineering Research Council NSERC of Canada.
References
1 M. Zenga, Inequality curve and inequality index based on the ratios between lower and upper
arithmetic means, Statistica & Applicazioni, vol. 5, pp. 327, 2007.
2 G. Pietra, Delle relazioni fra indici di variabilita, note I e II, Atti del Reale Istituto Veneto di Scienze,
Lettere ed Arti, vol. 74, pp. 775804, 1915.
3 M. O. Lorenz, Methods of measuring the concentration of wealth, Journal of the American Statistical
Association, vol. 9, pp. 209219, 1905.
4 C. Gini, Sulla misura della concentrazione e della variabilita dei caratteri, in Atti del Reale Istituto
Veneto di Scienze, Lettere ed Arti. Anno Accademico, vol. 48, part 2, pp. 12011248, Premiate Ocine
Grafiche Carlo Ferrari, Venezia, Italy, 1914.
5 V. Pareto, La legge della domanda, Giornale degli Economisti, vol. 10, pp. 5968, 1895.
6 V. Pareto, Ecrits sur la courve de la repartition de la richesse, in Complete Works of V. Pareto, G.
Busino, Ed., Librairie Droz, Geneve, Switzerland, 1965.
7 M. Zenga, Il contributo degli italiani allo studio della concentrazione, in La Distribuzione Personale
del Reddito: Problemi di Formazione, di Ripartizione e di Misurazione, M. Zenga, Ed., Vita e Pensiero,
Milano, Italy, 1987.
8 C. Dagum, A new model of personal distribution: specification and estimation, Economie Appliquee,
vol. 30, pp. 413437, 1977.
9 C. Dagum, The generation and distribution of income. The Lorenz curve and the Gini ratio,
Economie Appliquee, vol. 33, pp. 327367, 1980.
10 C. Dagum, A model of net wealth distribution specified for negative, null and positive wealth. A case
of study: Italy, in Income and Wealth Distribution, Inequality and Poverty, C. Dagum and M. Zenga, Eds.,
pp. 4256, Springer, Berlin, Germany, 1990.
11 J. L. Gastwirth, A general definition of the Lorenz curve, Econometrica, vol. 39, pp. 10371039, 1971.
12 F. Greselin and L. Pasquazzi, Asymptotic confidence intervals for a new inequality measure,
Communications in Statistics: Simulation and Computation, vol. 38, no. 8, pp. 17421756, 2009.
13 C. Kleiber and S. Kotz, Statistical Size Distributions in Economics and Actuarial Sciences, Wiley Series in
Probability and Statistics, Wiley-Interscience, Hoboken, NJ, USA, 2003.
14 Z. A. Karian and E. J. Dudewicz, Fitting Statistical Distributions: The Generalized Lambda Distribution
and Generalized Bootstrap Method, CRC Press, Boca Raton, Fla, USA, 2000.
15 B. L. Jones, M. L. Puri, and R. Zitikis, Testing hypotheses about the equality of several risk measure
values with applications in insurance, Insurance: Mathematics & Economics, vol. 38, no. 2, pp. 253270,
2006.
16 V. Brazauskas, B. L. Jones, M. L. Puri, and R. Zitikis, Nested L-statistics and their use in comparing
the riskiness of portfolios, Scandinavian Actuarial Journal, no. 3, pp. 162179, 2007.
17 A. C. Davison and D. V. Hinkley, Bootstrap Methods and Their Application, vol. 1 of Cambridge Series in
Statistical and Probabilistic Mathematics, Cambridge University Press, Cambridge, UK, 1997.
18 B. Efron, Better bootstrap confidence intervals, Journal of the American Statistical Association, vol. 82,
no. 397, pp. 171200, 1987.
19 Bank of Italy, Household income and wealth in 2004, Supplements to the Statistical Bulletin, Sample
Surveys, vol. 16, no. 7, 2006.
20 C. N. Chen, T. W. Tsaur, and T. S. Rhai, The Gini coecient and negative income, Oxford Economic
Papers, vol. 34, pp. 473478, 1982.
21 E. Maasoumi, Empirical analysis of welfare and inequality, in Handbook of Applied Econometrics,
Volume II: Microeconomics, M. H. Pesaran and P. Schmidt, Eds., Blackwell, Oxford, UK, 1994.
22 A. Necir, A. Rassoul, and R. Zitikis, Estimating the conditional tail expectation in the case of heavy-
tailed losses, Journal of Probability and Statistics. In press.
23 R. Zitikis, The Vervaat process, in Asymptotic Methods in Probability and Statistics (Ottawa, ON, 1997),
B. Szyszkowicz, Ed., pp. 667694, North-Holland, Amsterdam, The Netherlands, 1998.
24 Y. Davydov and R. Zitikis, Convex rearrangements of random elements, in Asymptotic Methods in
Stochastics, vol. 44 of Fields Institute Communications, pp. 141171, American Mathematical Society,
Providence, RI, USA, 2004.
25 F. Greselin, M. L. Puri, and R. Zitikis, L-functions, processes, and statistics in measuring economic
inequality and actuarial risks, Statistics and Its Interface, vol. 2, no. 2, pp. 227245, 2009.
doi:10.1155/2010/214358
Research Article
Risk Navigator SRM:
An Applied Risk Management Tool
Dana L. K. Hoag and Jay Parsons

Department of Agricultural and Resource Economics, Colorado State University, Campus 1172,
Fort Collins, CO 80523-1172, USA
Correspondence should be addressed to Dana L. K. Hoag, [email protected]
Received 3 November 2009; Revised 25 January 2010; Accepted 16 March 2010
Copyright q 2010 D. L. K. Hoag and J. Parsons. This is an open access article distributed under
Risk Navigator SRM is a ten-step risk management program for agricultural producers, which is
based on the strategic planning process. The ten steps are designed to integrate disparate and
dicult risk management concepts into a single system that is easy to use, yet still eective.
With the aid of computers and the internet, producers can work through each step toward a final
comprehensive plan. The website includes 25 decision tools that help producers accomplish each
step and provides links to complementary educational programs, like a national agricultural risk
education library, the award-winning risk management simulation program called Ag Survivor,
and a recently published book that describes the program and provides additional depth and
explanations. The ten-step program has been presented in over 200 workshops with over 90
percent approval by participants. The website has averaged over 1,000 unique visitors per month
from 120 countries.
1. Risk Navigator SRM:

An Applied Risk Management Tool for Agriculture
Risk management technology has drastically outpaced the ability of most practitioners to
adopt innovations. Studies of farmers and ranchers, for example, consistently show that they
do not utilize what is available; some even seem to virtually disregard risk altogether 1,
2. One survey of nearly 1,000 farmers for Farm Futures 3 found that only 5 percent use
available tools. That survey also showed that those that did manage risk tended to be high-
end managers with skills not typical of the industry. Agricultural producers were said to
resist change because there are too many decisions and too little time.
Congress considered agricultural risks so important that in 1996 it created the
Risk Management Agency RMA in the US Department of Agriculture. Its purpose
is to promote, support, and regulate sound risk management solutions for agricultural
Table 1: Risk management sources and controls; see source in 6.
Risk Defined Sources Management Controls

Uncontrollable events
such as weather, pests, or Diversification, insurance
disease make yields Weather, extreme crop, revenue, buildings,
unpredictable. Changing temperatures, pests, disease, storage, vaccines, extra labor,
Production technology makes a technology, genetics, inputs production contracts e.g.,
manager or capital availability, quality, price, ensure input supply and
obsolescent. Inputs are equipment failure, labor, . . . quality, new technologies
unavailable or low e.g., automate watering
quality
Prices of inputs or
outputs change after a Product quality genetics,
Futures and options, forward
producer commits to a disease, handling,
contracting, retained
plan of action. Price input/feed
Marketing and ownership, quality controls,
fluctuations stem from Product price quality,
Price storage timing, cooperatives,
domestic and timing, global market,
niche/value-added marketing,
international supplies or weather, government policy,
...
substantial changes in contracts, . . .
demand
Market, production, legal and
Stems form the way a Cash reserves, equity,
human risk, interest rate
business is financed borrowing capacity, reducing
changes, natural disasters
Financial Borrowed funds leverage other types of risk
drought, land market
business equity but production, marketing, etc.,
changes, foreign exchange,
increase business risks insurance
loan calls, . . .
Taxes, contract disputes,
Government or other Estate planning, tax planning,
regulations, government
institutional rules, contracts, bonds e.g.,
policies, law suits, ambiguous
Institutional regulations, and policies environmental liability,
and/or unwritten
eect profitability research and education about
agreements, neighbors,
through costs or returns local laws, . . .
environmental programs, . . .
The character, health, or
behavior of people
Ambiguous and/or Family planning, including
introduces risk. This
unwritten agreements, poor labor planning, clear contracts,
Human could include theft,
planning, miscommunication, training and goal setting,
Resources illness, death in the
health or other family communication, estate
family, loss of an
disasters, . . . planning, . . .
employee, or a divorce
for example
producers 4. According to RMA, there are at least five major forms of agricultural risks
5: production, market/price, financial, institutional, and human. As shown in Table 1,
agricultural producers face many sources of risk and a multitude of ways to manage them
across and within these five categories. Price risk, for example, can be aected by product
quality, exogenous supply, and government policies; it has at least seven management
options, including futures, forward pricing, and storage. RMA eorts have substantially
boosted the output of risk information and education to address these risks.
While there is little doubt about its importance for decision making, the challenge
is to know how to describe, measure, and communicate risk 7, page 4. Consider the
parallels to understanding and using probability, which is itself an important component
for risk management. Myerson 8 concludes that there is a disconnection between theory
and practice because formulas traditionally taught in probability courses are hard to apply to
real problems. He suggests that recent advances in computer technology can help overcome
these disconnects. Aven 7 goes further by suggesting that a common unifying framework is
also needed. For example, websites now combine a framework and computer technology to
make managing a stock portfolio relatively simple, even for those people with only a minimal
understanding of price analysis. These frameworks require integrating multiple fields, like
economics, finance, and statistics, and finding an acceptable balance between the precision
and usability.
Agricultural economics and related fields have contributed greatly to developing
innovative and eective tools for managing risk in agriculture. But, as is the case with
probability, these sometimes disparate theories and concepts can be dicult to understand,
which may make integration dicult. Numerous books, articles, and materials are available
but are generally inaccessible except by specialists 2. A 2007 study in Nevada, for example,
showed that after six months only 50% of program participants planned to incorporate
what they had learned in a risk management workshop 9. A 2007 study 10 showed that
older producers had less knowledge about risk management tools, compared to younger
producers, and had less interest in learning more about them. This supports the Farm Futures
survey 3 in that many producers view the human capital investment required to learn how
to properly incorporate the use of risk management tools into their operation as significant.
The purpose of this manuscript is to describe a new framework for risk management
called Risk Navigator SRM. The program is too involved to fully describe here, and it was
prepared for education and extension programs, rather than basic research. Nevertheless,
researchers might be interested in how disparate parts were integrated to strike a balance
between precision and usability, with the purpose of making the components of risk
management more usable through a synergistic and reinforcing framework. Precision can
interfere with usability and vice versa. That balance in Risk Navigator SRM is based on
interaction with producers at over 200 meetings in over a dozen states. This includes a
description of supporting software tools made available by website. Our focus here is on
showing how probability is integrated with other risk concepts to make risk management
more accessible to producers. All ten steps of the process were fully applied to a case study,
EWS Farms, which also may be of interest. EWS Farms produce primarily corn and wheat in
Northeastern Colorado 6.
We proceed with a description of Risk Navigator SRM and explain the SRM process,
which has ten steps. A brief summary is provided for all ten steps, but a more complete
description is provided about steps viewed to be of more interest to this readership. We also
provide examples of the computer tools available to help producers with each step where
appropriate.
2. Risk Navigator
Risk Navigator SRM is a program developed to make risk management accessible to typical
agricultural producers. The process is general enough that it can be applied to other
applications, but all of the examples and tools are customized for agriculture. SRM stands
for Strategic Risk Management. Strategic planning 11, 12 is an umbrella framework used
to organize and integrate risk management concepts and tools for farmers and ranchers. The
SRM process has been taught to hundreds of farmers and ranchers under the brand name Risk
Navigator SRM or under previous incarnations branded RightRisk. Risk Navigator SRM is
housed on a public website at http://www.RiskNavigatorSRM.com/. The website includes

detailed descriptions about a ten-step strategic risk management program and 25 customized
computer tools to help decision makers with each step. The tools are matched to the steps
and standardized through Xcelsius flash files from Excel where possible. Some files are pdf
or Excel spreadsheets. The site also includes links to complementary educational materials,
such as a risk practice simulator called Ag Survivor, http://www.AgSurvivor.com/, a
comprehensive book titled Applied Risk Management in Agriculture 6, and a general
education website called RightRisk, at http://www.RightRisk.org/. The book includes a
comprehensive discussion about risk management, matched to the ten steps, and includes a
case study to demonstrate each step. The RightRisk.org website has publications, workshops,
and access to resources such as the National Ag Risk Education Library.
3. The Strategic Risk Management Process

The framework chosen to make risk management more accessible is strategic planning 6.
The foundation of strategic planning is based on three major phases: strategic, tactical, and
operational 12, 13. The strategic phase of strategic planning is designed to set boundaries
based on resources, opportunities, and threats, and to set goals. The tactical phase is designed
to evaluate various alternatives for reaching the strategic vision and goals and to develop
a solid plan to achieve them. Implementation, evaluation, and replanning occur in the
operational stage.
The strategic, tactical, and operational phases have also been linked to agriculture in
previous studies. For example, Aull-Hyde and Tadesse 14 modeled strategic dierences
when long- and short-term risks are considered in agricultural production. Fountas et al. 15
modeled the eects of information flows for strategic, tactical, and operational dierences
in the context of precision agriculture. Other models review specific details of tactical or
operational implementation, while assuming that the strategic phase of the model has already
been conducted Ahumada and Villalobos 16.
Not surprisingly, risk management researchers have proposed frameworks that
intuitively capture the strategic planning process. For example, Clemens and Reilly 17
propose the following steps that are typical of many risk researchers e.g., 2, 18, 19:
1 identify decision,
2 identify alternatives,
3 decompose the problem,
4 choose best alternative,
5 conduct sensitivity analysis,
6 repeat if necessary.
These are the same principles found in the tactical and operational phases of strategic
planning. The strategic phase is more commonly addressed separately through risk
preference and tolerance e.g., 17, 19.
Risk Navigator SRM provides formality to agricultural risk management and
condenses this information in a way that is understandable to agricultural operators by fitting
risk management into the strategic management framework. Hoag 20 developed ten steps
that map existing risk management concepts into the three phases Figure 1. There are three
steps in the first phase: 1 determine financial health, 2 determine risk preference, and 3
Determine risk preference

Determine
financial health
Strategic Establish
risk goals
Replan Determine risk

sources
Strategic risk
management
Identify
management
Monitor & adjust
alternatives
Operational Tactical
Estimate
likelihoods
Implement Rank management alternatives

plans
Figure 1: Strategic risk management process; see source in the study of Hoag in 6.
establish risk goals. These steps are not typically covered in risk management frameworks,
but were inspired by the strategic planning process. The first two of these steps were chosen
to set boundaries on tactical choices, which are then used to set goals in Step 3, as set forth by
the strategic planning process.
There are four steps in the next phase, the tactical stage, which constitute Steps 4
7 in the SRM process: 4 determine risk sources, 5 identify management alternatives,
6 estimate risk probabilities, and 7 rank management alternatives. This tactical phase
is based on another framework, the payo matrix Table 2, which is commonly used in
risk analysis to capture frameworks like that shown above 17, 18. Each of these steps
is designed to elicit a component of a payo matrix. The payo matrix is a construct
that displays payos, usually profits, by management actions e.g., cash sale, contract sale,
hedging on the futures market, etc. and states of nature e.g., normal weather resulting
in a normal crop and typical crop prices, or bad weather, resulting in a short US crop
and high crop prices. Probability is displayed next to each state of nature. The matrix is
designed to show risk dimensions in a way that helps decision makers rank risks based
on their risk personality, which is further described in Step 7. Summary statistics can be
displayed at the bottom of the table to provide more information, such as expected value
and standard deviation. For the purposes of illustration, the EWS Farms case study manages
corn price risk. There are three marketing management alternatives: selling on the cash
market, forward pricing, or hedging. The source of risk is the likelihood of a short U.S.
crop.
One limitation of the payo matrix is that it only addresses one risk at a time.
It can accommodate complex problems with multiple management alternatives, but only
one source. Our attempts to discuss joint distributions in risk training workshops, such as
price with yield, reduced comprehension and acceptance by producers when presented;
therefore we chose to focus on addressing multiple management options for a single risk.
Joint distributions and extensions of the model are described by Hoag in 6. In addition, Ag
Survivor risk simulations are based on joint distributions, where appropriate.
Table 2: Payo matrix for EWS farms corn pricing decision; see source in 6.
Whole Farm Returns for Management Alternatives

no. 1 no. 2 no. 3
Outcome Forward
Risk states $/bushel corn Probability Cash market Contract Hedge
Normal crop $2.10/bu .35 $287,700 $305,900 $298,200
Short crop $2.50/bu .65 $342,500 $332,700 $339,000
Expected value $323,320 $323,320 $324,720
Standard deviation $38,749 $18,900 $28,850
The last phase, the operational stage, utilizes three steps intended to carry out the plans
made in the tactical stage: 8 implement plans, 9 monitor and adjust, and 10 replan. The
first step focuses on the day-to-day activities to assure that good planning eorts are carried
out. Monitoring can provide the information needed to determine whether plans should be
adjusted. Re-planning takes the decision maker around the circle to start over.
4. The Strategic Phase

We proceed with a brief description of each step. An application to the case study and
examples of tools available at the RiskNavigatorSRM.com website are included, where space
is appropriate for this audience.
Step 1 Determine Financial Health. The first step is to determine financial health in order to
determine a persons financial capability to take on risks. Financial health refers to assessing
the well-being of a businesss financial resources, with respect to their ability to take on
risk. Educational programs about financial management are widely available in agricultural
extension programs. There are six tools available at http://www.RiskNavigatorSRM.com/
that were based on an extension program in Montana developed by Duane Grif-
fith http://www.montana.edu/softwaredownloads/financialmgtdownloads.html. This
includes typical tools to develop not only commonly used financial statements, like a balance
sheet and cash flow statement, but also tools specifically designed to help people understand
how health aects risk resilience. Specifically, the RDfinancial tool and Sweet Sixteen Ratio
Analyzer tool identify strengths and weaknesses of a decision makers financial position.
RDfinancial and the Ratio Analyzer provide a plethora of financial information, including the
sixteen financial ratios commonly used to describe financial health. RDfinancial also contains
a credit scoring model.
Step 2 Determine Risk Preference. The second step involves assessing a persons risk
preferences, which aects a persons attitude about taking on risks. There are many
limitations regarding the elicitation of risk preferences 17, 18 but, sometimes, it is worth
living with these problems if producers need the extra information at the decision margin.
The Risk Navigator SRMs Risk Preference Calculator tool oers three dierent methods
to help people gauge their preferences. One method, shown in Figure 2, computes a relative
Pratt-Arrow risk preference score 21, 22. The coecient of absolute risk aversion, ra W,
is the negative of the second derivative of utility, U, for wealth, W, divided by the first
Estimate my risk preference

You have an opportunity to grow onions, but crops fail about half the time. Indicate in row 1 how much you would
spend to produce onions if you have a 50% chance of making no crop and a 50% chance of making $100 thousand
dollars. Do the same in rows 2, 3 and 4 where the odds are slightly dierent.
Decision to produce onions
Expected payo How much would you spend to produce onions?
No crop Crop
1 $0 $100 $50 $47

$0 $100
2 $0 $47 $24 $20
$0 $47
3 $47 $100 $74 $65
$47 $100
4 $20 $65 $43 $38
$20 $65
Your utility Your risk preference Risk preference key

1.2
Negative = risk taker
1
0.52
0.8 0 = risk neutral
Utility
0.6
1 = somewhat risk averse
0.4
0.2 Your risk tolerance 2 = rather risk averse
0
0 20 40 60 80 100 120 163.61 3 = very risk averse
Certainty equivalent (CE) 4 = almost paranoid about risk
Utility Predicted
Figure 2: Elicitation of risk preference/tolerance in Risk Preference Calculator. See source in 6.
derivative of utility for wealth:
U W
ra W . 4.1
U W
This coecient is positive for risk-averse individuals, zero for risk neutral individuals, and
negative for risk-loving individuals. People that are risk averse would pay a premium to
avoid risk. Risk-neutral individuals maximize expected values and ignore risk. Since r
changes with the size of the gamble, the concept of a coecient of relative risk aversion RAC
was created simply by multiplying the coecient of absolute risk aversion by wealth:
RAC Wra W. 4.2
The RAC equals 0 for someone that is risk neutral. It varies from about 0.5 to 4.0 for risk-
averse people, as suggested by Hardaker et al. 18:
i RACW 0.5: hardly risk averse at all,
ii RACW 1.0: somewhat risk averse normal,
iii RACW 2.0: rather risk averse,
iv RACW 3.0: very risk averse,
v RACW 4.0: almost paranoid about risk.
A common, yet limited, functional form used for utility is the negative exponential
utility function:
UW 1 exprW . 4.3
This is a convenient functional form for illustration since the coecient of absolute risk
aversion is r.
The RAC can be found by eliciting a persons utility function, which can be estimated
by the ELCE Equally Likely Certainty Equivalents method as described by Hardaker et
al. in 18 Figure 2. ELCE elicits equally likely certainty equivalents by asking a series of
questions that present 50-50 bets, which can be used to sketch a utility function like that
shown in the lower left quadrant of Figure 2. The certainty equivalent is the certain amount
that a person would be indierent to receiving compared to an expected value with risk. To
provide a more realistic scenario, onion production was chosen for the ELCE questions in
the Risk Preference Calculator tool, since the crop tends to produce boom or bust returns
so we could use the 50-50 technique. These bets are presented by slider bars as shown in
Figure 2. This allowed the program to remain simple and realistic, and fit well with a 50-50 bet
for agricultural producers. Using the Hardaker et al. scale, the EWS case study farmer turned
out to be quite risk tolerant, with a risk preference score of 0.52 somewhat risk averse.
This score is based on derivatives from the utility curve fitted in the lower left quadrant of
Figure 2.
The second risk preference assessment method not shown in the Risk Preference
Calculator provides users the option of taking a short risk quiz designed by Grable and
Lytton 23. The quiz asks 13 questions, which are tabulated to provide a score from low
risk tolerance to high risk tolerance. More than one method is oered to counteract
inherent diculties in measuring risk preferences. In this case, psychological research oers
a completely unique approach in the form of this quiz.
The third method links risk preference to risk tolerance, which is the amount $X,
where a person would be indierent between an equal chance of receiving $X and losing
$X/2 8. A risk-averse person that would pay a certainty equivalent of $7,000 for a risky,
50-50 bet of receiving either 0 or $20,000, for example, has a risk tolerance of $15,641;
risk tolerance increases to $99,833 for someone willing to pay $9,900 for that same bet.
A risk-neutral producer would of course be willing to pay $10,000. Risk tolerance uses a
dierent means than ELCE to elicit tolerance, but is closely linked to preference. Relative
risk tolerance RRT can be derived from the RAC, since absolute risk tolerance, R, is
the inverse of absolute risk preference, r, in the negative exponential utility function.
The scale of RRT spans from 0.25 1/4 for a person who is almost paranoid about
risk to 2 1/0.5 for someone who is rarely risk averse to 10 1/.1 for someone that is
almost risk neutral. This allows the Risk Preference Calculator to provide a comparable
estimate of the relative risk preference based on risk tolerance, which opens the door
to an entirely separate and more prevalent literature. For example, Howard 24 defined
R for firms that he looked at in terms of annual sales, equity, and income. He found
that R 1.24 multiplied by net income, or 6.4% of sales, or 15.7% of equity for the
businesses that he examined. The Risk Preference Calculator can therefore provide a
parallel estimate of risk preference by simply asking the producer for net income, sales, or
equity.
Utilizing three methods to elicit risk preferences makes the Risk Preference Calculator
tool more accessible to users with diering tastes. It also provides some continuity and
comparability for those users willing to apply all three methods, which helps combat the
inherent problems with estimating risk preferences 17.
Step 3 Establish Risk Goals. There are seven goal planning tools on the Risk Navigator SRM
website, including action planning, team roster, mission statement, time management, and
transition planning. These classic tools are updated to assist decision makers with developing
their goals in the Strategic Goal Worksheet. EWS Farms strategic goals were 6 the
following.
i Strategic Goal no. 1 Financial: Ensure short- and long-term financial success by
maintaining business profitability, while expanding the overall business financial
resource base.
ii Strategic Goal no. 2 Family: Continue to live, work, and grow our families
in a rural, agricultural environment. Encourage individual development and
exploration in a manner that is consistent and flexible in order to allow all
individuals to reach their full potential.
iii Strategic Goal no. 3 Organizational: Continue to pursue organizational structures

that fit the family dynamics of the operation, as well as allow for strategic goal
attainment. Also, increase the business activities eciency of the operation.
iv Strategic Goal no. 4 Integrated Farm Management: Manage our farm as a
cointegrated unit, while providing a step-by-step process for developing a strategic
risk management plan.
A comprehensive description of all of their goals and mission statement can be found in
Applied Risk Management in Agriculture 6.
5. The Tactical Phase

Step 4 Determine Risk Sources. The four steps in the tactical phase are those required to
build a payo matrix. The first step is to determine risk sources. In addition to determining
risk sources, it is also important to prioritize risks so that management eorts can be focused.
EWS Farms chose to focus on price risk for corn. This was interesting since it was not even
one of their risk management goals. However, it is not inconsistent with observations at risk
education extension meetings. Men that have attended the Risk Navigator SRM workshops
have overwhelmingly chosen price and yield risk revenue over all other types of risk when
doing this exercise. Decision makers are encouraged in this step to review their goals and
information with those available in the study by Hoag in 6, like those shown in Table 1.
Several methods are demonstrated in the book titled Applied Risk Management
in Agriculture 6, such as influence diagrams and SWOT Strengths, Weaknesses,
Opportunities, and Threats analysis. However, the method chosen for the Risk Navigator
tool, Risk Influence Calculator, uses a risk-influence diagram to help decision makers sort
out and prioritize risk. The Risk Influence Calculator is simple and only asks producers to list
each risk, then to rate it on a scale of 1 low to 10 high for A probability of occurring, B
impact if it occurs, and C influence to stop it from occurring.
After reviewing his goals and reading what others said about risks e.g., Table 1, EWS
Farms identified the following strategic risks:
i Market/Price
a Corn Price. Will my price cover my costs?
ii Production
a Weather. Will rainfall support my crop stand?

b Hail. Will hail destroy half my crop?
c Input seed. Will good corn seed be available at a reasonable price?
iii Financial
a Expansion. Can the operation generate enough profit to cover new land
payments?
iv Human
a Family. Will my dad retire?
v Institutional
a Water. Will irrigation water be restricted?
All risks identified are placed in a graph that plots influence against risk, so the
decision maker can prioritize risks where the biggest impact can be made. The Risk Influence
tool has been one of the most popular in Navigator risk management workshops, as it creates
a lot of discussion and rethinking about priorities.
Step 5 Identify Management Alternatives. The book, Applied Risk Management in Agriculture
6, describes four main techniques to manage risks: 1 avoid it, 2 transfer it, 3 assume or
retain it, or 4 reduce it. To keep things understandable, we use a risk profile. A risk profile
is a multidimensional representation of the consequences that stem from each management
action. For simplicity, we use a probability density function PDF as the risk profile, since
it is familiar to most people and it contains information that is relevant to managing risk,
such as mean, mode, maximum, and minimum. We can show the consequences of the four
basic management actions mentioned above through simple manipulations of the PDF such
as skewing, truncating, changing variance squishing, or changing the mean moving. This
simple representation of a risk profile and terms like squishing is meant to build basic skills
and understanding in the participants; however, perhaps more importantly, the graphical
depictions engage the audience in what may often be perceived as a dry subject. It also ties
the concept of management alternatives to PDFs for the next step concerning likelihood. Like
the previous step, the book also provides an extensive list and brief discussion about specific
techniques commonly used in agriculture, like crop insurance and the futures market e.g.,
Table 1.
Decision makers at EWS Farms identified the following possibilities to manage price
risk in their corn:
i Cash market sales selling everything in the market at harvest,

ii Forward contracting to a local grain elevator contracting for a fixed price before
harvest,
iii Hedging on the futures market selling a contract on the futures market for a fixed
price,
iv Spreading out crop sales across the year multiple marketing points,
v Maintaining flexibility on timing of sales storing, then selling opportunistically.
After more consideration, EWS Farms chose three management alternatives: cash
market, forward contracting, and hedging. They developed a comprehensive marketing
plan using the Marketing Plan tool in Risk Navigator SRM. Six components exist in
this marketing plan, and each component has its own worksheet: 1 The Relationship
between the Strategic Risk Management Plan and the Marketing Plan; 2 Production History
and Expectations; 3 Expected Prices; 4 Production Costs; 5 Price, Date, and Quantity
Targets; 6 Review and Evaluation. Each of their marketing alternatives is carefully planned,
including the distribution of the sale over time. The marketing alternatives are also based on
ten years of local data on EWS Farms and for their local elevator including basis adjustments
for prices.
To help EWS Farms prepare for the remaining steps, the book recommends putting
their information into a decision tree and then a payo matrix. The main value of starting
with the decision tree, shown in Figure 3, is in its visual construction 25, which requires the
decision maker to identify all relevant courses of action, events, and payos in a clear and
understandable manner. It also makes it easier to process information to put into a payo
matrix, as shown in Table 2.
Step 6 Estimate Probabilities. The concept of probability has been cultivated throughout
the ten-step process by encouraging decision makers to represent risk with a risk profile
and to think of a risk profile in terms that can be easily understood from the basic shape of
a PDF. They are encouraged in Step 2 to determine their risk personality preference so
they can find the risk profile that best suits them. In Step 5 they are shown how a risk profile
is aected by risk management alternatives, where the random variable is usually income or
cost. In this section we show how to tie the concept of the risk profile to a probability density
function PDF or a cumulative density function CDF more directly. Much of the book
chapter 6 is dedicated toward a basic lesson about the PDF and CDF, including concepts
like mean, mode, median, variance, standard deviation, and coecient of variation. Also
included are descriptions about how to interpret basic shapes of the PDF. The discussion is
very basic, aimed at making the PDF and CDF concepts that people could use to quickly and
intuitively interpret the basic statistical components that are important for risk management
e.g., measures of central tendency, spread, and range. For example, it is easy to convey that
management tools like insurance or the futures market squish, move, or truncate, a
PDF.
The Risk Profiler tool makes it relatively easy for producers to build probability
density functions PDFs, which then can be used to provide information for probability in
the payo matrix. The art of eliciting probability is fraught with limitations e.g., 1719,
Short US crop
$342, 500
P = 0.65 Harvest price = $2.50
Cash market
Harvest = 137, 000 bu.
P = 0.35 Normal US crop
$287, 700
Harvest price = $2.10
Short US crop
$332, 700
P = 0.65 Harvest price = $2.50
Marketing Forward contract

decision 70, 000 bu. @ $2.36
P = 0.35 Normal US crop

$305, 900
Short US crop
$339, 000
P = 0.65 (w/$0.25 basis)
Hedge
70,000 bu. @ $2.70
($0.34 basis) Normal US crop
P = 0.35
$298, 200
(w/$0.45 basis)
Figure 3: Decision tree for EWS Farms corn pricing decision; see source in 6.
but the PDFs are used here to simplify the process of estimating probability in the payo
matrix and to provide a stable mechanism to tie the steps together. Furthermore, Hoag 6
discusses some of the common problems, like anchoring, and how to avoid them.
There are at least two major ways to elicit probability information 6,10: asking the
expert about his or her degree of belief about an outcome or present the expert with a lottery
that reveals their probability values. Risk Profiler provides three options to elicit a PDF by
asking the expert about what they believe will happen. The most straightforward method
assumes that the future will look like the past ten periods. Figure 4 shows ten annual corn
price entries for EWS Farms and resulting PDF, CDF, histogram, and summary statistics. The
PDF is assumed to be Normal for simplicity. However, for increased education, a histogram is
also drawn to help decision makers understand what might be hidden behind the normalized
function. For example, a bimodal distribution is hidden by the normal distribution in the case
of corn price on EWS farms. A few summary statistics are also provided for each estimation
method.
The second method oered by Risk Profiler to elicit a PDF involves having the user
describe profile features. In this case, a PDF can be drawn based only on the minimum,
most likely, and maximum values elicited from the decision maker for the random variable.
The PERT distribution is applied as follows:
PERTa, b, c Beta1 , 2 c a a, 5.1

Risk profiler
Describe Profile Features Describe Profile PDF Build Profile w/Data Help Save, Load, Delete
Density profile
Enter ten data values for your variable below.
0.035 PDF
2.55 Statistics
2.47 0.03
Probability
Measure Value 0.025
1.91
Mean 1.97 0.02
1.48 Variance 0.14 0.015
1.56 Standard 0.01
1.84 deviation 0.38
0.005
2.12 Coecient of 0
variation 0.19 0.5 1 1.5 2 2.5 3
2.33
Max 2.55 Values
1.70 Min 1.48
1.77
Histogram Cumulative probability profile

3.5
1 CDF
3
Probability 0.8
Frequency
2.5
0.6
2
0.4
1.5 0.2
1 0
0.5 1 1.5 2 2.5 3
0.5 Values
1.51.7 1.71.9 1.92.1 2.12.3 2.32.6
Intervals
Figure 4: Building a PDF with Risk Profiler; see source in 6.
where a is the minimum value, b is the most likely value, c is the maximum value, and
Beta is the beta distribution 19.
Furthermore,

a 2b a c
1 ,
b c a
5.2
1 c
2 .
a
In this case, the resulting PDF and related information are pictured on a screen exactly like
that shown in Figure 4, but the upper left quadrant is replaced with a section to collect the
PERT input data.
The last method is describe profile PDF. In this method, people provide five sets of
probabilities and values for the random variable, see Table 3.
This is a variation on the fractal method 17, 18; in eliciting the PDF for prices for a
given crop, for example, a decision maker might be asked to pick several price values and
give an associated probability for each one. Asking for the producer to supply the probability
and values simultaneously is a combination of what Frey 26 called the fixed value method
and fixed probability method. The fixed value method asks an expert the probability that
Table 3
Probability Price
.30 $2.00
.20 $2.25
.40 $2.50
.20 $2.75
.10 $3.00
the actual value is higher or lower than some arbitrary number for the random variable.
The fixed probability method has the decision maker specify the value range for a random
variable that corresponds to a specific probability. The Risk Profiler method lets the user
simultaneously mix and match the fixed value and fixed probability methods by allowing
them to enter either probabilities or values.
Step 7 Rank Management Alternatives. The final step of the tactical stage is to rank
the various alternatives considered to this point and select those with the most desirable
outcomes. Risk Navigator SRM oers three tools to help decision makers rank risks. Value
at Risk VaR is a popular method for capturing the downside risk in financial decision
making. It is an evaluation of what you stand to lose with an investment. VaR answers the
questions What is my worst-case scenario? or How much could I potentially lose in a
really bad month? 27. This strategic tool considers only the undesirable parts of dispersion,
those on the negative side of the mean, as opposed to the standard deviation, for example.
The VaR tool is simple to use and involves only one screen not pictured here. A second
tool, the Risk Eciency Tool, uses Stochastic Eciency with Respect to a Function to rank
outcomes for all levels of risk preference, except risk preferring.
The easiest and most eective tool, Risk Ranker, uses the payo matrix to link risk
personalities with the risk profiles. This tool allows a user to directly compare risk profiles,
which can provide a lot of information by itself, and oers an instant comparison of the
management alternatives under consideration with seven ranking rules that cater to dierent
risk personalities e.g., someone that is avoiding risk as compared to someone that wishes to
maximize expected value.
After filling out a payo matrix for up to five management alternatives, the program
can be used to compare risk profiles. For example, on the second tab, Compare Profiles,
shown in Figure 5, the payo matrix entered by the decision maker is reprinted in the upper
left corner. The first column paired with any management alternative e.g., cash replicates
the information entered in Risk Profiler, which provides continuity to the program and
reinforces how the risk profile integrates with the payo matrix. All five PDFs and CDFs
for EWS Farms are plotted in one graph and summary statistics are provided in tabular form.
Decision makers are provided with information about how to rank alternatives with methods
that use only distributions, such as stochastic dominance 6, and may therefore use this
method alone to rank risks.
In many, if not most, cases risks cannot be ranked by visual inspection of the PDF
or CDF. Therefore, the next tab over, Risk Ranker displays seven dierent risk ranking
measures for the payo matrix, all on one screen: Maximize EV, Maximax, Most likely,
Minimax regret, Hurwicz, Maximin, and the Laplace Insucient Reason Index Table 4.
Each of these techniques ranks risks based on dierent aspects of the payo that might, or
might not match a decision makers risk personality. For example, Maximize EV expected
By: Dana Hoag and Eihab Fathelrahman

Contact: [email protected] Risk ranker
Payo Matrix Compare Profiles Risk Indexing Risk Ranker Help Save, Load, Delete
Use the probability density function PDF and cumulative probability function CDF charts to compare profiles
0.035 Probability density function

Probability Cash Hedge 1 Hedge 2 F price 1 F price 2 Cash
0.03
Hedge 1
Probability
0.1 1.26 1.45 1.40 1.66 1.36 0.025
0.02 Hedge 2
0 1.45 1.32 1.38 1.76 1.50 0.015 F price 1
0.9 1.97 2.11 2.10 2.59 1.80 0.01 F price 2
0.005
0 1.67 1.21 1.49 1.71 1.71 0
0.5 1 1.5 2 2.5 3 3.5
0 2.8 1.67 2.01 1.86 2.00 Values
Statistics Cash Hedge 1Hedge 2F price 1F price 2 Cumulative density function

1
Expected value 1.83 1.55 1.68 1.92 1.67
0.8 Cash
Probability
Variance 0.05 0.04 0.04 0.08 0.02
Standard 0.6 Hedge 1
0.21 0.20 0.21 0.28 0.13
deviation 0.4 Hedge 2
Coecient of 8.59 7.84 7.98 6.87 12.68 F price 1
variation 0.2
F price 2
Max 2.80 2.11 2.10 2.59 2.00 0
Min 1.26 1.21 1.38 1.66 1.36 0.5 1 1.5 2 2.5 3 3.5
Values
Figure 5: Risk Rankercompare profilessee source in 6.
Table 4: Attributes of basic decision rulesa .
Decision Rule Mean Variation Low High Other

Maximized Expected ValueChoose alternative with highest x
expected value
MaximaxChoose alternative with best outcome x
Most LikelyChoose alternative that is most likely to occur x
subjective
Minimax RegretChoose alternative with least hindsight regret x x
HurwiczWeighted average of maximax and minimax x x
MaximinChoose alternative with best of the worst outcomes x
LaplaceSimple average x
a
See a more detailed description in 6. An x is placed in the column where the rule is primarily focused.
value is for risk neutral producers. Maximin chooses the alternative with the maximum,
minimum outcome and is therefore designed for very risk-averse individuals.
Table 4 shows that each of the rules focuses on dierent areas. None is comprehensive,
and many are oversimplified. For example, Maximin would choose A in the two five-year
income streams A and B shown below, since it focuses only on the minimum; however, most
people, if not all, would choose B if they could:
A 10, 10, 10, 10, 10,

5.3
B 100, 100, 100, 9, 100.
The table presented in the tool makes it handy to compare rankings quickly and easily so that
many dimensions can be considered, and is very eective at getting across the message that
people need to match their risk management personality to their risk ranking techniques.
6. The Operational Phase

Steps 810.
The operational stage involves putting plans into action. The first step, implementation,
involves acquiring the necessary resources, scheduling the tasks to be completed, and
overseeing all aspects of the plan 6. As plans are implemented, the second step in this
phase kicks in as resources need to be monitored so adjustments can be made as needed.
The last step in the phase, and step ten overall, is re-planning. If not actively brought
to the attention of those involved with the program, this step may be easily ignored by
many managers since it could highlight what was not achieved. Re-planning is also the
first step in preparation for starting the cycle over. There are five SRM tools available on
the website for the operational phase: 1 Action Planning Worksheet, 2 Critical Success
Indicator Worksheet, 3 Resource Flow Plan, 4 Risk Management Worksheet, and 5 Time
Management Worksheet. A detailed description of how EWS Farms applied this step is
provided by Hoag in 6.
7. Conclusion
Risk Navigator SRM integrates risk management techniques into one place and provides
farmers and ranchers the resources to use the products from their home oce. It is both
a learning tool as well as a means to help producers actually manage risks, financial and
other, in their operations. It is also very practical for students in economics and business. The
manageable steps allow for farmers and ranchers to learn the tools at their own speed, while
providing the opportunities to customize the data for their own farm/ranch. We find that
the dierence between this and other programs designed to manage risk is the integration
of basic risk management principles into a structured and easy to learn format. This allows
people to use concepts that have been individually available for decades, but inaccessible,
because they are most valuable in a framework, which typically requires expertise to
build. The book, Applied Risk Management in Agriculture 6, can supplement the website
by providing detailed descriptions of each step and by providing additional educational
opportunities. Navigator workshops have been presented in formats from 45 minutes to two
days. The program has grown and the tools can be continuously upgraded since they are
oered at the Navigator website. A blog has been added and a free, ten-step online education
program is nearly complete.
Acknowledgments
The authors thank to the USDA Risk Management Agency for partial funding and the
anonymous reviewers for their helpful comments.
References
1 J. Harwood, R. Heifner, K. Coble, J. Perry, and A. Somwaru, Managing risk in farming: concepts,
research and analysis, Report 774, U.S. Department of Agriculture-Economic Research Service,
Market and Trade Economics Division and Resource Economics Division, Washington, DC, USA,
1999.
2 B. Fleisher, Agricultural Risk Management, Lynne Rienner Publishers, Boulder, Colo, USA, 1990.
3 Risk Management AgencyU.S. Department of Agriculture, Building a Risk Management Plan: Risk-
Reducing Ideas that Work, U.S. Department of Agriculture, Washington, DC, USA, 1998.
4 Risk Management AgencyU.S. Department of Agriculture, A Risk Management Agency Fact Sheet
About the Risk Management Agency, Program Aid 1667-02, U.S. Department of Agriculture, Washington,
DC, USA, 2009.
5 Risk Management AgencyU.S. Department of Agriculture, Introduction to Risk Management:
Understanding Agricultural Risks: Production, Marketing, Financial, Legal and Human Resources, U.S.
Department of Agriculture, Washington, DC, USA, 1997.
6 D. L. Hoag, Applied Risk Management in Agriculture, CRC Press-Taylor Francis, Boca Rotan, Fla, USA,
2009.
7 T. Aven, Foundations of Risk Analysis: A Knowledge and Decision-Oriented, John Wiley & Sons, Chichester,
UK, 2003.
8 R. Myerson, Probability Models for Economic Decisions, Thomson, Brooks/Cole, Belmont, Calif, USA,
2005.
9 K. Curtis, Risk management programming: is it eective at creating change, in Proceedings
of the National Extension Risk Management Education Conference, Phoenix, Ariz, USA, April 2007,
http://www.agrisk.umn.edu/conference/agenda.aspx?ConfID4&OutputSum.
10 G. F. Patrick, A. J. Peiter, T. O. Knight, K. H. Coble, and A. E. Baquet, Hog producers risk
management attitudes and desire for additional risk management education, Journal of Agricultural
and Applied Economics, vol. 39, no. 3, pp. 671687, 2007.
11 G. Steiner, A Step by Step Guide to Strategic Planning: What Every Manager Must Know, The Free Press,
12 G. Toft, Synoptic one best way approaches of strategic management, in Handbook of Strategic
Management, J. Rabin, G. Miller, and W. B. Hildreth, Eds., pp. 545560, Marcel Dekker, New York,
NY, USA, 2nd edition, 2000.
13 B. Bilgen and I. Ozkarahan, Strategic tactical and operational production-distribution models: a
review, International Journal of Technology Management, vol. 28, no. 2, pp. 151171, 2004.
14 R. Aull-Hyde and S. Tadesse, A strategic agricultural production model with risk and return
considerations, Agricultural and Resource Economics Review, vol. 23, no. 1, pp. 3746, 1994.
15 S. Fountas, D. Wulfsohn, B. S. Blackmore, H. L. Jacobsen, and S. M. Pedersen, A model of decision-
making and information flows for information-intensive agriculture, Agricultural Systems, vol. 87,
no. 2, pp. 192210, 2006.
16 O. Ahumada and J. R. Villalobos, Planning the Production and Distribution of Fresh Produce,
Annals of Operations Research. In Press.
17 R. T. Clemens and T. Reilly, Making Hard Decisions with Decision Tools, Duxbury, Pacific Grove, Calif,
USA, 2001.
18 J. B. Hardaker, R. M. Huirne, J. R. Anderson, and G. Lien, Coping with Risk in Agriculture, CABI
Publishing, Wallingford, UK, 3rd edition, 2004.
19 D. Vose, Risk Analysis: A Quantitative Guide, John Wiley & Sons, New York, NY, USA, 2nd edition,
2000.
20 D. L. Hoag, The strategic risk management process, Unpublished proposal, Department of
Agricultural and Resource Economics, Colorado State University, Fort Collins, Colo, USA, 2005.
21 K. J. Arrow, Aspects of the Theory of Risk-Bearing, Academic Bookstore, Helsinki, Finland, 1965.
22 J. W. Pratt, Risk aversion in the small and the large, Econometrica, vol. 32, pp. 122136, 1964.
23 J. E. Grable and R. H. Lytton, Financial risk tolerance revisited: the development of a risk assessment
instrument, Financial Services Review, vol. 8, no. 3, pp. 163181, 1999.
24 R. A. Howard, Decision analysis: practice and promise, Management Science, vol. 34, no. 6, pp. 679
695, 1988.
25 A. N. Rae, Agricultural Management Economics: Activity Analysis and Decision Making, CABI Publishing,
Wallingford, UK, 1994.
26 C. Frey, Briefing paper part 1: introduction to uncertainty analysis, Department of Civil Engi-
neering, North Carolina State University, 1998, http://legacy.ncsu.edu/classes/ce456001/www/
Background1.html.
27 D. Harper, Introduction to value at risk VaR, 2008, http://www.investopedia.com/articles/04/
092904.asp.
doi:10.1155/2010/707146
Research Article
Estimating L-Functionals for Heavy-Tailed
Distributions and Application
Abdelhakim Necir and Djamel Meraghni

Laboratory of Applied Mathematics, Mohamed Khider University of Biskra, 07000 Biskra, Algeria
Correspondence should be addressed to Abdelhakim Necir, [email protected]
Copyright q 2010 A. Necir and D. Meraghni. This is an open access article distributed under
L-functionals summarize numerous statistical parameters and actuarial risk measures. Their
sample estimators are linear combinations of order statistics L-statistics. There exists a class
of heavy-tailed distributions for which the asymptotic normality of these estimators cannot be
obtained by classical results. In this paper we propose, by means of extreme value theory,
alternative estimators for L-functionals and establish their asymptotic normality. Our results may
be applied to estimate the trimmed L-moments and financial risk measures for heavy-tailed
distributions.
1. Introduction
1.1. L-Functionals
Let X be a real random variable rv with continuous distribution function df F. The
corresponding L-functionals are defined by
1
LJ : JsQsds, 1.1
0
where Qs : inf{x R : Fx s}, 0 < s 1, is the quantile function pertaining to df F and
J is a measurable function defined on 0, 1 see, e.g. Serfling, 1. Several authors have used
the quantity LJ to solve some statistical problems. For example, in a work by Cherno
et al. 2 the L-functionals have a connection with optimal estimators of location and scale
parameters in parametric families of distributions. Hosking 3 introduced the L-moments as
a new approach of statistical inference of location, dispersion, skewness, kurtosis, and other
aspects of shape of probability distributions or data samples having finite means. Elamir and
Seheult 4 have defined the trimmed L-moments to answer some questions related to heavy-
tailed distributions for which means do not exist, and therefore the L-moment method cannot
be applied. In the case where the trimming parameter equals one, the first four theoretical
trimmed L-moments are
1
mi : Ji sQsds, i 1, 2, 3, 4, 1.2
0
where
Ji s : s1 si s, 0 < s < 1, 1.3
with i polynomials of order i 1 see Section 4. A partial study of statistical estimation of
trimmed L-moments was given recently by Hosking 5.
Deriving asymptotics of complex statistics is a challenging problem, and this was
indeed the case for a decade since the introduction of the distortion risk measure by
Denneberg 6 and Wang 7; see also Wang 8. The breakthrough in the area was oered
by Jones and Zitikis 9, who revealed a fundamental relationship between the distortion
risk measure and the classical L-statistic, thus opening a broad gateway for developing
statistical inferential results in the area see, e.g., Jones and Zitikis 10, 11; Brazauskas et al.
12, 13 and Greselin et al. 14. These works mainly discuss CLT-type results. We have been
utilizing the aforementioned relationship between distortion risk measures and L-statistics to
develop a statistical inferential theory for distortion risk measures in the case of heavy-tailed
distributions.
Indeed L-functionals have many applications in actuarial risk measures see, e.g.,
Wang 8, 15, 16. For example, if X 0 represents an insurance loss, the distortion risk
premium is defined by

X : g1 Fxdx, 1.4
0
where g is a non decreasing concave function with g0 0 and g1 1. By a change of
variables and integration by parts, X may be rewritten into
1
X g 1 sQsds, 1.5
0
where g denotes the Lebesgue derivative of g. For heavy-tailed claim amounts, the empirical
estimation with confidence bounds for X has been discussed by Necir et al. 17 and Necir
and Meraghni 18. If X R represents financial data such as asset log-returns, the distortion
risk measures are defined by
0

HX : g1 Fx 1 dx g1 Fxdx. 1.6
0
Likewise, by integration by parts it is shown that
1
HX g 1 sQsds. 1.7
0
Wang 8 and Jones and Zitikis 9 have defined the risk two-sided deviation by
1
r X : Jr sQsds, 0 < r < 1, 1.8
0
with
r s1r 1 s1r
Jr s : , 0 < s < 1. 1.9
2 s1r 1 s1r
As we see, X, HX, and r X are L-functionals for specific weight functions. For more
details about the distortion risk measures one refers to Wang 8, 16. A discussion on their
empirical estimation is given by Jones and Zitikis 9.
1.2. Estimation of L-Functionals and Motivations

p D
In the sequel let and , respectively, stand for convergence in probability and
convergence in distribution and let N0, 2 denote the normal distribution with mean 0 and
variance 2 .
The natural estimators of quantity LJ are linear combinations of order statistics
called L-statistics. For more details on this kind of statistics one refers to Shorack and Wellner
19, page 260. Indeed, let X1 , . . . , Xn be a sample of size n 1 from an rv X with df F, then
the sample estimator of LJ is
1
n J :
L JsQn sds, 1.10
0
where Qn s : inf{x R : Fn x s}, 0 < s 1, is the empirical quantile function that

corresponds to the empirical df Fn x : n1 ni1 I{Xi x} for x R, pertaining to the
n J may be
sample X1 , . . . , Xn with I denoting the indicator function. It is clear that L
rewritten into

n
n J
L ai,n Xi,n , 1.11
i1
i/n
where ai,n : i1/n Jsds, i 1, . . . , n, and X1,n Xn,n denote the order statistics
based upon the sample X1 , . . . , Xn . The first general theorem on the asymptotic normality of
n J is established by Cherno et al. 2. Since then, a large number of authors have studied
L
the asymptotic behavior of L-statistics. A partial list consists of Bickel 20, Shorack 21, 22,
Stigler 23, 24, Ruymgaart and Van Zuijlen 25, Sen 26, Boos 27, Mason 28, and Singh
29. Indeed, we have
n L n J LJ D N 0, 2 J , as n , 1.12
provided that
1 1
2 J : mins, t stJsJtdQsQt < . 1.13
0 0
In other words, for a given function J, condition 1.13 excludes the class of distributions F
for which 2 J is infinite. For example, if we take J 1, LJ is equal to the expected value
EX and hence the natural estimator of L n J is the sample mean X n . In this case, result 1.12
corresponds to the classical central limit theorem which is valid only when the variance of F
is finite. How then can be construct confidence bounds for the mean of a df when its variance
is infinite? This situation arises when df F belongs to the domain of attraction of -stable
laws heavy-tailed with characteristic exponent 1, 2; see Section 2. This question was
answered by Peng 30, 31 who proposed an alternative asymptotically normal estimator for
the mean. Remark 3.3 below shows that this situation also arises for the sample trimmed L-
moments mi when 1/2 < < 2/3 and for the sample risk two-sided deviation r X when
1/r 1/2 < < 1/r for any 0 < r < 1. To solve this problem in a more general setting,
we propose, by means of the extreme value theory, asymptotically normal estimators of L-
functionals for heavy-tailed distributions for which 2 J .
The remainder of this paper is organized as follows. Section 2 is devoted to a brief
introduction on the domain of attraction of -stable laws. In Section 3 we define, via the
extreme value approach, a new asymptotically normal estimator of L-functionals and state
our main results. Applications to trimmed L-moments, risk measures, and related quantities
are given in Section 4. All proofs are deferred to Section 5.
2. Domain of Attraction of -Stable Laws

A df is said to belong to the domain of attraction of a stable law with stability index 0 < 2,
notation: F D, if there exist two real sequences An > 0 and Cn such that

n
D
A1
n Xi Cn S , , , as n , 2.1
i1
where S , , is a stable distribution with parameters 0 < 2, 1 1, > 0

and < < see, e.g., Samorodnitsky and Taqqu 32. This class of distributions was
introduced by Levy during his investigations of the behavior of sums of independent random
variables in the early 1920s 33. S , , is a rich class of probability distributions that
allow skewness and thickness of tails and have many important mathematical properties.
As shown in early work by Mandelbrot 1963 and Fama 34, it is a good candidate to
accommodate heavy-tailed financial series and produces measures of risk based on the tails
of distributions, such as the Value-at-Risk. They also have been proposed as models for
many types of physical and economic systems, for more details see Weron 35. This class of
distributions have nice heavy-tail properties. More precisely, if we denote by Gx : P|X|
x Fx Fx, x > 0, the df of Z : |X|, then the tail behavior of F D, for 0 < < 2,
may be described by the following
i The tail 1 G is regularly varying at infinity with index . That is
1 Gxt
lim x , for any x > 0. 2.2
t 1 Gt
ii There exists 0 p 1 such that
1 Fx Fx
lim p, lim 1 p : q. 2.3
x1 Gx x1 Gx
Let, for 0 < s < 1, Ks : inf{x > 0 : Gx s} be the quantile function pertaining to G and
Q1 s : maxQ1 s, 0 and Q2 s : maxQs, 0. Then Proposition A.3 in a work by
Csorgo et al. 36 says that the set of conditions above is equivalent to the following.
i K1 is regularly varying at 0 with index 1/. That
K1 xs
lim x1/ , for any x > 0. 2.4
s0 K1 s
ii There exists 0 p 1 such that
Q1 1 s Q2 1 s 1/
lim p1/ , lim 1p : q1/ . 2.5
s0 K1 s s0 K1 s
Our framework is a second-order condition that specifies the rate of convergence in statement
i . There exists a function A, not changing sign near zero, such that

K1 xs x 1
limAs1 x1/ x1/ , for any x > 0, 2.6
s0 K1 s
where 0 is the second-order parameter. If 0, interpret x 1/ as log x. The

second-order condition for heavy-tailed distributions has been introduced by de Haan and
Stadtmuller 37.
3. Estimating L-Functionals When F D

3.1. Extreme Quantile Estimation
The right and left extreme quantiles of small enough level t of df F, respectively, are two reals
xR and xL defined by 1 FxR t and FxL t, that is, xR Q1 t and xL Qt.
The estimation of extreme quantiles for heavy-tailed distributions has got a great deal of
interest, see for instance Weissman 38, Dekkers and de Haan 39, Matthys and Beirlant
40 and Gomes et al. 41. Next, we introduce one of the most popular quantile estimators.
Let k kn and n be sequences of integers called trimming sequences satisfying 1 < k <
n, 1 < < n, k , , k/n 0 and /n 0, as n . Weissmans estimators of
extreme quantiles xR and xL are defined, respectively, by
1/L
L t : k
xL Q Xk,n t1/L , as t 0,
n
3.1
1/R

xR QR 1 t : Xn,n t1/R , as t 0,
n
where
1
1 k
L
L k : log Xi,n log Xk,n ,
k i1
3.2
1
1

R
R : log Xni 1,n log Xn,n
i1
are two forms of Hills estimator 42 for the stability index which could also be estimated,
using the order statistics Z1,n Zn,n associated to a sample Z1 , . . . , Zn from Z, as
follows:
1
1 m

m : log Zni 1,n log Znm,n , 3.3
m i1
with log u : max0, log u and m mn being an intermediate sequence fulfilling the same
conditions as k and . Hills estimator has been thoroughly studied, improved, and even
generalized to any real-valued tail index. Its weak consistency was established by Mason 43
assuming only that the underlying distribution is regularly varying at infinity. The almost
sure convergence was proved by Deheuvels et al. 44 and more recently by Necir 45. The
asymptotic normality has been investigated, under various conditions on the distribution tail,
by numerous workers like, for instance, Csorgo and Mason 46, Beirlant and Teugels 47,
and Dekkers et al. 48.
3.2. A Discussion on the Sample Fractions k and

Extreme value-based estimators rely essentially on the numbers k and of lower- and
upper-order statistics used in estimate computation. Estimators R have, in general,
L and
substantial variances for small values of k and and considerable biases for large values of k
and . Therefore, one has to look for optimal values for k and , which balance between these
two vices.
Numerically, there exist several procedures for the thorny issue of selecting the
numbers of order statistics appropriate for obtaining good estimates of the stability index ;
2.5
1.5
0.5
0 20 40 60 80 100
Figure 1: Plots of Hill estimators, as functions of the number of extreme statistics, solid line,
R dashed
L dotted line of the characteristic exponent of a stable distribution skewed to the right, based
line, and
on 1000 observations with 50 replications. The horizontal line represents the true value of 1.2.
2.5
1.5
0.5
0 20 40 60 80 100
Figure 2: Plots of Hill estimators, as functions of the number of extreme statistics, solid line,
R dashed
L dotted line of the characteristic exponent of a stable distribution skewed to the left, based
line, and
on 1000 observations with 50 replications. The horizontal line represents the true value of 1.2.
see, for example, Dekkers and de Haan 49, Drees and Kaufmann 50, Danielsson et al. 51,
Cheng and Peng 52 and Neves and Alves 53. Graphically, the behaviors of L , R , and
as
functions of k, and m, respectively, are illustrated by Figures 1, 2, and 3 drawn by means of
the statistical software R 54. According to Figure 1, R is much more suitable than L when
estimating the stability index of a distribution which is skewed to the right > 0 whereas
Figure 2 shows that L is much more reliable than R when the distribution is skewed to the
left < 0. In the case where the distribution is symmetric 0, both estimators seem to be
equally good as seen in Figure 3. Finally, it is worth noting that, regardless of the distribution
skewness, estimator , based on the top statistics pertaining to the absolute value of X, works
well and gives good estimates for the characteristic exponent .
2.5
1.5
0.5
0 20 40 60 80 100
Figure 3: Plots of Hill estimators, as functions of the number of extreme statistics,

solid line,
R dashed
L dotted line of the characteristic exponent of a symmetric stable distribution, based on 1000
line and
observations with 50 replications. The horizontal line represents the true value of 1.2.
1000 2000 3000 4000 5000
Figure 4: Plots of the ratios of the numbers of extreme statistics, as functions of the sample size, for a
stable symmetric distribution S1.2 1, 0, 0 solid line, a stable distribution skewed to the right S1.2 1, 0.5, 0
dashed line and a stable distribution skewed to the left S1.2 1, 0.5, 0 dotted line.
It is clear that, in general, there is no reason for the trimming sequences k and to be
equal. We assume that there exists a positive real constant such that /k as n . If
the distribution is symmetric, the value of is equal to 1; otherwise, it is less or greater than
1 depending on the sign of the distribution skewness. For an illustration, see Figure 4 where
we plot the ratio for several increasing sample sizes.
3.3. Some Regularity Assumptions on J

For application needs, the following regularity assumptions on function J are required:
H1 J is dierentiable on 0, 1,
J1 s
H2 : lims0 < ,
Js
H3 both Js and J1 s are regularly varying at zero with common index R.
H4 there exists a function a not changing sign near zero such that
Jxt/Jt x x 1
limt0 x , for any x > 0, 3.4
at
where 0 is the second-order parameter.
The three remarks below give more motivations to this paper.
Remark 3.1. Assumption H3 has already been used by Mason and Shorack 55 to establish
the asymptotic normality of trimmed L-statistics. Condition H4 is just a refinement of H3
called the second order condition that is required for quantile function K in 2.6.
Remark 3.2. Assumptions H1H4 are satisfied by all weight functions Ji i2,4 with ,
1, 1 see Section 4.1 and by function Jr in 1.9 with , r 1, 1. These two examples
show that the constants and may be positive or negative depending on application needs.
Remark 3.3. L-functionals LJ exist for any 0 < < 2 and R such that 1/ < 1.
However, Lemma 5.4 below shows that for 1/ > 1/2 we have 2 J . Then, recall
1.3; whenever 1/2 < < 2/3, the trimmed L-moments exist however 2 Ji , i 1, . . . , 4.
Likewise, recall 1.9; whenever 1/r 1/2 < < 1/r, the two-sided deviation r X exists
while 2 Jr .
3.4. Defining the Estimator and Main Results

We now have all the necessary tools to introduce our estimator of LJ, given in 1.1, when
F D with 0 < < 2. Let k kn and n be sequences of integers satisfying 1 < k < n,
1 < < n, k , , k/n 0, /n 0, and the additional condition /k <
as n . First, we must note that since 1 1/ > 0 see Remark 3.3 and since both L
and R are consistent estimators of see, Mason 43, then we have for all large n

1 1
P 1 >0 P 1 > 0 1 o1. 3.5
L
R

Observe now that LJ defined in 1.1 may be split in three integrals as follows:
k/n 1/n 1
LJ JtQtdt JtQtdt JtQtdt : TL,n TM,n TR,n . 3.6
0 k/n 1/n
L t and Q
Substituting Q R 1t for Qt and Q1t in TL,n and TR,n , respectively and making
use of assumption H3 and 3.5 yield that for all large n
k/n 1/L k/n

k
L tdt k/nJk/n
JtQ Xk,n t1/L Jtdt 1 o1 Xk,n ,
0 n 0 1 1/
L
/n 1/R /n
3.7
J1 tQR 1 tdt Xn ,n t1/R J1 tdt
0 n 0
/nJ1 /n
1 o1 Xn,n .
1 1/
R
Hence we may estimate TL,n and TR,n by
k/nJk/n /nJ1 /n

TL,n : Xk,n , TR,n : Xn,n , 3.8
1 1/
L 1 1/
R
respectively. As an estimator of TM,n we take the sample one that is
1/n
n
TM,n : JtQn tdt ai,n Xi,n , 3.9
k/n ik 1
with the same constants ai,n as those in 1.11. Thus, the final form of our estimator is

n
/nJ1 /n
k, J k/nJk/n Xk,n
L ai,n Xi,n Xn,n . 3.10
1 1/
L ik 1
1 1/
R
A universal estimator of LJ may be summarized by
L k, JI 2 J L

n J L n JI 2 J < , 3.11
n J is as in 1.11. More precisely

where L

L k, JI A , L
n J L n JI A , , 3.12
where A, : {, 0, 2 R : 1/2 < 1/ < 1} and A, is its complementary in
0, 2 R.
Note that for the particular case k and J 1 the asymptotic normality of the
trimmed mean TM,n has been established in Theorem 1 of Csorgo et al. 56. The following
theorem gives the asymptotic normality of TM,n for more general trimming sequences k and
and weighting function J. For convenience, we set, for any 0 < x < 1/2 and 0 < y < 1/2,
1y 1y

2
x, y; J : mins, t stJsJtdQsQt < , 3.13
x x
and let n2 J : 2 k/n, /n; J.
Theorem 3.4. Assume that F D with 0 < < 2. For any measurable function J satisfying
assumption (H3) with index R such that 0 < 1/ < 1 and for any sequences of integers k
and such that1 < k < n, 1 < < n, k , , k/n 0, and /n 0, as n ,
there exists a probability space , A, P carrying the sequence X1 , X2 , . . . and a sequence of Brownian
bridges {Bn s, 0 s 1, n 1, 2, . . .} such that one has for all large n

1/n
n TM,n TM,n JsBn sds

k/n
op 1, 3.14
n J n J
and therefore
n TM,n TM,n D
N0, 1 as n . 3.15
n J
The asymptotic normality of our estimator is established in the following theorem.
Theorem 3.5. Assume that F D with 0 < < 2. For any measurable function J satisfying
assumptions (H1)(H4) with index R such that 1/2 < 1/ < 1, and for any sequences of
that 1 < k < n, 1 < < n, k , , k/n 0, /n 0,
integers k and such
/k < , and kak/nAk/n 0 as n , one has
n Lk, J LJ
D
N 0, 02 , as n , 3.16
n J
where
2
22 1 2 1 1
02 02 , : 1 2 2 4 1.
2 1 1 1 1
3.17
The following corollary is more practical than Theorem 3.5 as it directly provides
confidence bounds for LJ.
Corollary 3.6. Under the assumptions of Theorem 3.5 one has
n Lk, J LJ D

N 0, V 2 , as n , 3.18
/n1/2 J1 /nXn,n
where
2/
q2 2 2/1
V V
2 2
, , , , p : 1
p
2 3.19
22 1 2 1 1
4 1,
2 1 1 1 1
with p, q as in statement (ii) of Section 2 and , as in assumptions (H2)-(H3) of Section 3.
3.5. Computing Confidence Bounds for LJ

The form of the asymptotic variance V 2 in 3.20 suggests that, in order to construct
confidence intervals for LJ, an estimate of p is needed as well. Using the intermediate order
statistic Znm,n , de Haan and Pereira 57 proposed the following consistent estimator for p :
1 n
pn pn m : I{Xi > Znm,n }, 3.20
m i1
where m mn is a sequence of integers satisfying 1 < m < n, m , and m/n 0, as

n the same as that used in 3.3.
Let J be a given weight function satisfying H1H4 with fixed constants and .
Suppose that, for n large enough, we have a realization x1 , . . . , xn of a sample X1 , . . . , Xn
from rv X with df F fulfilling all assumptions of Theorem 3.5. The 1 -confidence intervals
for LJ will be obtained via the following steps.
Step 1. Select the optimal numbers k , , and m of lower- and upper-order statistics used in
3.2 and 3.3.
Step 2. Determine Xk ,n , Xn ,n , Jk /n, J1 /n, and : /k .
L :
Step 3. Compute, using 3.2, R :
L k and R . Then deduce, by 3.10, the
k , J.
estimate L
:
Step 4. Use 3.3 and 3.20 to compute m and pn : pn m . Then deduce, by 3.19,
the asymptotic standard deviation

V : , , , , pn .
V2 3.21
Finally, the lower and upper 1 -confidence bounds for LJ, respectively, will be

k , J z/2 V Xn ,n J1 /n ,

L
n
3.22

k , J z/2 V Xn ,n J1 /n ,

L
n
where z/2 is the 1 /2 quantile of the standard normal distribution N0, 1 with 0 < < 1.
4. Applications
4.1. TL-Skewness and TL-Kurtosis When F D
When the distribution mean EX exists, the skewness and kurtosis coecients are,
respectively, defined by L1 : 3 /3/2
2 and L2 : 4 /22 with k : EX EXk , k 2, 3, and 4
being the centered moments of the distribution. They play an important role in distribution
classification, fitting models, and parameter estimation, but they are sensitive to the behavior
of the distribution extreme tails and may not exist for some distributions such as the Cauchy
distribution. Alternative measures of skewness and kurtosis have been proposed; see, for
instance, Groeneveld 58 and Hosking 3. Recently, Elamir and Seheult 4 have used the
trimmed L-moments to introduce new parameters called TL-skewness and TL-kurtosis that
are more robust against extreme values. For example, when the trimming parameter equals
one, the TL-skewness and TL-kurtosis measures are, respectively, defined by
m3 m4
1 : , 2 : , 4.1
m2 m2
where mi , i 2, 3, 4, are the trimmed L-moments defined in Section 1. The corresponding

weight functions of 1.3 are defined as follows:
J2 s : 6s1 s2s 1,

20
J3 s : s1 s 5s2 5s 1 , 4.2

3
15
J4 s : s1 s 14s3 21s2 9s 1 .

2
If we suppose that F D with 1/2 < < 2/3, then, in view of the results above,
asymptotically normal estimators for 1 and 2 will be, respectively,
3
m 4
m
1 : , 2 : , 4.3
2
m 2
m
where

6k/n2
n
i 6/n2

Xk,n aj,n Xj,n Xn,n , for i 2,

2 1/L 2 1/
R

jk 1

20k/n2
n
i 20/n2
i
m Xk,n aj,n Xj,n Xn,n , for i 3, 4.4

32 1/
L 32 1/
R

jk 1

15k/n2 n
i 15/n2

X a X Xn,n , for i 4,
22 1/
L
k,n j,n j,n
22 1/
R
jk 1
i j/n
with aj,n : J sds,
j1/n i
i 2, 3, 4, and j 1, . . . , n.
Theorem 4.1. Assume that F D with 1/2 < < 2/3. For any sequences of integers k and
such that 1 < k < n, 1 < < n, k , , k/n 0, /n 0, /k < , and
kak/nAk/n 0 as n , one has, respectively, as n ,

1 1
n D
3/2
N 0, V12 ,
/n Xn,n
4.5
2 2
n D

N 0, V22 ,
/n3/2 Xn,n
where

36 9m3 2 2 225 4m4 2 2
V12 : 2 1 , V22 : 1 , 4.6
m2 10m2 4m22 5m2
with

2/
22 12 2 1 1
2
: 1 q/p 2/3
1. 4.7
22 14 2 1
4.2. Risk Two-Sided Deviation When F D

Recall that the risk two-sided deviation is defined by
1
r X : Jr sQsds, 0 < r < 1, 4.8
0
where
r s1r 1 s1r
Jr s : , 0 < s < 1. 4.9
2 s1r 1 s1r
An asymptotically normal estimator for r X, when 1/r 1/2 < < 1/r, is
r
n
r/nr
r X rk/n Xk,n

r
aj,n Xj,n Xn,n , 4.10
2r 4/
L jk 1
2r 4/
R
where
r r
r 1 i i 1 r i i 1 r
aj,n 1 1 , 4.11
2 n n n n
j 1, . . . , n.
Theorem 4.2. Assume that F D with 0 < < 2 such that 1/r 1/2 < < 1/r, for
any 0 < r < 1. Then, for any sequences of integers k and such that1 < k < n, 1 < < n,
k , , k/n 0, /n 0, /k < , and kak/nAk/n 0 as n ,
one has, as n ,
n r X r X D

N 0, Vr2 , 4.12
/nr1/2 Xn,n
where
2/
r2 q 22 r 12 2r 1 1
Vr2 : 1 2/2r 1
1.
4 p 2r 14 r 1
4.13
5. Proofs
First we begin by the following three technical lemmas.
Lemma 5.1. Let f1 and f2 be two continuous functions defined on 0, 1 and regularly varying at
zero with respective indices > 0 and < 0 such that < . Suppose that f1 is dierentiable at zero,
then
1/2
f1 sdf2 s
lim x
. 5.1
x0 f1 xf2 x
Lemma 5.2. Under the assumptions of Theorem 3.5, one has
1/n 1/
s1 s1/2 JsdQs 1 1/2 1/ q/p
lim
k/n
, 5.2
n k/n1/2 Jk/nQk/n 1/2 1
for any 0 < < 1/4.

Lemma 5.3. For any 0 < x < 1/2 and 0 < y < 1/2 one has
1y 1y 2

x, y; J xc x yc 1 y
2 2 2
c tdt
2
xcx yc 1 y ctdt ,
x x
5.3
s
where cs : 1/2
JtdQt, 0 < s < 1/2.
Lemma 5.4. Under the assumptions of Theorem 3.5, one has
k/nJ 2 k/nQ2 k/n

lim w2 ,
n n2 J
2/ 5.4
/nJ 2 1 /nQ2 1 /n 2 q
lim 22/ 1 w2 ,
n n2 J p
where

1 2 2
w :
2
2/ 22/ 1
. 5.5
2 1 2 q/p
Proof of Lemma 5.1. Let f1 denote the derivative of f1 . Applying integration by parts, we get,
for any 0 < x < 1/2,
1/2 1/2
1 1
f1 sdf2 s f1 f2 f1 xf2 x f1 sf2 sds. 5.6
x 2 2 x
Since the product f1 f2 is regularly varying at zero with index < 0, then f1 xf2 x 0
as x 0. Therefore
1/2 1/2
f1 sdf2 s f1 sf2 sds
lim x
1 lim x
. 5.7
x0 f1 xf2 x x0 f1 xf2 x
By using Karamatas representation see, e.g., Seneta 59, it is easy to show that
xf1 x 1 o1f1 x, as x 0. 5.8
Hence
1/2 1/2
f1 sdf2 s f1 sf2 sds
lim x
1 lim x
. 5.9
x0 f1 xf2 x x0 xf1 xf2 x
It is clear that 5.8 implies that f1 is regularly varying at zero with index 1; therefore f1 f2
is regularly varying with index 1 < 0. Then, Theorem 1.2.1 by de Haan 60, page 15
yields
1/2
f1 sf2 sds 1
lim x
. 5.10
x0 xf1 xf2 x
This completes the proof of Lemma 5.1.
Proof of Lemma 5.2. We have
1/n 1/2
In : s1 s1/2 JsdQs s1 s1/2 JsdQs
k/n k/n
1/2 5.11
s1 s1/2 J1 sdQ1 s : I1n I2n .
1/n
By taking, in Lemma 5.1, f1 s s1 s1/2 Js and f2 s Qs with 1/2 ,
1/, and x k/n, we get
I1n 1/
lim . 5.12
n k/n 1/2
Jk/nQk/n 1/2 1/
Likewise if we take f1 s s1 s1/2 J1 s and f2 s Q1 s with 1/2 ,
1/, and x /n, we have
I2n 1/
lim . 5.13
n /n1/2 J1 /nQ1 /n 1/2 1/
Note that statement ii of Section 2 implies that
1/
q
lim Q1 s/Qs . 5.14
s0 p
The last two relations, together with assumption H2 and the regular variation of Q1 s,
imply that
1/ 1/2 1/
I2n q/p /
lim . 5.15
n k/n1/2 Jk/nQk/n 1/2 1/
This achieves the proof of Lemma 5.2.

Proof of Lemma 5.3. We will use similar techniques to those used by Csorgo et al. 36,
Proposition A.2. For any 0 < s < 1/2, we set

c 1y for 1 y t < 1,

Wx,y t : ct for x < t < 1 y, 5.16

cx for 0 < t x.
Then 2 x, y; J may be rewritten into
1 1 2

x, y; J
2 2
Wx,y sds Wx,y sds , 5.17
0 0
and the result of Lemma 5.3 follows immediately.
Proof of Lemma 5.4. From Lemma 5.3 we may write
n2 J
Tn1 Tn2 Tn3 Tn4 , 5.18
k/nJ 2 k/nQ2 k/n
where
k/nc2 k/n /nc2 1 /n

Tn1 : , Tn2 : ,
k/nJ k/nQ k/n
2 2 k/nJ 2 k/nQ2 k/n
1/n 2
k/n
c tdt
Tn3 : , 5.19
k/nJ 2 k/nQ2 k/n
1/n
2
k/nck/n /nc1 /n k/n ctdt
Tn4 : .
k/nJ 2 k/nQ2 k/n
By the same arguments as in the proof of Lemma 5.2, we infer that
ck/n 1
lim
,
n Jk/nQk/n 1
1/ 1/ 5.20
c1 /n q/p
lim .
n Jk/nQk/n 1
Therefore
2/ 22/ 1
1 2 q/p
lim Tn1 2 , lim Tn2 2 . 5.21
n n
1 1
Next, we consider the third term Tn3 which may be rewritten into
1/2 1/n
k/n tdt
c2 c2 tdt
Tn3
1/2
. 5.22
k/nJ 2 k/nQ2 k/n k/nJ 2 k/nQ2 k/n
Observe that
1/2 2 1/2 c2 tdt
k/n tdt
c2 ck/n

k/n
. 5.23
k/nJ 2 k/nQ2 k/n Jk/nQk/n k/nc2 k/n
It is easy to verify that function c2 is regularly varying at zero with index 2 1/. Thus,
by Theorem 1.2.1 by de Haan 60 we have
1/2
c2 tdt
lim
k/n
. 5.24
n k/nc2 k/n 2 2
Hence
1/2
k/n tdt
c2
lim 2 . 5.25
n k/nJ 2 k/nQ2 k/n
1 2 2
By similar arguments we show that
1/n 2/ 2 22/ 1

1/2
c2 tdt q/p
lim 2 . 5.26
n k/nJ 2 k/nQ2 k/n
1 2 2
Therefore
2/ 2 22/ 1
1 q/p
lim Tn3 2 . 5.27
n
1 2 2
By analogous techniques we show that Tn4 0 as n ; we omit details. Summing up

the three limits of Tni , i 1, 2, 3, achieves the proof of the first part of Lemma 5.4. As for the
second assertion of the lemma, we apply a similar procedure.
5.1. Proof of Theorem 3.4

Csorgo et al. 36 have constructed a probability space , A, P carrying an infinite sequence
1 , 2 , . . . of independent rvs uniformly distributed on 0, 1 and a sequence of Brownian
bridges {Bn s, 0 s 1, n 1, 2, . . .} such that, for the empirical process,
n s : n1/2 {n s s}, 0 s 1, 5.28

where n is the uniform empirical df pertaining to the sample 1 , . . . , n ; we have for any
0 < 1/4 and for all large n

n s Bn s
sup 1/2
Op n . 5.29
1/ns11/n s1 s
For each n 1, let 1,n n,n denote the order statistics corresponding to 1 , . . . , n .
Note that for each n, the random vector Q1,n , . . . , Qn,n has the same distribution as
X1,n , . . . , Xn,n . Therefore, for 1 i n, we shall use the rvs Qi,n to represent the rvs
Xi,n , and without loss of generality, we shall be working, in all the following proofs, on the
probability space above. According to this convention, the term TM,n defined in 3.9 may be
rewritten into
n,n
TM,n Qsdn s, 5.30
k,n
s
where s : 0
Jtdt. Integrating by parts yields
n1/2 TM,n TM,n

1,n 2,n 3,n , 5.31
n J
where
1/n
n1/2 k/n {n s s}dQs
1,n : ,
n J
k,n
n1/2 k/n {n
s k/n}dQs
2,n : , 5.32
n J
1/n
n1/2 n,n {n s 1 /n}dQs
3,n : .
n J
Next, we show that
D
1,n N0, 1 as n , 5.33
p
i,n 0 as n for i 2, 3. 5.34
Making use of the mean-value theorem, we have for each n
n s s n s sJn s, 5.35

where {n s}n1 is a sequence of rvs with values in the open interval of endpoints s 0, 1
and n s. Therefore
1/n
n sJn sdQs
1,n
k/n
. 5.36
n J
This may be rewritten into
1/n 1/n
k/n
n sJsdQs k/n
n sJs{Jn s Js}dQs
1,n
n J n J 5.37
: 1,n
1,n .
Note that
1/n
n s Bn sJsdQs
k/n
n J
5.38
n s Bn s 1/n
sup 1/2
s1 s1/2 |Js|dQs/n J,
k/ns1/n s1 s k/n
for 0 < < 1/4, which by 5.29 is equal to
1/n
Op n s1 s1/2 |Js|dQs
k/n
. 5.39
n J
Since we have, from Lemmas 5.2 and 5.3,
1/n
n
s1 s1/2 |Js|dQs/n J O1, as n , 5.40
k k/n
then the right-hand side of the last inequality is equal to Op k which in turn tends to zero
as n . This implies that as n
1/n
Bn sJsdQs
1,n
k/n
op 1. 5.41
n J
Next, we show that 1,n op 1. Indeed, function J is dierentiable on 0, 1; then by
the mean-value theorem, there exists a sequence {n s}n1 of rvs with values in the open
interval of endpoints s 0, 1 and n s such that for each n we have
1/n
n sJs{n s s}J n sdQs

k/n 5.42
1,n .
n J
From inequalities 3.9 and 3.10 by Mason and Shorack 55, we infer that, for any 0 < < 1,
there exists 0 < M < such that for all large n we have
M |Js|
J n s , 5.43
s1 s
for any 0 < s 1/2. On the other hand, we have for any 0 < s < 1
|n s s| |n s s|. 5.44
Therefore
1/2
M n1/2 n s2 |Js|/s1 s dQs

k/n 5.45
1,n .
n J
This implies, since, for each n 1, E|n s|2 < s1 s, that
1/2
M n1/2
k/n |Js|dQs 5.46
E
1,n ,
n J
which tends to zero as n .

Next, we consider the term 2,n which may be rewritten into
k,n k,n
n1/2 k/n {n s
s}dQs n1/2 k/n {s
k/n}dQs
2,n . 5.47
n J n J
Making use of the mean-value theorem, we get
k,n k,n
sJ
k/n n
n s dQs n1/2 k/n
s k/nJsn dQs
2,n , 5.48
n J n J
where n s is a sequence of rvs with values in the open interval of endpoints s k/n, k,n
and n s and sn a sequence of rvs with values in the open interval of endpoints s
k/n, k,n and k/n. Again we may rewrite 2,n into
k,n k,n
k/n n s
J n s Js dQs k/n n sJsdQs

2,n
n J n J
k,n
k k Jsn
n1/2 J s 1 dQs/n J 5.49
n k/n n Jk/n
k,n
k k
n1/2 J s dQs/n J.
n k/n n
Recall that, as n , both k and tend to infinity with k/n 0 and /n 0. This implies
that

n k D
k,n N0, 1 as n , 5.50
k1/2 n

n l D
nl,n 1 N0, 1 as n , 5.51
l1/2 n
see, e.g., Balkema and de Haan 61, page 18. Next, we use similar arguments to those used
in the proof of Theorem 1 by Csorgo et al. 56. For any 0 < c < write
k/n ck1/2 /n

n sJ n s JsdQs
1 k/nck1/2 /n
2,n c :
n J
k/n ck1/2 /n
k/nck 1/2 /n
|Js|n sdQs
,
n J
5.52
k/n ck1/2 /n
2 k/nck1/2 /n
n1/2 |s k/n|Jsn dQs
2,n c :
n J
k/n ck1/2 /n
n1/2 Jk/n k/nck1/2 /n
s k/nJsn /Jk/n 1dQs
.
n J
Notice that by 5.49

k k1/2
1 2
lim lim inf P |2,n | 2,n c 2,n c lim lim inf P k,n c . 5.53
c n c n n n
In view of 5.50, this last quantity equals 1. Therefore to establish 5.34 for i 2, it suces
to show that for each 0 < c <
1 p 2
2,n c 0, 2,n c 0 as n . 5.54
By the mean-value theorem, there exists {n s}n1 a sequence of rvs with values in the open
interval of endpoints s and n s such that for each n we have

J n s Js n s s J n s . 5.55
Since |n s s| |n s s|, then by inequality 5.43 we infer that, for any 0 < < 1,
there exists 0 < M < such that for all large n we have

M n1/2 n s|Js|
J n s Js . 5.56
s1 s
1
This implies that the first term in 2,n c is less than or equal to
k/n ck1/2 /n

2n s|Js|
M n1/2 dQs/n J . 5.57
k/nck1/2 /n s1 s
Since E2n s s1 s, then the expected value of the previous quantity is less than or
equal to
1 ck/n
M n1/2 1ck/n
|Js|dQs
. 5.58
n J
1
Likewise the expected value of the second term in 2,n c is less than or equal to
k/n ck1/2 /n 1/2 k/n ck1/2 /n

k/nck1/2 /n
s1 s1/2 |Js|dQs k/n ck1/2 /n k/nck1/2 /n
|Js|dQs
.
n J n J
5.59
Fix 1 < < . It is readily verified that, for all large n, the quantity 5.58 is less than
k/n
M n1/2 |Js|dQs
k/n
, 5.60
n J
and the right-hand side of 5.59 is less than

1/2 k/n
k
2 |Js|dQs/n J. 5.61
n k/n
Therefore for all large n

k 1/2 k/n
1
E 2,n c 2 M k1/2 |Js|dQs/n J. 5.62
n k/n
By routine manipulations, as in the proofs of Lemmas 5.1 and 5.2 we omit details, we easily
show that
1/2 k/n
k w 1/
lim |Js|dQs/n J 1/ . 5.63

n n k/n 1
Since k , then for any 0 < c <
1
lim supE 2,n c 2 1/ 1/ w, 5.64
n
for any fixed 1 < < . This implies that for all 0 < c <
1
lim E 2,n c 0. 5.65
n
Therefore, by Markov inequality, we have the first result of 5.34.

2
Now, consider the term 2,n c. Observe that ns /k is a sequence of rvs with values in
p
the open interval of endpoints nk,n /k and 1. On the other hand, 5.49 implies that nk,n /k

p
1 as n . Hence ns /k 1 as well. Then, it is readily checked that, in view of relation
3.9 by Mason and Shorack 55, we have
supsHn Jsn
Op 1 as n . 5.66
Jk/n
Therefore
k/n ck1/2 /n

k k
n1/2 s dQs/n J.
2
2,n c Op 1J 5.67
n k/nck1/2 /n n
Observe that
k/n ck1/2 /n 1/2 k/n ck1/2 /n

k k
n1/2 s dQs c dQs. 5.68
k/nck1/2 /n n n k/nck1/2 /n
Hence for all large n
k/n
2 k
2,n c Op 1J dQs/n J, 5.69
n k/n
for any fixed 1 < < , we have
2
2,n c Op 1 1/ 1/ w. 5.70
2
this means that 2,n c 0 as n . By the same arguments and making use of 5.51 we
p
show that 3,n 0 as n we omit details, which achieves the proof of Theorem 3.4.

Recall 3.8, 3.9, and 3.10 and write
k, J LJ TL,n TL,n TM,n TM,n TR,n TR,n .

L 5.71
It is easy to verify that
k/n
L
k/nJk/nXk ,n
TR,n TR,n JsQtdt SLn1 SLn2 SLn3 , 5.72
1 L 1 0
where

k k L
SLn1 : J Xk ,n ,
n n 1 L 1 1 1

k/nQk/nJk/n Xk , n
L
Sn2 : 1 , 5.73
1 1 Qk/n
k/n
k/nJk/nQk/n
L
Sn3 : JsQtdt.
1 1 0
Likewise we have
/n
R
/nJ1 /nXn ,n
TR,n TR,n JsQ1 tdt SRn1 SRn2 SRn3 , 5.74
1 R 1 0
where

L

SRn1: k/nJk/nXk ,n ,
1 L 1 1 1

/nQ1 /nJ1 /n Xn ,n
R
Sn2 : 1 , 5.75
1 1 Q1 /n
/n
/nQ1 /nJ1 /n
SRn3 : JsQ1 tdt.
1 1 0
It is readily checked that SRn1 may be rewritten into

L k/nJk/nXk ,n
1 1
SLn1 . 5.76
1 L 1 1 1 L
L is a consistent estimator of , then for all large n

Since

2 k/nJk/nXk ,n 1 1
SLn1 1 op 1 2 . 5.77
1 1 L

In view of Theorems 2.3 and 2.4 of Csorgo and Mason 46, Peng 30, and Necir et al. 17
has been shown that under the second-order condition 2.6 and for all large n
1 k/n
1 n k n Bn s
k Bn ds op 1,
L
k n k 0 s
Xk ,
n k 5.78
k n
1 1 Bn op 1,
Qk/n k n

Xk ,n k
1 op 1,
Q n
where {Bn s, 0 s 1, n 1, 2, . . .} is the sequence of Brownian bridges defined in

Theorem 3.4. This implies that for all large n
k/n
k1/2 /n Jk/nQk/n n k n Bn s
SLn1 1 op 1 2 Bn ds op 1 ,
1 1 k n k 0 s

k1/2 /n Qk/nJk/n n k
SLn2 Bn op 1 .
1 1 k n
5.79
Then, in view of Lemma 5.3, we get for all large n
L k/n
n Sn1 SLn2 w n k n Bn s
2 k Bn n k ds
n J 1 1 0 s
5.80

w n k
Bn op 1.
1 1 k n
By the same arguments we omit details, we show that for all large n
R 1
n Sn1 SRn2 wR n n Bn s
2 Bn 1 ds
n J 1 1 n 1/n 1 s
5.81

wR n
Bn 1 op 1,
1 1 n
where wR : ||q/p1/ 1/ 1/2 w. Similar arguments as those used in the proof of Theorem
1 by Necir et al. 17 yield that
R
nSRn3 nSn3
J o1 as n . 5.82
n J n
Then, by 5.80, 5.81, and 5.82 we get

k/n
n Lk, J LJ w n k n Bn s
2 k Bn n k ds
n J 1 1 0 s
1/n
w n k k/n
JsBn sds
Bn op 1
1 1 k n n J
5.83
1
wR n n Bn s
2 Bn 1 ds
1 1 n 1/n 1 s

wR n
Bn 1 op 1.
1 1 n
k, J LJ/n J will be computed by

The asymptotic variance of nL
k/n k/n
2 n mins, t st
02 lim 2
w 4 k ds dt
n st
1 1 0 0
2 1/n dcs 1/n mins, t stdcs

1 nk k k/n k/n
w
2
4 k n 1 n
1 1 n J
2
1 n
wR
2
4 1
1 1 n n
1
2 n 1 mins, t st
wR2 4 ds dt
1 1 1/n 1/n st

1 n k/n t k/nt
2w
2
4 dt
1 1 k 0 t
k/n 1/n
n s st
2w 2 ds dct/n J
1 1 k 0 k/n s
k/n
1 n n s 1 /ns
2wwR 4 k ds
1 1 0 s
k/n 1
2 n n mins, t st
2wwR 4 k ds dt
1 1 0 1/n st
1/n
1 n k k
2w 2 s dct/n J
1 1 k k/n n n
1/n
1 n k k
2w 2 s dct/n J
1 1 k k/n n n
2
1 n n k k
2wwR 4 k n n 1
1 1 n
1
1 k/n k/ns
2wwR 4 ds
1 1 1/n 1s
1/n
1 n
2wR 2 ss 1 dct/n J
1 1 k/n n
1 1/n
n t st
2wR 2 ds dct/n J
1 1 1/n k/n 1 s

1
1 n 1 /n 1 /ns
2wR
2
4 ds .
1 1 1/n 1s
5.84
After calculation we get
2 2
22 1 1 22
02 w
2
4 w 2
4 1 w 2
R 4 w 2
R 4
1 1 1 1 1 1 1 1

1
2w2 4 2w 2 w
1 1 1 1

1
2w 2 w 2wR 2 w R
1 1 1 1

1 2
2wR 2 wR wR
2
4
1 1 1 1

22 12 2 1 2
w wR
2 2
4 1.
1 1 1 1
5.85
Finally, it is easy to verify that

1 2 2
w 2
wR2 . 5.86
2
This completes the proof of Theorem 3.5.
5.3. Proof of the Corollary

Straightforward by combining Theorem 3.5 and Lemma 5.4, we omit details.

We will only present details for the proof concerning the first part of Theorem 4.1. The proof
for the second part is very similar. For convenience we set
m2 : m
2 m2 , m3 : m
3 m3 . 5.87
Then we have
3 m3 m2 m3 m3 m2
m
1 1 . 5.88
2 m2
m 2 m2
m
2 is consistent estimator of m2 , then for all large n

Since m

m3 m3 m2
1 1 1 oP 1 , 5.89
m2 m22
and therefore

1 1
n 1
C1n C2n op 1, 5.90
n J3 m2
where

nm3 m3 nm2
C1n : , C2n : . 5.91
n J3 m2 n J3
In view of 5.83 we may write that for all large n
k/n
w1 n k n Bn s
C1n Bn ds
2 12 k n k 0 s
1/n
w1 n k k/n
JsBn sds
Bn op 1
2 1 k n n J
5.92

wR,1 n n 1 Bn s
Bn 1 ds
2 12 n 1/n 1 s

wR,1 n
Bn 1 op 1,
2 1 n
and, by Lemma 5.4, we infer that n J3 /n J2 10/9 as n ,
k/n
10m2 w1 n k n Bn s
C2n Bn ds
9m3 2 12 k n k 0 s
1/n
w1 n k k/n
JsBn sds
Bn op 1
2 1 k n n J
1 5.93
wR,1 n n Bn s
2 Bn 1 ds
1 1 n 1/n 1 s

wR,1 n
Bn 1 op 1,
2 1 n
with
12 2 1/ 3/21/

w1 : 2/ 32/
, wR,1 : q/p w1 . 5.94
2 1 q/p
By the same arguments as the proof of Theorem 3.5 we infer that
1 1
n D
N 0, V12 as n . 5.95
/n3/2 Xn,n
This achieves the proof of Theorem 4.1.

Theorem 4.2 is just an application of Theorem 3.5, we omit details.
Acknowledgments
The authors thank anonymous referees for their insightful comments and valuable
suggestions. The first author is indebted to Ricardas Zitikis for his constructive advice and
also for having introduced him the first author to the Actuarial Research Group ARG in
London, Ontario.
References
1 R. J. Serfling, Approximation Theorems of Mathematical Statistics, John Wiley & Sons, New York, NY,
USA, 1980.
2 H. Cherno, J. L. Gastwirth, and M. V. Johns, Asymptotic distribution of linear combinations of
functions of order statistics with applications to estimation, Annals of Mathematical Statistics, vol. 38,
pp. 5272, 1967.
3 J. R. M. Hosking, L-moments: analysis and estimation of distributions using linear combinations of
order statistics, Journal of the Royal Statistical Society Serie B, vol. 52, pp. 105124, 1990.
4 E. A. H. Elamir and A. H. Seheult, Trimmed L-moments, Computational Statistics and Data Analysis,
vol. 43, no. 3, pp. 299314, 2003.
5 J. R. M. Hosking, Some theory and practical uses of trimmed L-moments, Journal of Statistical
Planning and Inference, vol. 137, no. 9, pp. 30243039, 2007.
6 D. Denneberg, Non-Additive Measure and Integral, Kluwer Academic Publishers, Dordrecht, The
Netherlands, 1994.
7 S. S. Wang, Insurance pricing and increased limits ratemaking by proportional hazards transforms,
8 S. Wang, An actuarial index of the right-tail risk, North American Actuarial Journal, vol. 2, no. 2, pp.
88101, 1998.
9 B. L. Jones and R. Zitikis, Empirical estimation of risk measures and related quantities, North
10 B. L. Jones and R. Zitikis, Testing for the order of risk measures: an application of L-statistics in
actuarial science, Metron, vol. 63, pp. 193211, 2005.
11 B. L. Jones and R. Zitikis, Risk measures, distortion parameters, and their empirical estimation,
Insurance: Mathematics and Economics, vol. 41, no. 2, pp. 279297, 2007.
12 V. Brazauskas, B. L. Jones, M. L. Puri, and R. Zitikis, Nested L-statistics and their use in comparing
the riskiness of portfolios, Scandinavian Actuarial Journal, vol. 2007, pp. 162179, 2007.
13 V. Brazauskas, B. L. Jones, and R. Zitikis, Robust fitting of claim severity distributions and the
method of trimmed moments, Journal of Statistical Planning and Inference, vol. 139, no. 6, pp. 2028
2043, 2009.
14 F. Greselin, M. L. Puri, and R. Zitikis, L-functions, processes, and statistics in measuring economic
inequality and actuarial risks, Statistics and Its Interface, vol. 2, pp. 227245, 2009.
15 S. S. Wang, Ordering of risks under PH-transforms, Insurance: Mathematics and Economics, vol. 18,
no. 2, pp. 109114, 1996.
16 S. S. Wang, A class of distortion operators for pricing financial and insurance risks, Journal of Risk
and Insurance, vol. 67, no. 1, pp. 1536, 2000.
17 A. Necir, D. Meraghni, and F. Meddi, Statistical estimate of the proportional hazard premium of
loss, Scandinavian Actuarial Journal, vol. 3, pp. 147161, 2007.
18 A. Necir and D. Meraghni, Empirical estimation of the proportional hazard premium for heavy-
tailed claim amounts, Insurance: Mathematics and Economics, vol. 45, no. 1, pp. 4958, 2009.
19 G. R. Shorack and J. A. Wellner, Empirical Processes with Applications to Statistics, John Wiley & Sons,
20 P. J. Bickel, Some contributions to the theory of order statistics, in Proceedings of the 5th Berkeley
Symposium on Mathematical Statistics and Probability, vol. 1, pp. 575591, University of California Press,
Berkeley, Calif, USA, 1967.
21 G. R. Shorack, Asymptotic Normality of Linear Combinations of Function of Order Statistics, Annals
of Mathematical Statistics, vol. 40, pp. 20412050, 1969.
22 G. R. Shorack, Functions of order statistics, Annals of Mathematical Statistics, vol. 43, pp. 412427,
1972.
23 S. M. Stigler, Functions of order statistics, Annals of Mathematical Statistics, vol. 43, pp. 412427, 1969.
24 S. M. Stigler, Linear functions of order statistics with smooth weight functions, Annals of Statistics,
vol. 2, pp. 676693, 1974.
25 F. H. Ruymgaart and M. C. A. Van Zuijlen, Asymptotic normality of linear combinations of functions
of order statistics in the non-i.i.d. case, Indagationes Mathematicae, vol. 39, no. 5, pp. 432447, 1977.
26 P. K. Sen, An invariance principle for linear combinations of order statistics, Zeitschrift fur
Wahrscheinlichkeitstheorie und Verwandte Gebiete, vol. 42, no. 4, pp. 327340, 1978.
27 D. D. Boos, A dierential for L-statistics, Annals of Statistics, vol. 7, no. 5, pp. 955959, 1979.
28 D. M. Mason, Asymptotic normality of linear combinations of order statistics with a smooth score
function, Annals of Statistics, vol. 9, pp. 899908, 1981.
29 K. Singh, On asymptotic representation and approximation to normality of L-statistics. I, Sankhya
Series A, vol. 43, no. 1, pp. 6783, 1981.
30 L. Peng, Estimating the mean of a heavy tailed distribution, Statistics and Probability Letters, vol. 52,
no. 3, pp. 255264, 2001.
31 L. Peng, Empirical-likelihood-based confidence interval for the mean with a heavy-tailed distribu-
tion, Annals of Statistics, vol. 32, no. 3, pp. 11921214, 2004.
32 G. Samorodnitsky and M. S. Taqqu, Stable Non-Gaussian Random Processes: Stochastic Models with
Infinite Variance, Chapman & Hall, New York, NY, USA, 1994.
33 P. Levy, Calcul des Probabilites, Gauthier-Villars, Paris, France, 1925.
34 E. F. Fama, The Behavior of stock market prices, Journal of Business, vol. 38, pp. 34105, 1965.
35 R. Weron, Levy-stable distributions revisited: tail index > 2 does not exclude the Levy-stable
regime, International Journal of Modern Physics C, vol. 12, no. 2, pp. 209223, 2001.
36 M. Csorgo, S. Csorgo, L. Horvath, and D. M. Mason, Normal and stable convergence of integral
functions of the empirical distribution function, Annals of Probability, vol. 14, pp. 86118, 1986.
37 L. de Haan and U. Stadtmuller, Generalized regular variation of second order, Journal of the
Australian Mathematical Society, vol. 61, no. 3, pp. 381395, 1996.
38 I. Weissman, Estimation of parameters and large quantiles based on the k largest observations,
Journal of American Statistical Association, vol. 73, pp. 812815, 1978.
39 A. L. M. Dekkers and L. de Haan, On the estimation of the extreme-value index and large quantile
estimation, Annals of Statistics, vol. 17, pp. 17951832, 1989.
40 G. Matthys and J. Beirlant, Estimating the extreme value index and high quantiles with exponential
regression models, Statistica Sinica, vol. 13, no. 3, pp. 853880, 2003.
41 M. I. Gomes, F. Figueiredo, and S. Mendonca, Asymptotically best linear unbiased tail estimators
under a second-order regular variation condition, Journal of Statistical Planning and Inference, vol.
134, pp. 409433, 2005.
42 B. M. Hill, A simple approach to inference about the tail of a distribution, Annals of Statistics, vol. 3,
pp. 11361174, 1975.
43 D. M. Mason, Laws of the large numbers for sums of extreme values, Annals of Probability, vol. 10,
pp. 754764, 1982.
44 P. Deheuvels, E. Haeusler, and D. M. Mason, Almost sure convergence of the Hill estimator,
Mathematical Proceedings of the Cambridge Philosophy Society, vol. 104, pp. 371381, 1988.
45 A. Necir, A functional law of the iterated logarithm for kernel-type estimators of the tail index,
Journal of Statistical Planning and Inference, vol. 136, no. 3, pp. 780802, 2006.
46 M. Csorgo and D. M. Mason, Central limit theorems for sums of extreme values, Mathematical
Proceedings of the Cambridge Philosophical Society, vol. 98, pp. 547558, 1985.
47 J. Beirlant and J. Teugels, Asymptotic normality of Hills estimator, in Extreme Value Theory
(Oberwolfach, 1987), vol. 51 of Lecture Notes in Statistics, pp. 148155, Springer, New York, NY, USA,
1989.
48 A. L. M. Dekkers, J. H. J. Einmahl, and L. de Haan, A moment estimator for the index of an extreme-
value distribution, Annals of Statistics, vol. 17, pp. 18331855, 1989.
49 A. L. M. Dekkers and L. Dehaan, Optimal choice of sample fraction in extreme-value estimation,
Journal of Multivariate Analysis, vol. 47, no. 2, pp. 173195, 1993.
50 H. Drees and E. Kaufmann, Selecting the optimal sample fraction in univariate extreme value
estimation, Stochastic Processes and Their Applications, vol. 75, no. 2, pp. 149172, 1998.
51 J. Danielsson, L. de Haan, L. Peng, and C. G. de Vries, Using a bootstrap method to choose the sample
fraction in tail estimation, Journal of Multivariate Analysis, vol. 76, no. 2, pp. 226248, 2001.
52 S. Cheng and L. Peng, Confidence intervals for the tail index, Bernoulli, vol. 7, no. 5, pp. 751760,
2001.
53 C. Neves and M. I. F. Alves, Reiss and Thomas automatic selection of the number of extremes,
Computational Statistics and Data Analysis, vol. 47, no. 4, pp. 689704, 2004.
54 R. Ihaka and R. Gentleman, R: a language for data analysis and graphics, Journal of Computational
and Graphical Statistics, vol. 5, no. 3, pp. 299314, 1996.
55 D. M. Mason and G. R. Shorack, Necessary and sucient conditions for asymptotic normality of
trimmed L-statistics, Annals of Probability, vol. 20, pp. 17791804, 1992.
56 S. Csorgo, L. Horvath, and D. M. Mason, What portion of the sample makes a partial sum
asymptotically stable or normal? Probability Theory and Related Fields, vol. 72, no. 1, pp. 116, 1986.
57 L. de Haan and T. T. Pereira, Estimating the index of a stable distribution, Statistics and Probability
Letters, vol. 41, no. 1, pp. 3955, 1999.
58 R. A. Groeneveld, A class of quantile measures for kurtosis, American Statistician, vol. 52, no. 4, pp.
325329, 1998.
59 E. Seneta, Regularly Varying Functions, vol. 508 of Lecture Notes in Mathematics, Springer, Berlin,
Germany, 1976.
60 L. de Haan, On Regular Variation and Its Application to the Weak Convergence of Sample Extremes, vol.
32 of Mathematical Center Tracts, Center for Mathematics and Computer Science, Amsterdam, The
Netherlands, 1970.
61 A. A. Balkema and L. de Haan, Limit laws for order statistics, in Limit Theorems of Probability Theory
(Colloquium, Keszthely, 1974), vol. 11 of Colloquia Mathematica Societatis Janos Bolyai, pp. 1722, North-
Holland, Amsterdam, The Netherlands, 1975.
62 M. Csorgo and D. M. Mason, On the asymptotic distribution of weighted uniform empirical and
quantile processes in the middle and on the tails, Stochastic Processes and their Applications, vol. 21,
no. 1, pp. 119132, 1985.
doi:10.1155/2010/965672
Research Article
POT-Based Estimation of the Renewal Function of
Interoccurrence Times of Heavy-Tailed Risks
Abdelhakim Necir,1 Abdelaziz Rassoul,2 and Djamel Meraghni1

1
Laboratory of Applied Mathematics, University Mohamed Khider, P.O. Box 145, Biskra 07000, Algeria
2
Ecole Nationale Superieure dHydraulique, Guerouaou, BP 31, Blida 09000, Algeria
Correspondence should be addressed to Abdelaziz Rassoul, rsl [email protected]
Copyright q 2010 Abdelhakim Necir et al. This is an open access article distributed under the
Making use of the peaks over threshold POT estimation method, we propose a semiparametric
estimator for the renewal function of interoccurrence times of heavy-tailed insurance claims with
infinite variance. We prove that the proposed estimator is consistent and asymptotically normal,
and we carry out a simulation study to compare its finite-sample behavior with respect to the
nonparametric one. Our results provide actuaries with confidence bounds for the renewal function
of dangerous risks.
1. Introduction
Let X1 , X2 , . . . be independent and identically distributed iid positive random variables
rvs, representing claim interoccurrence times of an insurance risk, with common
distribution function df F having finite mean and variance 2 . Let

X1 Xm , m 1, 2, . . . ,
Sm : 1.1
0, m0
be the claim occurrence times, and define the number of claims recorded over the time
interval 0, t by
Nt : max{m 0, Sm t}. 1.2

The corresponding renewal function is defined by

Rt : ENt F k t, t > 0, 1.3
k1
where F k is the k-fold convolution of F for k 1.

The renewal theory has proved to be a powerful tool in stochastic modeling in a wide
variety of applications such as reliability theory, where a renewal process is used to model the
successive repairs of a failed machine see 1, risk theory, where a renewal process is used
to model the successive occurrences of risks see 2, 3, inventory theory, where a renewal
process is used to model the successive times between demand points see 4, manpower
planning, where a renewal process is used to model the sequence of resignations from a
given job see 5, and warranty analysis, where a renewal process is used to model the
successive purchases of a new item following the expiry of a free-replacement warranty see
6. Therefore, the need for renewal function estimates seems more than pressing in many
practical problems. For a summary of renewal theory, one refers to Feller 7, Asmussen 8,
and Resnick 9.
Statistical estimation of the renewal function has been considered in several ways.
Using a nonparametric approach, Frees 10 introduced two estimators based on the
empirical counterparts of F and F k by suitably truncating the sum in 1.3. Zhao and Subba
Rao 11 proposed an estimation method based on the kernel estimate of the density and the
renewal equation. A histogram-type estimator, resembling to the second estimator of Frees,
was given by Markovich and Krieger 12.
When EX 2 , Sgibnev 13 gave an asymptotic approximation of 1.3 as follows:
t
t 1
Rt Fxdx dy, 1.4
EX 2 0 y
with F : 1 F being the tail of F.

By replacing F by its empirical counterpart Fn in 1.4, Bebbington et al. 14 recently
proposed a nonparametric estimator for Rt in the case where F is of infinite variance, given
by
t
n t : t 1
R F n xdx dy, 1.5

2 0 y
where and 2 , respectively, represent the first and second sample moments of F. Their main
result says that whenever F belongs to the domain of attraction of a stable law S with 1/2 <
< 1 see, e.g., 15, the df of R n t converges, for suitable normalizing constants, to S . This
result provides confidence bounds for Rt with respect to the quantiles of S .
In general, people prefer estimators having simple formulas and carrying some kind
of asymptotic normality property in order to facilitate confidence interval construction. From
this point of view, the estimator R n t may not be as satisfactory to the users as it should be.

Then an alternative estimator to Rn t would be more useful in practice. Our task is to use the
extreme value theory tools to construct such an alternative estimator.
Indeed, an important class of models having infinite second-order moments is the set
of heavy-tailed distributions e.g., Pareto, Burr, Student, etc.. A df F is said to be heavy-tailed
with tail index > 0 if

Fx cx1/ 1 x Lx , as x , 1.6
for 0, 1, > 0, and some real constant c, with L a slowly varying function at infinity,
that is, Ltx/Lx 1 as x for any t > 0. For details on these functions, see Chapter 0
in Resnick 16 or Seneta 17. Notice that when 1/2, 1 we have < and EX 2 .
In this case, an asymptotic approximation of the renewal function Rt is given in 1.4.
Prior to Sgibnev 13, Teugels 18 obtained an approximation of Rt when F is heavy-
tailed with tail index 1/2, 1:
t 2 t2 Ft
Rt 2 , as t . 1.7
1 2 1
Extreme value theory allows for an accurate modeling of the tails of any unknown
distribution, making the semiparametric statistical inference more accurate for heavy-tailed
distributions. Indeed, the semiparametric approach permits extrapolating beyond the largest
value of a given sample while the nonparametric one does not since the empirical df vanishes
outside the sample. This represents a big handicap for those dealing with heavy-tailed data.
Extreme value theory has two aspects. The first one consists in approximating the tail
distribution by the generalized extreme value GEV distribution, thanks to Fisher-Tippett
theorem see 19, 20. The second aspect commonly known as POT method is based on
Balkema-de Haan result which says that the distribution of the excesses over a fixed threshold
is approximated by the generalized Pareto distribution GPD see 21, 22. Those interested
in extreme value theory and its applications are referred to the textbooks of de Haan and
Ferriera 23 and Embrechts et al. 24. In our situation, we have a fixed threshold equal
to the horizon t tn see Section 3. Therefore, the POT method would be the appropriate
choice to derive an estimator for Rt by exploiting the heavy-tail property of df F used in
approximation 1.4. The asymptotic normality of our estimator is established under suitable
assumptions.
The remainder of the paper is organized as follows. In Section 2, we introduce the GPD
approximation, mostly known as the POT method. A new estimator of the renewal function
Rt is proposed in Section 3, along with two main results on its limiting behavior. Section 4
is devoted to a simulation study. The proofs are postponed until Section 5.
2. GPD Approximation
The distribution of the excesses, over a fixed threshold t, pertaining to df F is defined by

Ft y : P X1 t y | X1 > t , for y > 0. 2.1
It is shown, in Balkema and de Haan 21 and Pickands 22, that Ft is approximated by a
generalized Pareto distribution GPD function G, with shape parameter R and scale
parameter t > 0, in the following sense:

supFt y G, y O t Lt , as t , 2.2
y>0
where t Lt 0 as t for any > 0. The GPD function G, is a two-parameter df

defined by

y 1/

1 1 ,
/ 0,

G, y 2.3

y
1 exp , 0,

for 0 y < if 0 and 0 y < / if < 0.

Let Y1 , . . . , YN be iid rvs with exact GPD G, . It is well known by standard arguments
see, e.g., 25, Chapter 9 that there exists, with probability 1 as N tends to infinity, a local
maximum N , N for the Log-Likelihood of G, s density based on the sample Y1 , . . . , YN .
In this case, by Theorem 3.7 page 447 in the work of Lehmann and Casella 26, we infer that
N and N are consistent estimators of and . Moreover, these estimators are asymptotically
normal provided that > 1/2. The extension to 1/2 was investigated by Smith 27.
Suppose now that Y1 , . . . , YN are drawn not from G, , but from Ft . In view of
the asymptotic approximation 2.2, Smith 27 has proposed estimates for , via the
Maximum Likelihood approach. The obtained estimators N , N are solutions of the
following system:

1 N
yi
log 1 ,
N i1
2.4
1 N
yi / 1
,
N i1 1 yi / 1
where y1 , . . . , yN is a realization of Y1 , . . . , YN .

Letting t tN as N and N tN and making use of 2.2, Smith 28
established, in Theorem 3.2, the asymptotic normality of N , N as follows:

N

1 D
N
N N2 0, Q1
as N , 2.5
N
where

1
2 1
Q 1 , 2.6
1 1

provided that Nt LtN 0 as N and x x Lx is nonincreasing near infinity.
N D
In the case NtN LtN 0, the limiting distribution in 2.5 is biased. Here denotes
convergence in distribution and N2 , stands for the bivariate normal distribution with
mean vector and covariance matrix .
3. Estimating the Renewal Function in Infinite Time

Since we are interested in the renewal function in infinite time, we must assume that time
t is large enough and for asymptotic considerations, we will assume that t depends on the
sample size n. That is, t tn , with tn as n . Relation 1.7 suggests that in order
to construct an estimator of Rtn , we need to estimate , and Ftn . Let n nt be the
number of Xi s, which are observed on horizon tn and denoted by
Ntn : card{Xi > tn : 1 i n}, 3.1
the number of exceedances over tn , with cardK being the cardinality of set K. Notice that
Ntn is a binomial rv with parameters n and pn : Ftn for which the natural estimator is
pn : Nt /n.
Select, from the sample X1 , . . . , Xn , only those observations Xi1 , . . . , XiNtn that exceed
tn . The Nt excesses
Ej:n : Xij tn , j 1, . . . , Ntn 3.2
are iid rvs with common df Ftn . As seen in Section 2, the maximum likelihood estimators
n , n are solutions of the following system:

1 vn ej:n
log 1 ,
vn j1
3.3
1 vn ej:n / 1
,
v j1 1 ej:n / 1
where vn is an observation of Ntn and the vector e1:n , . . . , evn :n a realization of

E1:n , . . . , ENtn :n . Regarding the distribution mean EX1 , we know that, for
0, 1/2, X1 has finite variance and therefore could naturally be estimated by the sample
mean X : n1 Sn which, by the Central Limit Theorem CLT, is asymptotically normal.
Whereas for 1/2, 1, X1 has infinite variance, in which case the CLT is no longer
valid. This case is frequently met in real insurance data see, e.g., 29. Using the GPD

approximation, Johansson 30 has proposed an alternative estimator for 0
xdFx. For
each n 1, we write as the sum of two components:
tn
n : xdFx, n : xdFx tn sdFtn s. 3.4
0 tn 0
Johansson 30 defined his estimator of , by estimating both Fx and Ftn s, as follows:
tn
J 3.5
n :
xdFn x tn sdFtn s,
0 0

where Fn is the empirical df based on the sample X1 , . . . , Xn and Ftn s is an estimate of
Ftn s obtained from the relation
Ftn s
F tn s , s > 0, 3.6
Ftn
which implies that Ftn s pn F tn s, s > 0. Approximation 2.2 motivates us to estimate

F tn s by F tn s : Gn ,n s, s > 0. Hence, an estimate of Ftn s is
3.7
Ftn s : pn Gn ,n s, s > 0.
By integrating 3.5, we get

J 1 n
n
n
Xi 1{Xi tn } pn tn
n i1 1 n 3.8
n n ,
:
with n 0, 1 with large probability. Here, 1K denotes the indicator function of set K.
n , n , and pn for , and Ftn in 1.7 yields the following
J
Respectively, substituting
estimator for the renewal function Rtn
n2 t2n pn

n tn : tn
R

. 3.9
J
1 n 2n 1
J2
n
n

n tn is given by the following two theorems.

The asymptotic behavior of R
Theorem 3.1. Let F be a df fulfilling 1.6 with 1/2, 1. Suppose that L is locally bounded in
x0 , for x0 0 and x x Lx is nonincreasing near infinity, for some > 0. Then, for any
tn On /4
with 0, 1, one has

n tn Rtn OP n/21/41/2 ,
R as n . 3.10
Theorem 3.2. Let F be as in Theorem 3.1. Then for any tn On/4 with 4/1 2, 1, we
have

D
n
Rn tn Rtn N0, 1, as n , 3.11
sn tn
where
2 2
pn 1 pn n pn 1 pn n
s2n :
12 2 1 tn 2 3
n2 1 n 1 2
3.12
12 n2 pn3 1 n pn2 1 pn n
3
n2 1 2 n2 1 2
with
1 22 tn pn
1 : ,
2 3 1 2 1
2 tn
2 : , 3.13
2 1 2 1

tn pn 43 32
3 : 2 2 ,
1 2 1 1 2 1
pn : Ftn , n : tn , and n2 : VarX1 1{X1 tn } .
4. Simulation Study
In this section, we carry out a simulation study by means of the statistical software R, see
31 to illustrate the performance of our estimation procedure, through its application to
sets of samples taken from two distinct Pareto distributions Fx 1 x1/ , x > 1 with tail
indices 3/4 and 2/3. We fix the threshold at 4, which is a value above the intermediate
statistic corresponding to the optimal fraction of upper-order statistics in each sample. The
latter is obtained by applying the algorithm of Cheng and Peng 32. For each sample size,
we generate 200 independent replicates. Our overall results are then taken as the empirical
means of the values in the 200 repetitions.
A comparison with the nonparametric estimator is done as well. In the graphical
illustration, we plot both estimators versus the sample size ranging from 1000 to 20000.
6 6
4 4
2 2
0 0
0 5000 10000 15000 20000 0 5000 10000 15000 20000
New estimator New estimator

Sample estimator Sample estimator
a b
Figure 1: Plots of the new and sample estimators of the renewal function, of interoccurrence times of
Pareto-distributed claims with tail indices 2/3 a and 3/4 b, versus the sample size. The horizontal line
represents the true value of the renewal function Rt evaluated at t 4.
Table 1: Semiparametric and nonparametric estimates of the renewal function of interoccurrence times
of Pareto-distributed claims with shape parameter 3/4. Simulations are repeated 200 times for dierent
sample sizes.
True value R 1.708

Semiparametric R
Nonparametric R
Sample size
Mean Bias RMSE Mean Bias RMSE
1000 1.696 0.013 0.250 2.141 0.433 0.553
2000 1.719 0.011 0.183 1.908 0.199 0.288
5000 1.705 0.003 0.119 1.686 0.022 0.168
Figure 1 clearly shows that the new estimator is consistent and that it is always
better than the nonparametric one. For the numerical investigation, we take samples of sizes
1000, 2000 and 5000. In each case, we compute the semiparametric estimate R as well as the

nonparametric estimate R. We also provide the bias and the root mean squared error rmse.
The results are summarized in Tables 1 and 2 for 3/4 and 2/3 respectively.
We notice that, regardless of the tail index value and the sample size, the semiparametric
estimation procedure is more accurate than the nonparametric one.
5. Proofs
The following tools will be instrumental for our needs.
Table 2: Semiparametric and nonparametric estimates of the renewal function of interoccurence times
of Pareto-distributed claims with shape parameter 2/3. Simulations are repeated 200 times for dierent
sample sizes.
True value R 2.222

Semiparametric R
Nonparametric R
Sample size
Mean Bias RMSE Mean Bias RMSE
1000 2.265 0.042 0.185 2.416 0.193 0.229
2000 2.247 0.024 0.157 2.054 0.167 0.223
5000 2.223 0.001 0.129 2.073 0.149 0.192
Proposition 5.1. Let F be a df fulfilling 1.6 with 1/2, 1, > 0, and some real c. Suppose that
L is locally bounded in x0 , for x0 0. Then for n large enough and for any tn On/4 ,
0, 1, one has
pn c1 o1n/4 ,

n2 O n/21/2 ,
5.1
s2n O n/21/2 ,
/8/41/2
npn t
n Ltn O n ,
where pn , n2 , and s2n are those defined in Theorem 3.2.
Lemma 5.2. Under the assumptions of Theorem 3.2, one has, for any real numbers u1 , u2 , u3 and
u4 ,

n
n 1 n
p n p n
Eexp iu1 n n i npn u2 , u3 n

iu4

n
n pn 1 pn
5.2
% 2 &
u1 1 1
u2 u24
exp u2 , u3 Q , as n ,
2 2 u3 2
where i2 1.
Proof of the Proposition. We will only prove the second result, the other ones are straightfor-
ward from 1.6. Let x0 > 0 be such that Fx cx1/ 1 x Lx, for x > x0 . Then for n
large enough, we have
tn x0 tn
' (
E X1 1{X1 tn } xdFx xdFx xdFx. 5.3
0 0 x0
x0
Recall that < , hence 0
xdFx < . Making use of the proposition assumptions, we get
21/
EX1 1{X1 tn } O1 and EX12 1{X1 tn } Otn and therefore n2 On/2/21 .
Proof of Lemma 5.2. See Johansson 30.

Proof of Theorem 3.1. We may readily check that for all large n,

n tn Rn tn
R
An Bn Cn , 5.4
tn
where

J
2 tn p n n

1 J
An : J J2 n ,
n
n 2 1 2 1
n2 tn
Bn :

pn pn , 5.5
1 n 2n 1
J2
n

2
2 3 2

tn pn n

Cn :

n n .
n

J2
1 n 2n 1 1 2 1
Johansson 30 proved that there exists a bounded sequence kn such that
)
kn
OP n
J
n
, 5.6
n
J
n OP n/41/21/2 . The first result of the proposition yields that
hence

J
tn p n n OP n/423/21/2 . 5.7
J
Since /42 3/2 1/2 < 0, then tn pn
n oP 1. On the other hand, by the CLT
we have
*
pn
pn pn OP , 5.8
n
pn pn OP n/41/21/2 oP 1. On the other hand, Smith 28, yields

then tn
n OP t
n Ltn , 5.9
it follows that, n2 tn

pn pn OP n/4121/21/2 oP 1, therefore

n2 tn pn pn

oP 1,
1 n 2n 1
J2
n

tn pn n OP n/411 oP 1, 5.10

n tn pn n OP n/4121 oP 1,

pn n OP n/41 oP 1.
Thus,

n
P

tn pn n 0,
1 n 2n 1
J2
n

5.11
tn pn 2 n 32
2

P

n 0 as n .
1 n 2n 1 1 2 1
J2
n

n Rtn OP n/21/41/2 , as sought.

Therefore for all large n, we get Rt
Proof of Theorem 3.2. From the proof of Theorem 3.1, for all large n, it is easy to verify that

n tn Rn tn
R

J
1 1 oP 1 n
tn
5.12
2 1 oP 1 pn pn

3 1 oP 1 n ,
where
1 22 tn pn
1 ,
2 3 1 2 1
2 tn
2 , 5.13
2 1
2 1

tn p n 43 32
3 2 2 .
1 2 1 1 2 1

Multiplying by n/n and using the proposition and the lemma together with the continuous
mapping theorem, we find that

n n J
Rn tn Rn tn 1 1 oP 1 n

n tn n

n
2 1 oP 1 pn pn 5.14
n

n
3 1 oP 1 n .
n
On the other hand, from Johansson 30, we have for all large n

n
n J n n
n
n n tn
pn pn
n n 1 n n

5.15
pn n n pn n
n n n oP 1.
1 2 n 1 n

n tn Rn tn into
This enables us to rewrite n/n tn R

pn 1 pn
n n n pn pn
1 n

n 2 1 tn
n n 1 pn 1 pn
)
n pn pn n n
1 1
n 1 pn n
5.16
)
pn pn n n

3 1 2
n oP 1,
n 1 pn

1
2 1
Q 1 .
1 1
In view of Lemma 5.2, we infer that for all large n, the previous quantity is

pn 1 pn
n

1 W 1 2 1 tn W2
n 1
5.17

21 1 n pn pn 1 pn 1 pn n
W3 3 W4 oP 1,
n 1 n 1 2
where Wi i1,4 are standard normal rvs with EWi Wj 0 for every i, j 1, . . . , 4 with i /
j,
except for
+ ) ) ,
n n n

EW3 W4 E 21 1 1 n
pn n pn
+) ) ,
n n n
5.18
21 1 E 1 n
pn n pn

21 1 2 .

n tn Rn tn is Gaussian with mean zero with asymptotic
Therefore, the rv n/n tn R
variance
2
pn 1 pn n 21 12 n2 pn3
Kn2 :
12 2 1 t n
n2 1 n2 1 2
2 5.19
1 2 pn 1 pn n 21 n pn2 1 4 1 pn n
3 3 oP 1.
n2 1 2 1 n2 1 2
Observe now that Kn2 s2n oP 1, where s2n is that in 3.12, this completes the proof of
Theorem 3.2.
6. Conclusion
In this paper, we have proposed a new estimator for the renewal function of heavy-tailed
claim interoccurence times, via a semiparametric approach. Our considerations are based on
one aspect of the extreme value theory, namely, the POT method. We have proved that our
estimator is consistent and asymptotically normal. Moreover, simulations show that it is more
accurate than the nonparametric estimator given by Bebbington et al. 14.
Acknowledgment
The authors are grateful to the referees whose suggestions led to an improvement of the
paper.
References
1 D. R. Cox, Renewal Theory, Methuen, London, UK, 1962.
2 P. Embrechts, M. Maejima, and E. Omey, Some limit theorems for generalized renewal measures,
Journal of the London Mathematical Society, vol. 31, no. 1, pp. 184192, 1985.
3 P. Embrechts, M. Maejima, and J. Teugels, Asymptotic behavior of compound distributions, Astin
Bulletin, vol. 15, pp. 4548, 1985.
4 S. M. Ross, Stochastic Processes, Wiley Series in Probability and Mathematical Statistics: Probability
and Mathematical Statistics, John Wiley & Sons, New York, NY, USA, 1983.
5 D. J. Bartholomew and A. F. Forbes, Statistical Techniques for Man-Power Planning, Wiley, Chichester,
UK, 1979.
6 W. R. Blischke and E. M. Scheuer, Applications of renewal theory in analysis of the free-replacement
warranty, Naval Research Logistics Quarterly, vol. 28, no. 2, pp. 193205, 1981.
7 W. Feller, An Introduction to Probability Theory and Its Applications, vol. 2, Wiley, New York, NY, USA,
2nd edition, 1971.
8 S. Asmussen, Applied Probability and Queues, Wiley Series in Probability and Mathematical Statistics:
Applied Probability and Statistics, John Wiley & Sons, Chichester, UK, 1987.
9 S. Resnick, Adventures in Stochastic Processes, Birkhauser, Boston, Mass, USA, 1992.
10 E. W. Frees, Nonparametric renewal function estimation, The Annals of Statistics, vol. 14, no. 4, pp.
13661378, 1986.
11 Q. Zhao and S. Subba Rao, Nonparametric renewal function estimation based on estimated
densities, Asia-Pacific Journal of Operational Research, vol. 14, no. 1, pp. 115126, 1997.
12 N. M. Markovich and U. R. Krieger, Nonparametric estimation of the renewal function by empirical
data, Stochastic Models, vol. 22, no. 2, pp. 175199, 2006.
13 M. S. Sgibnev, On the renewal theorem in the case of infinite variance, Sibirski Matematicheski
Zhurnal, vol. 22, no. 5, pp. 178189, 1981.
14 M. Bebbington, Y. Davydov, and R. Zitikis, Estimating the renewal function when the second
moment is infinite, Stochastic Models, vol. 23, no. 1, pp. 2748, 2007.
15 V. M. Zolotarev, One-Dimensional Stable Distributions, vol. 65 of Translations of Mathematical Monographs,
American Mathematical Society, Providence, RI, USA, 1986.
16 S. I. Resnick, Extreme Values, Regular Variation, and Point Processes, vol. 4 of Applied Probability. A Series
of the Applied Probability Trust, Springer, New York, NY, USA, 1987.
17 E. Seneta, Regularly Varying Functions, Lecture Notes in Mathematics, Vol. 508, Springer, Berlin,
Germany, 1976.
18 J. L. Teugels, Renewal theorems when the first or the second moment is infinite, Annals of
Mathematical Statistics, vol. 39, pp. 12101219, 1968.
19 R. A. Fisher and L. H. C. Tippett, Limiting forms of the frequency distribution of the largest or
smallest member of a sample, Proceedings of the Cambridge Philosophical Society, vol. 24, pp. 180190,
1928.
20 B. Gnedenko, Sur la distribution limite du terme maximum dune serie aleatoire, Annals of
Mathematics, vol. 44, pp. 423453, 1943.
21 A. A. Balkema and L. de Haan, Residual life time at great age, Annals of Probability, vol. 2, pp.
792804, 1974.
22 J. Pickands III, Statistical inference using extreme order statistics, The Annals of Statistics, vol. 3, pp.
119131, 1975.
23 L. de Haan and A. Ferreira, Extreme Value Theory: An Introduction, Springer Series in Operations
Research and Financial Engineering, Springer, New York, NY, USA, 2006.
24 P. Embrechts, C. Kluppelberg, and T. Mikosch, Modelling Extremal Events for Insurance and Finance, vol.
33 of Applications of Mathematics, Springer, Berlin, Germany, 1997.
25 D. R. Cox and D. V. Hinkley, Theoretical Statistics, Chapman and Hall, London, UK, 1974.
26 E. L. Lehmann and G. Casella, Theory of Point Estimation, Springer Texts in Statistics, Springer, New
York, NY, USA, 2nd edition, 1998.
27 R. L. Smith, Maximum likelihood estimation in a class of nonregular cases, Biometrika, vol. 72, no.
1, pp. 6790, 1985.
28 R. L. Smith, Estimating tails of probability distributions, The Annals of Statistics, vol. 15, no. 3, pp.
11741207, 1987.
29 J. Beierlant, G. Matthys, and G. Dierckx, Heavy-tailed distributions and rating, Astin Bulletin, vol.
31, no. 1, pp. 3758, 2001.
30 J. Johansson, Estimating the mean of heavy-tailed distributions, Extremes, vol. 6, no. 2, pp. 91109,
2003.
31 R. Ihaka and R. Gentleman, R: a language for data analysis and graphics, Journal of Computational
and Graphical Statistics, vol. 5, no. 3, pp. 299314, 1996.
2001.
doi:10.1155/2010/596839
Research Article
Estimating the Conditional Tail Expectation in
the Case of Heavy-Tailed Losses
Abdelhakim Necir,1 Abdelaziz Rassoul,2 and Ricardas Zitikis3

1
Laboratory of Applied Mathematics, Mohamed Khider University of Biskra, 07000, Algeria
2
Ecole Nationale Superieure dHydraulique, Guerouaou, BP 31, Blida, 09000, Algeria
3
ON, Canada N6A5B7
Academic Editor: Edward Furman
Copyright q 2010 Abdelhakim Necir et al. This is an open access article distributed under the
The conditional tail expectation CTE is an important actuarial risk measure and a useful tool
in financial risk assessment. Under the classical assumption that the second moment of the loss
variable is finite, the asymptotic normality of the nonparametric CTE estimator has already been
established in the literature. The noted result, however, is not applicable when the loss variable
follows any distribution with infinite second moment, which is a frequent situation in practice.
With a help of extreme-value methodology, in this paper, we oer a solution to the problem by
suggesting a new CTE estimator, which is applicable when losses have finite means but infinite
variances.
1. Introduction
One of the most important actuarial risk measures is the conditional tail expectation CTE
see, e.g., 1, which is the average amount of loss given that the loss exceeds a specified
quantile. Hence, the CTE provides a measure of the capital needed due to the exposure to
the loss, and thus serves as a risk measure. Not surprisingly, therefore, the CTE continues
to receive increased attention in the actuarial and financial literature, where we also find its
numerous extensions and generalizations see, e.g., 28, and references therein. We next
present basic notation and definitions.
Let X be a loss random variable with cumulative distribution function cdf F.
Usually, the cdf F is assumed to be continuous and defined on the entire real line,
with negative loss interpreted as gain. We also assume the continuity of F throughout
the present paper. The CTE of the risk or loss X is then defined, for every t 0, 1,
by
CTEF t EX | X > Qt, 1.1
where Qt inf{x : Fx t} is the quantile function corresponding to the cdf F. Since the
cdf F is continuous, we easily check that
1
1
CTEF t Qsds. 1.2
1t t
Naturally, the CTE is unknown since the cdf F is unknown. Hence, it is desirable to
establish statistical inferential results such as confidence intervals for CTEF t with specified
confidence levels and margins of error. We shall next show how to accomplish this task,
initially assuming the classical moment assumption EX 2 < . Namely, suppose that we
have independent random variables X1 , X2 , . . . , each with the cdf F, and let X1:n < < Xn:n
denote the order statistics of X1 , . . . , Xn . It is natural to define an empirical estimator of
CTEF t by the formula
1
n t 1
CTE Qn sds, 1.3
1t t
where Qn s is the empirical quantile function, which is equal to the ith order statistic Xi:n
for all s i 1/n, i/n, and for all i 1, . . . , n. The asymptotic behavior of the estimator
n t has been studied by Brazauskas et al. 9, and we next formulate their most relevant
CTE
result for our paper as a theorem.
Theorem 1.1. Assume that EX 2 < . Then for every t 0, 1, we have the asymptotic normality
statement

n t CTEF t 1 td N 0, 2 t ,
n CTE 1.4
when n , where the asymptotic variance 2 t is given by the formula
1 1

t
2
min x, y xy dQxdQ y . 1.5
t t
The assumption EX 2 < is, however, quite restrictive as the following example
shows. Suppose that F is the Pareto cdf with index > 0, that is, 1 Fx x1/ for all x 1.
Let us focus on the case < 1, because when 1, then CTEF t for every t 0, 1.
Theorem 1.1 covers only the values 0, 1/2 in view of the assumption EX 2 < . When
1/2, 1, we have EX 2 but, nevertheless, CTEF t is well defined and finite since
EX < . Analogous remarks hold for other distributions with Pareto-like tails, an we shall
indeed work with such general distributions in this paper.
Namely, recall that the cdf F is regularly varying at infinity with index 1/ < 0 if
1 Ftx
lim x1/ 1.6
t 1 Ft
for every x > 0. This class includes a number of popular distributions such as Pareto,
generalized Pareto, Burr, Frechet, Student, and so forth, which are known to be appropriate
models for fitting large insurance claims, fluctuations of prices, log-returns, and so forth
see, e.g., 10. In the remainder of this paper, therefore, we restrict ourselves to this class
of distributions. For more information on the topic and, generally, on extreme value models
and their manifold applications, we refer to the monographs by Beirlant et al. 11, Castillo et
al. 12, de Haan and Ferreira 13, Resnick 14.
The rest of the paper is organized as follows. In Section 2 we construct an alternative,
called new, CTE estimator by utilizing an extreme value approach. In Section 3 we establish
the asymptotic normality of the new CTE estimator and illustrate its performance with a
little simulation study. The main result, which is Theorem 3.1 stated in Section 3, is proved in
Section 4.
2. Construction of a New CTE Estimator

We have already noted that the old estimator CTE n t does not yield the asymptotic
normality in the classical sense beyond the condition EX 2 < . Indeed, this follows by
n t becomes the sample mean of X1 , . . . , Xn , and thus the
setting t 0, in which case CTE

asymptotic normality of CTEn 0 is equivalent to the classical Central Limit Theorem CLT.
Similar arguments show that the finite second moment is necessary for having the asymptotic
n t at any fixed level t 0, 1. Indeed, note that
normality in the classical sense of CTE
the asymptotic variance 2 t in Theorem 1.1 is finite only if EX 2 < .
For this reason, we next construct an alternative CTE estimator, which takes into
account dierent asymptotic properties of moderate and high quantiles in the case of heavy-
tailed distributions. Hence, from now on we assume that 1/2, 1. Before indulging
ourselves into construction details, we first formulate the new CTE estimator:
1k/n
n t 1 kXnk ,n
CTE Qn sds , 2.1
1t t n1 t 1
where we use the simplest yet useful and powerful Hills 15 estimator
1
k
n log Xni1:n log Xnk:n 2.2
k i1
of the tail index 1/2, 1. Integers k kn {1, . . . , n} are such that k and k/n 0
when n , and we note at the outset that their choices present a challenging task. In
Figures 1 and 2, we illustrate the performance of the new estimator CTE n t with respect to
the sample size n 1, with the integers k kn chosen according to the method proposed by
Cheng and Peng 16. Note that when t increases through the values 0.25, 0.50, 0.75, and 0.90
10 14
8
10
6
4 6
2 2
0 5000 10000 15000 20000 0 5000 10000 15000 20000

a b
25 30
20
15
10
5 0
0 5000 10000 15000 20000 0 5000 10000 15000 20000

c d
Figure 1: Values of the CTE estimator CTE n t vertical axis versus sample sizes n horizontal axis
evaluated at the levels t 0.25, t 0.50, t 0.75, and t 0.90 panels ad, resp. in the Pareto case with
the tail index 2/3.
panels ad, resp., the vertical axes of the panels also increase, which reflects the fact that
the larger the t gets, the more erratic the new and old estimators become. Note also that
the empirical i.e., old estimator underestimates the theoretical CTEF t, which is a well
known phenomenon see 17.
We have based the construction of CTE n t on the recognition that one should estimate
moderate and high quantiles dierently when the underlying distribution is heavy-tailed. For
this, we first recall that the high quantile qs is, by definition, equal to Q1 s for suciently
small s. For an estimation theory of high quantiles in the case of heavy-tailed distributions
we refer to, for example, Weissman 18, Dekkers and de Haan 19, Matthys and Beirlant
20, Gomes et al. 21, and references therein. We shall use the Weissman estimator

k k
qs Xnk:n s , 0<s< , 2.3
n n
of the high quantile qs . Then we write CTEF t as the sum CTE1,n t CTE2,n t with the
two summands defined together with their respective empirical estimators CTE 1,n t and
2,n t as follows:
CTE
1k/n 1k/n
1 1 1,n t,
CTE1,n t Qsds Qn sds CTE
1t t 1t t
1 1 2.4
1 1 2,n t.
CTE2,n t Qsds q1s ds CTE
1t 1k/n 1 t 1k/n
14 20
15
10
10
6
5
2
0 5000 10000 15000 20000 0 5000 10000 15000 20000

a b
25
20 40
15
20
10
5 0
0 5000 10000 15000 20000 0 5000 10000 15000 20000

c d
Figure 2: Values of the CTE estimator CTE n t vertical axis versus sample sizes n horizontal axis
evaluated at the levels t 0.25, t 0.50, t 0.75, and t 0.90 panels ad, resp. in the Pareto case with
the tail index 3/4.
Simple integration gives the formula
2,n t kXnk

CTE
,
n . 2.5
n 1 1 t
Consequently, the sum CTE 1,n t CTE

2,n t is an estimator of CTEF t, and this is exactly the

estimator CTEn t introduced above. We shall investigate asymptotic normality of the new
estimator in the next section, accompanied with an illustrative simulation study.
3. Main Theorem and Its Practical Implementation

We start this section by noting that Hills estimator n has been thoroughly studied, improved,
and generalized in the literature. For example, weak consistency of n has been established
by Mason 22 assuming only that the underlying distribution is regularly varying at infinity.
Asymptotic normality of has been investigated under various conditions by a number of
researchers, including Csorgo and Mason 23, Beirlant and Teugels 24, Dekkers et al. 25,
see also references therein.
The main theoretical result of this paper, which is Theorem 3.1 below, establishes
asymptotic normality of the new CTE estimator CTE n t. To formulate the theorem, we need
to introduce an assumption that ensures the asymptotic normality of Hills estimator n .
Namely, the cdf F satisfies the generalized second-order regular variation condition with
second-order parameter 0 see 26, 27 if there exists a function at which does not
change its sign in a neighbourhood of infinity and is such that, for every x > 0,

1 1 Ftx 1/ x/ 1
lim x x1/ . 3.1
t at 1 Ft /
When 0, then the ratio on the right-hand side of 3.1 is interpreted as log x. For
statistical inference concerning the second-order parameter , we refer, for example, to Peng
and Qi 28, Gomes et al. 21, Gomes and Pestana 29. Furthermore, in the formulation of
Theorem 3.1, we shall also use the function Az 2 aUz, where Uz Q1 1/z.
Theorem 3.1. Assume that the cdf F satisfies condition 3.1 with 1/2, 1. Then for any
sequence of integers k kn such that k/n 0 and k1/2 An/k 0 when n ,
we have that for any fixed t 0, 1,

n CTEn t CTEF t 1 t
d N 0, 2 , 3.2
k/n1/2 Xnk:n
where the asymptotic variance 2 is given by the formula
4
2 4 . 3.3
1 2 1
The asymptotic variance 2 does not depend on t, unlike the variance in Theorem 1.1.
This is not surprising because the heaviness of the right-most tail of F makes the asymptotic
1 t
behaviour of t Qn sQsds heavier than the classical CLT-type behaviour of 0 Qn s
Qsds, for any fixed t. This in turn implies that under the conditions of Theorem 3.1,
statement 3.2 is equivalent to the same statement in the case t 0. The latter statement
concerns estimating the mean EX of a heavy-tailed distribution. Therefore, we can view
Theorem 3.1 as a consequence of Peng 30, and at the same time we can view results of
Peng 30 as a consequence of Theorem 3.1 by setting t 0 in it. Despite this equivalence, in
Section 4 we give a proof of Theorem 3.1 for the sake of completeness. Our proof, however,
is crucially based on a powerful technique called the Vervaat process see 3133, for details
and references.
To discuss practical implementation of Theorem 3.1, we first fix a significance level
0, 1 and use the classical notation z/2 for the 1 /2-quantile of the standard
normal distribution N0, 1. Given a realization of the random variables X1 , . . . , Xn e.g.,
claim amounts, which follow a cdf F satisfying the conditions of Theorem 3.1, we construct
a level 1 confidence interval for CTEF t as follows. First, we choose an appropriate
number k of extreme values. Since Hills estimator has in general a substantial variance
for small k and a considerable bias for large k, we search for a k that balances between the
two shortcomings, which is indeed a well-known hurdle when estimating the tail index. To
resolve this issue, several procedures have been suggested in the literature, and we refer to,
for example, Dekkers and de Haan 34, Drees and Kaufmann 35, Danielsson et al. 36,
Cheng and Peng 16, Neves and Fraga Alves 37, Gomes et al. 38, and references therein.
n t and 95% confidence intervals for CTEF t when 2/3.
Table 1: Point estimates CTE
t 0.75 CTEF t 7.005

n k n t
CTE error rmse lower upper cover length
1000 054 6.876 0.045 0.303 6.356 7.397 0.839 1.041
2000 100 6.831 0.025 0.231 6.463 7.199 0.882 0.736
5000 219 7.119 0.016 0.194 6.881 7.357 0.895 0.476
t 0.90 CTEF t 12.533
n k n t
1000 054 12.753 0.017 0.534 12.241 13.269 0.847 1.028
2000 100 12.487 0.003 0.294 12.137 12.838 0.841 0.701
5000 219 12.461 0.005 0.236 12.246 12.676 0.887 0.430
n t and 95% confidence intervals for CTEF t when 3/4.
Table 2: Point estimates CTE
t 0.75 CTEF t 9.719

n k n t
1000 051 9.543 0.018 0.582 8.589 9.543 0.854 0.954
2000 104 9.808 0.009 0.466 9.150 10.466 0.888 1.316
5000 222 9.789 0.007 0.410 9.363 10.215 0.915 0.852
t 0.90 CTEF t 18.494
n k n t
1000 051 18.199 0.015 0.989 17.437 18.960 0.874 1.523
2000 104 18.696 0.011 0.858 18.052 19.340 0.895 1.288
5000 222 18.541 0.002 0.798 18.092 18.990 0.925 0.898
In our current study, we employ the method of Cheng and Peng 16 for an appropriate value
k of the parameter k. Having computed Hills estimator and consequently determined
Xnk :n , we then compute the corresponding values of CTE n t and 2 , and denote them by
n
n t and 2 , respectively. Finally, using Theorem 3.1 we arrive at the following 1 -
CTE n
confidence interval for CTEF t:
k /n1/2 Xnk :n 2n
n t z/2
CTE . 3.4
1 t n
To illustrate the performance of this confidence interval, we have carried out a small-
scale simulation study based on the Pareto cdf Fx 1 x1/ , x 1, with the tail index
set to 2/3 and 3/4, and the level t set to 0.75 and 0.90. We have generated 200 independent
replicates of three samples of sizes n 1000, 2000, and 5000. For every simulated sample, we
have obtained estimates CTE n t. Then we have calculated the arithmetic averages over the
values from the 200 repetitions, with the absolute error error and root mean squared error
rmse of the new estimator C n t reported in Table 1 2/3 and Table 2 3/4. In
the tables, we have also reported 95%-confidence intervals 3.4 with their lower and upper
bounds, coverage probabilities, and lengths.
We note emphatically that the above coverage probabilities and lengths of confidence
intervals can be improved by employing more precise but, naturally, considerably more
complex estimators of the tail index. Such estimators are described in the monographs by
Beirlant et al. 11, Castillo et al. 12, de Haan and Ferreira 13, and Resnick 14. Since the
publication of these monographs, numerous journal articles have appeared on the topic. Our
aim in this paper, however, is to present a simple yet useful result that highlights how much
Actuarial Science and developments in Mathematical Statistics, Probability, and Stochastic
Processes are interrelated, and thus benefit from each other.
4. Proof of Theorem 3.1

We start the proof of Theorem 3.1 with the decomposition

n t CTEF t 1 t An,1 t An,2 ,
CTE 4.1
where
1k/n
An,1 t Qn s Qsds,
t
1 4.2
k/n
An,2 Xnk:n Qsds.
1 n 1k/n
We shall show below that there are Brownian bridges Bn such that
1k/n
nAn,1 t 0
Bn sdQs
oP 1, 4.3
k/n1/2 Q1 k/n k/n1/2 Q1 k/n

nAn,2 2 n k
2 k nB 1
k/n1/2 Q1 k/n 1 n
1 4.4
n Bn s
2 k ds oP 1.
1 1k/n 1 s
Assuming for the time being that statements 4.3 and 4.4 hold, we next complete the proof
of Theorem 3.1. To simplify the presentation, we use the following notation:
1k/n
0
Bn sdQs
W1,n 1/2
,
k/nQ1 k/n

2 n k
W2,n 2 k nB 1 , 4.5
1 n
1
n Bn s
W3,n 2 ds.
1 k 1k/n 1s
Hence, we have the asymptotic representation

n CTEn t CTEF t 1 t
W1,n W2,n W3,n oP 1. 4.6
k/n1/2 Q1 k/n
The sum W1,n W2,n W3,n is a centered Gaussian random variable. To calculate its asymptotic
variance, we establish the following limits:
2 4 2 2
E W1,n
2
, E W2,n
2
4 , E W3,n
2
4 ,
2 1 1 1
2 2 2
2EW1,n W2,n 2 , 2EW1,n W3,n 2 , 4.7
1 1
2 3
2EW2,n W3,n 4 .
1
Summing up the right-hand sides of the above six limits, we obtain 2 , whose expression in
terms of the parameter is given in Theorem 3.1. Finally, since Xnk:n /Q1 k/n converges
in probability to 1 see, e.g., the proof of Corollary in 39, the classical Sultskys lemma
completes the proof of Theorem 3.1. Of course, we are still left to verify statements 4.3 and
4.4, which make the contents of the following two subsections.
4.1. Proof of Statement 4.3

If Q were continuously dierentiable, then statement 4.3 would follow easily from the proof
of Theorem 2 in 39. We do not assume dierentiability of Q and thus a new proof is required,
which is crucially based on the Vervaat process see 3133, and references therein
t Qt
Vn t Qn s Qsds Fn x Fxdx. 4.8
0
Hence, for every t such that 0 < t < 1 k/n, which is satisfied for all suciently large n since
t is fixed, we have that
1k/n t
An,1 t Qn s Qsds Qn s Qsds
0 0
Q1k/n 4.9
k
Fn x Fxdx Vn 1 Vn t.
Qt n
It is well known see 3133 that Vn t is nonnegative and does not exceed Fn Qt
tQn t Qt. Since the cdf F is continuous by assumption, we therefore have that

nVn t |en t||Qn t Qt|, 4.10

where en t is the uniform empirical process nFn Qt FQt, which for large n looks
like the Brownian bridge Bn t. Note also that with the just introduced notation en , the
Q1k/n
integral on the right-hand side of 4.9 is equal to Qt en Fxdx. Hence,
Q1k/n
nAn,1 t Qt
en Fxdx

k/n1/2 Q1 k/n k/n1/2 Q1 k/n
|en 1 k/n||Qn 1 k/n Q1 k/n| 4.11
OP 1
k/n1/2 Q1 k/n
|en t||Qn t Qt|
OP 1 .
k/n1/2 Q1 k/n
We shall next replace the empirical process en by an appropriate Brownian bridge Bn in

the first integral on the right-hand side of 4.11 with an error term of magnitude oP 1,
and we shall also show that the second and third summands on the right-hand side of
4.11 are of the order oP 1. The replacement of en by Bn can be accomplished using, for
example, Corollary 2.1 on page 48 of Csorgo et al. 40, which states that on an appropriately
constructed probability space and for any 0 < 1/4, we have that
|en s Bn s|

sup 1/2
OP n . 4.12
1/ns11/n s1/2 1 s
This result is applicable in the current situation since we can always place our original
problem into the required probability space, because our main results are in probability.
Furthermore, since Qt x Q1 k/n, we have that t Fx 1 k/n. Hence, statement
4.12 implies that
Q1k/n Q1k/n Q1k/n

Qt
en Fxdx Qt
Bn Fxdx Qt 1 Fx1/2 dx
OP 1 .
k/n1/2 Q1 k/n k/n1/2 Q1 k/n n k/n1/2 Q1 k/n
4.13
Changing the variables of integration and using the property k/n1/2 Q1 k/n when
n , we obtain that
Q1k/n 1k/n
Bn Fxdx Bn sdQs

Qt
0
oP 1. 4.14
1/2 1/2
k/n Q1 k/n k/n Q1 k/n
The main term on the right-hand side of 4.14 is W1,n . We shall next show that the right-most
summand of 4.13 converges to 0 when n .
Changing the variable of integration and then integrating by parts, we obtain the
bound
Q1k/n 1k/n
1k/n
Qt 1 Fx1/2 dx 1 s1/2 Qs 1 s1/2 Qsds
t t
O1 .
n k/n1/2 Q1 k/n n k/n1/2 Q1 k/n n k/n1/2 Q1 k/n
4.15
We want to show that the right-hand side of bound 4.15 converges to 0 when n . For
this, we first note that
1k/n

1 s1/2 Qs 1 1 t1/2 Qt
t
0. 4.16
1/2
n k/n Q1 k/n k n k/n1/2 Q1 k/n
Next, with the notation u Q1 u/u1/2 , we have that
1k/n 1t
1 s1/2 Qsds 1 k/n sds
t
0 4.17
n k/n1/2 Q1 k/n k k/nk/n
when n , where the convergence to 0 follows from Result 1 in the Appendix of Necir
and Meraghni 39. Taking statements 4.154.17 together, we have that the right-most
summand of 4.13 converges to 0 when n .
Consequently, in order to complete the proof of statement 4.3, we are left to show that
the second and third summands on the right-hand side of 4.11 are of the order oP 1. The
third summand is of the order oP 1 because |en t
Qn t Qt| OP 1 and k/n1/2 Q1
k/n . Hence, we are only left to show that the second summand on the right-hand side
of equation 4.11 is of the order oP 1, for which we shall show that

|en 1 k/n| Qn 1 k/n

1/2 Q1 k/n 1 oP 1. 4.18
k/n
To prove statement 4.18, we first note that
|en 1 k/n| |en 1 k/n Bn 1 k/n| |Bn 1 k/n|

1/2
1/2
. 4.19
k/n k/n k/n1/2
The first summand on the right-hand side of bound 4.19 is of the order OP 1 due to
statement 4.12 with 0. The second summand on the right-hand side of bound 4.19
is of the order OP 1 due to a statement on page 49 of Csorgo et al. 40 see the displayed
bound just below statement 2.39 therein. Hence, to complete the proof of statement 4.18,
we need to check that
Qn 1 k/n
1 oP 1. 4.20
Q1 k/n
Observe that, for each n, the distribution of Qn 1k/n is the same as that of QEn1 1k/n,
where En1 is the uniform empirical quantile function. Furthermore, the processes {1 En1 1
s, 0 s 1} and {En1 s, 0 s 1} are equal in distribution. Hence, statement 4.20 is
equivalent to

Q 1 En1 k/n
1 oP 1. 4.21
Q1 k/n
From the Glivenko-Cantelli theorem we have that En1 k/n k/n 0 almost surely, which
also implies that En1 k/n 0 since k/n 0 by our choice of k. Moreover, we know from
Theorem 0 and Remark 1 of Wellner 41 that

sup s1 En1 s s oP 1, 4.22
1/ns1
from which we conclude that
nEn1 k/n
1 oP 1. 4.23
k
Since the function s

Q1 s is slowly varying at zero, using Potters inequality see the
5th assertion of Proposition B.1.9 on page 367 of de Haan and Ferreira 13,we obtain that

Q 1 En1 k/n nEn1 k/n
1 oP 1 4.24
Q1 k/n k
for any 0, . In view of 4.23, the right-hand side of 4.24 is equal to 1 oP 1, which
implies statement 4.21 and thus finishes the proof of statement 4.3.
4.2. Proof of Statement 4.4

The proof of statement 4.4 is similar to that of Theorem 2 in Necir et al. 42, though some
adjustments are needed since we are now concerned with the CTE risk measure. We therefore
present main blocks of the proof together with pinpointed references to Necir et al. 42 for
specific technical details.
We start the proof with the function Uz Q1 1/z that was already used in the
formulation of Theorem 3.1. Hence, if Y is a random variable with the distribution function
Gz 1 1/z, z 1, then UY QGY d X because GY is a uniform on the interval
0, 1 random variable. Hence,
k/n
k/n 1
An,2 UYnk:n U ds, 4.25
1 n 0 s
and so we have
k/n
nAn,2 1 UYnk:n n/k 0 U1/sds
k
k/n1/2 Q1 k/n 1 n Un/k Un/k
1 UYnk:n 1

k 4.26
1 n Un/k 1
2
1 s Uns/kds
1
k .
1 Un/k
We next show that the right-most term in 4.26 converges to 0 when n . For this reason,
we first rewrite the term as follows:

1 s2 Uns/kds 1 Uns/k
1
k k s
ds. 4.27
1 Un/k 1 s
2 Un/k
The right-hand side of 4.27 converges to 0 see notes on page 149 of Necir et al. 42 due to
the second-order condition 3.1, which can equivalently be rewritten as

1 Uzs s 1
lim s s

4.28
z Az Uz

for every s > 0, where Az 2 aUz. Note that kAn/k 0 when n . Hence, in
order to complete the proof of statement 4.4, we need to check that

1 UYnk:n 1 2 n k
k 2 Bn 1
1 n Un/k 1 1 k n
1 4.29
n Bn s
2 k ds oP 1.
1 1k/n 1 s
With Hills estimator written in the form

1
k
UYni1:n
n log , 4.30
k i1 UYnk:n
we proceed with the proof of statement 4.29 as follows:
1 UYnk:n
k 1
1 n Un/k
1 UYnk:n Ynk:n 1 Ynk:n
n
k k 1 k .
1 n Un/k n/k 1 n n/k 1 n
4.31
Furthermore, we have that
n k
1 1
UYni1:n Yni1:n
k log log
1 n 1 n k i1 UYnk:n Ynk:n
4.32
k
1
Yni1:n
log 1 .
1 n k i1 Ynk:n
Arguments on page 156 ofNecir et al. 42 imply that the first term on the right-hand side of
4.32
is of the order OP kAYnk:n , and a note on page 157 of Necir et al. 42 says that
k AYnk:n oP 1. Hence, the first term on the right-hand side of 4.32 is of the order
oP 1. Analogous considerations using bound 2.5 instead of 2.4 on page 156 of Necir et al.
42 imply that the first term on the right-hand side of 4.31 is of the order oP 1. Hence, in
summary, we have that
1 UYnk:n
1

Ynk:n

k 1 k 1
1 n Un/k 1 n n/k
4.33
k
1
Yni1:n
log 1 oP 1.
1 n k i1 Ynk:n
We now need to connect the right-hand side of 4.33 with Brownian bridges Bn . To this
end, we first convert the Y -based order statistics into U-based i.e., uniform on 0, 1 order
statistics. For this we recall that the cdf of Y is G, and thus Y is equal in distribution to G1 U,
which is 1/1 U. Consequently,
1 UYnk:n
1

1

k 1 k 1
1 n Un/k 1 n n/k1 Unk:n
4.34
k
1
1 Unk:n
log 1 oP 1.
1 n k i1 1 Uni1:n
Next we choose a sequence of Brownian bridges Bn see pages 158-159 in 42 and references
therein such that the following two asymptotic representations hold:

1 n k
k 1 Bn 1 oP 1,
n/k1 Unk:n k n
k
1
1 Unk:n n k
log 1 Bn 1 4.35
k i1 1 Uni1:n k n
1
n Bn s
ds oP 1.
k 1k/n 1 s
Using these two statements on the right-hand side of 4.34 and also keeping in mind that n
is a consistent estimator of see 22, we have that
1 UYnk:n
2 n k
k 1 Bn 1
1 n Un/k 1 k n
1 4.36
n Bn s
ds oP 1.
1 k 1k/n 1 s
Dividing both sides of equation 4.36 by 1 , we arrive at 4.29. This completes the proof
of statement 4.4 and of Theorem 3.1 as well.
Acknowledgments
Our work on the revision of this paper has been considerably influenced by constructive
criticism and suggestions by three anonymous referees and the editor in charge of the
manuscript, Edward Furman, and we are indebted to all of them. Results of the paper were
first announced at the 44th Actuarial Research Conference at the University of Wisconsin,
Madison, Wisconsin, July 30August 1, 2009. The authors are grateful to participants of this
most stimulating conference, organized by the Society of Actuaries, for generous feedback.
The research has been partially supported by grants from the Society of Actuaries SOA and
the Natural Sciences and Engineering Research Council NSERC of Canada.
References
1 M. Denuit, J. Dhaene, M. Goovaerts, and R. Kaas, Actuarial Theory for Dependent Risks: Measures, Orders
and Models, Wiley, Chichester, UK, 2005.
2 Z. M. Landsman and E. A. Valdez, Tail conditional expectations for elliptical distributions, North
3 M. R. Hardy and J. L. Wirch, The iterated CTE: a dynamic risk measure, North American Actuarial
Journal, vol. 8, no. 4, pp. 6275, 2004.
4 J. Cai and H. Li, Conditional tail expectations for multivariate phase-type distributions, Journal of
Applied Probability, vol. 42, no. 3, pp. 810825, 2005.
5 B. J. Manistre and G. H. Hancock, Variance of the CTE estimator, North American Actuarial Journal,
vol. 9, no. 2, pp. 129156, 2005.
6 E. Furman and Z. Landsman, Tail variance premium with applications for elliptical portfolio of
risks, Astin Bulletin, vol. 36, no. 2, pp. 433462, 2006.
7 E. Furman and R. Zitikis, Weighted risk capital allocations, Insurance: Mathematics & Economics, vol.
43, no. 2, pp. 263269, 2008.
8 E. Furman and R. Zitikis, Weighted pricing functionals with applications to insurance: an overview,
North American Actuarial Journal, vol. 13, pp. 483496, 2009.
9 V. Brazauskas, B. L. Jones, M. L. Puri, and R. Zitikis, Estimating conditional tail expectation with
actuarial applications in view, Journal of Statistical Planning and Inference, vol. 138, no. 11, pp. 3590
3604, 2008.
10 J. Beirlant, G. Matthys, and G. Dierckx, Heavy-tailed distributions and rating, Astin Bulletin, vol.
31, no. 1, pp. 3758, 2001.
11 J. Beirlant, Y. Goegebeur, J. Teugels, and J. Segers, Statistics of Extremes, Wiley Series in Probability and
Statistics, John Wiley & Sons, Chichester, UK, 2004.
12 E. Castillo, A. S. Hadi, N. Balakrishnan, and J. M. Sarabia, Extreme Value and Related Models with
Applications in Engineering and Science, Wiley Series in Probability and Statistics, Wiley-Interscience,
Hoboken, NJ, USA, 2005.
13 L. de Haan and A. Ferreira, Extreme Value Theory: An Introduction, Springer Series in Operations
14 S. I. Resnick, Heavy-Tail Phenomena: Probabilistic and Statistical Modeling, Springer Series in Operations
15 B. M. Hill, A simple general approach to inference about the tail of a distribution, The Annals of
Statistics, vol. 3, no. 5, pp. 11631174, 1975.
2001.
17 J. T. Kim and M. R. Hardy, Quantifying and correcting the bias in estimated risk measures, Astin
Bulletin, vol. 37, no. 2, pp. 365386, 2007.
18 I. Weissman, Estimation of parameters and large quantiles based on the k largest observations,
Journal of the American Statistical Association, vol. 73, no. 364, pp. 812815, 1978.
19 A. L. M. Dekkers and L. de Haan, On the estimation of the extreme-value index and large quantile
estimation, The Annals of Statistics, vol. 17, no. 4, pp. 17951832, 1989.
20 G. Matthys and J. Beirlant, Estimating the extreme value index and high quantiles with exponential
regression models, Statistica Sinica, vol. 13, no. 3, pp. 853880, 2003.
21 M. I. Gomes, F. Figueiredo, and S. Mendonca, Asymptotically best linear unbiased tail estimators
under a second-order regular variation condition, Journal of Statistical Planning and Inference, vol.
134, no. 2, pp. 409433, 2005.
22 D. M. Mason, Laws of large numbers for sums of extreme values, The Annals of Probability, vol. 10,
no. 3, pp. 754764, 1982.
23 S. Csorgo and D. M. Mason, Central limit theorems for sums of extreme values, Mathematical
Proceedings of the Cambridge Philosophical Society, vol. 98, no. 3, pp. 547558, 1985.
24 J. Beirlant and J. L. Teugels, Asymptotic normality of Hills estimator, in Extreme Value Theory
(Oberwolfach, 1987), J. Husler and R.-D. Reiss, Eds., vol. 51 of Lecture Notes in Statistics, pp. 148155,
Springer, New York, NY, USA, 1989.
25 A. L. M. Dekkers, J. H. J. Einmahl, and L. de Haan, A moment estimator for the index of an extreme-
value distribution, The Annals of Statistics, vol. 17, no. 4, pp. 18331855, 1989.
26 L. de Haan and U. Stadtmuller, Generalized regular variation of second order, Australian
Mathematical Society Journal Series A, vol. 61, no. 3, pp. 381395, 1996.
27 J. Geluk, L. de Haan, S. Resnick, and C. Starica, Second-order regular variation, convolution and the
central limit theorem, Stochastic Processes and Their Applications, vol. 69, no. 2, pp. 139159, 1997.
28 L. Peng and Y. Qi, Estimating the first- and second-order parameters of a heavy-tailed distribution,
Australian & New Zealand Journal of Statistics, vol. 46, no. 2, pp. 305312, 2004.
29 M. I. Gomes and D. Pestana, A simple second-order reduced bias tail index estimator, Journal of
Statistical Computation and Simulation, vol. 77, no. 5-6, pp. 487504, 2007.
30 L. Peng, Estimating the mean of a heavy tailed distribution, Statistics & Probability Letters, vol. 52,
no. 3, pp. 255264, 2001.
31 R. Zitikis, The Vervaat process, in Asymptotic Methods in Probability and Statistics, B. Szyszkowicz,
Ed., pp. 667694, North-Holland, Amsterdam, The Netherlands, 1998.
32 Y. Davydov and R. Zitikis, Generalized Lorenz curves and convexifications of stochastic processes,
Journal of Applied Probability, vol. 40, no. 4, pp. 906925, 2003.
33 Y. Davydov and R. Zitikis, Convex rearrangements of random elements, in Asymptotic Methods
in Stochastics, B. Szyszkowicz, Ed., vol. 44 of Fields Institute Communications, pp. 141171, American
Mathematical Society, Providence, RI, USA, 2004.
34 A. L. M. Dekkers and L. de Haan, Optimal choice of sample fraction in extreme-value estimation,
Journal of Multivariate Analysis, vol. 47, no. 2, pp. 173195, 1993.
35 H. Drees and E. Kaufmann, Selecting the optimal sample fraction in univariate extreme value
estimation, Stochastic Processes and Their Applications, vol. 75, no. 2, pp. 149172, 1998.
36 J. Danielsson, L. de Haan, L. Peng, and C. G. de Vries, Using a bootstrap method to choose the sample
fraction in tail index estimation, Journal of Multivariate Analysis, vol. 76, no. 2, pp. 226248, 2001.
37 C. Neves and M. I. Fraga Alves, Reiss and Thomas automatic selection of the number of extremes,
Computational Statistics & Data Analysis, vol. 47, no. 4, pp. 689704, 2004.
38 M. I. Gomes, D. Pestana, and F. Caeiro, A note on the asymptotic variance at optimal levels of a
bias-corrected Hill estimator, Statistics & Probability Letters, vol. 79, no. 3, pp. 295303, 2009.
39 A. Necir and D. Meraghni, Empirical estimation of the proportional hazard premium for heavy-
tailed claim amounts, Insurance: Mathematics & Economics, vol. 45, no. 1, pp. 4958, 2009.
40 M. Csorgo, S. Csorgo, L. Horvath, and D. M. Mason, Weighted empirical and quantile processes,
The Annals of Probability, vol. 14, no. 1, pp. 3185, 1986.
41 J. A. Wellner, Limit theorems for the ratio of the empirical distribution function to the true
distribution function, Zeitschrift fur Wahrscheinlichkeitstheorie und verwandte Gebiete, vol. 45, no. 1,
pp. 7388, 1978.
42 A. Necir, D. Meraghni, and F. Meddi, Statistical estimate of the proportional hazard premium of
loss, Scandinavian Actuarial Journal, no. 3, pp. 147161, 2007.
doi:10.1155/2010/543065
Research Article
An Analysis of the Influence of Fundamental
Values Estimation Accuracy on Financial Markets
Hiroshi Takahashi
Graduate School of Business Administration, Keio University, 4-1-1 Hiyoshi, Kohoku-ku,
Yokohama-city 223-8572, Japan
Correspondence should be addressed to Hiroshi Takahashi, [email protected]
Received 14 September 2009; Revised 17 December 2009; Accepted 17 February 2010
Academic Editor: Edward Furman
Copyright q 2010 Hiroshi Takahashi. This is an open access article distributed under the Creative
This research analyzed the influence of the dierences in the forecast accuracy of fundamental
values on the financial market. As a result of intensive experiments in the market, we made
the following interesting findings: 1 improvements in forecast accuracy of fundamentalists can
contribute to an increase in the number of fundamentalists; 2 certain situations might occur,
according to the level of forecast accuracy of fundamentalists, in which fundamentalists and
passive management coexist, or in which fundamentalists die out of the market, and furthermore;
3 where a variety of investors exist in the market, improvements in the forecast accuracy could
increase the number of fundamentalists more than the number of investors that employ passive
investment strategy. These results contribute to clarifying the mechanism of price fluctuations in
financial markets and also indicate one of the factors for the low ratio of passive investors in asset
management business.
1. Introduction
A growing body of studies regarding asset pricing have been conducted, and many
prominent theories have been proposed 14. Along with the advancement of these theories,
many arguments regarding securities investment in practical business aairs in finance have
been actively discussed. The theory of asset pricing and investment strategy for shares are
also currently being discussed with enthusiasm. The accurate valuation of fundamental
values of investment grade assets is one of significant interest for those investors that actually
make transactions in real financial markets. For example, many of institutional investors
have a number of security analysts in their own companies in order to try to evaluate the
fundamental values of each security.
Market eciency is a central hypothesis in traditional asset pricing theories and
there has been a large amount of discussion regarding it 5. For example, in the Capital
Asset Pricing Model CAPM, which is one of the most popular asset pricing theories,
equilibrium asset prices are derived on the assumption of ecient markets and rational
investors. CAPM indicates that the optimal investment strategy is to hold market portfolio1
2. Since it is very dicult for investors to get an excess return in an ecient market,
it is assumed to be dicult to beat market portfolio even though fundamental values
are estimated correctly based on public information 2, 6. On the other hand, passive
investment strategy, which tries to maintain an average return using benchmarks based
on market indices, is consistent with traditional asset pricing theories and is considered
to be an eective method in ecient markets. On the basis of such arguments, there
has been growing interest in passive investment strategy in the practical business aairs
of asset management. Many investors employ the passive investment strategy for their
investment.2
Recently, however, traditional financial theories have been criticized in terms of
their explanation power and the validity of their assumptions. Behavioral finance has
recently been in the limelight and many reports indicate that deviation from rational
decision-making can explain anomalies which cannot be explained with traditional financial
theories 710. Generally, investor behavior which is assumed in behavioral finance has
complicated rules for decision making compared to decision making based on expected
utility maximization. For this reason, in many cases, it is dicult to derive the influence
of investor behavior on prices analytically 11. In order to take such investors behavior
into account in analyzing financial markets, we need to introduce a dierent analytical
method.
In the area of computer science, Agent-Based Modeling has been proposed as an
eective approach to analyze the relation between microrules and macrobehavior 12. This
is a bottom-up approach that tries to describe macrobehavior of the entire system using
local rules. This approach is appropriate for analyzing a multiagent system in which a
great number of agents that act autonomously gather together.3 The agent-based approach is
applied in a wide variety of study fields such as engineering and biology, and many reports
have been made about analyses adopting this approach in the field of social science 13
17.
In the background of the above-mentioned arguments, the purpose of this research
is to clarify the influence of the dierence in the forecast accuracy of fundamental values
on financial markets by using the agent-based model for analysis. This analysis includes
the relationship between fundamentalists that invest based on fundamentals and passive
investment strategy. Section 2 describes the model used in this analysis. Section 3 shows the
results of the analysis. Section 4 summarizes this paper.
2. Model
A computer simulation of the financial market involving 1000 investors was used as
the model for this research. Shares and risk-free assets were the two types of assets
used along with the possible transaction methods. Several types of investors exist in the
market, each undertaking transactions based on their own stock evaluations. This market
was composed of three major stages, 1 generation of corporate earnings, 2 formation
of investor forecasts, and 3 setting transaction prices. The market advances through
repetition of these stages. The following sections describe negotiable transaction assets,
modeling of investor behavior, setting transaction prices, and natural selection rules in the
market.
Table 1: List of investor types.
No. Investor types

1 Fundamentalist
2 Forecasting by past average most recent 10 days
3 Forecasting by trend most recent 10 day
4 Passive investor
2.1. Negotiable Assets in the Market

This market has risk-free and risk assets. There are risky assets in which all profits gained
during each term are distributed to the shareholders. Corporate earnings yt are expressed
as yt yt1 1 t . However, they are generated according to the process of t N0, y2 with
share trading being undertaken after the public announcement of profits for the term 18.
Each investor is given common asset holdings at the start of the term with no limit placed on
debit and credit transactions 1000 in risk-free assets and 1000 in stocks. Investors adopt the
buy-and-hold method for the relevant portfolio as a benchmark4 to conduct decision-making
by using a one-term model.
2.2. Modeling Investor Behavior

Each type of investor handled in this analysis is organized in Table 1.5 The investors in
this market evaluate transaction prices based on their own forecasts for market tendency,
taking into consideration both risk and return rates when making decisions. Each investor
determines the investment ratio wti based on the maximum objective function fwi t as
shown below6 19:
2 2
s,i
f wti rt 1
int,i
wti rf 1 wti t1 wti . 2.1
int,i s,i
Here, rt 1 and t1 express the expected rate of return and risk for stocks as estimated by
each investor i. rf indicates the risk-free rate. wti represents the stock investment ratio of the
investor i for term t.
The expected rate of return for shares is calculated as follows 19:
2 f,i 2
s,i s,i
1 c1 t1 rt 1 1 t1 rtim
int,i
rt 1 2 2 . 2.2
s,i s,i
1 c1 t1 1 t1
f,i
Here, rt 1 Crtim expresses the expected rate of return, calculated from short-term expected rate
of return, and risk and gross current price ratio of stocks, respectively. c is a coecient that
adjusts the dispersion level of the expected rate of return calculated from risk and gross
current price ratio of stocks 19.
f,i f,i f,i
The short-term expected rate of return rt is obtained where Pt 1 , yt 1 is the equity
price and profit forecast for term t 1 is estimated by the investor, as shown below:

f,i f,i
f,i Pt 1 yt 1
rt 1 1 1 ti . 2.3
Pt
The short-term expected rate of return includes the error term ti N0, n2 reflecting that
even investors using the same forecast model vary slightly in their detailed outlook. The stock
f,i f,i
price Pt 1 , profit forecast yt 1 , and risk estimation methods are described in Section 2.2.2.
The expected rate of return obtained from stock risk and so forth is calculated from
i
stock risk t1 , benchmark equity stake Wt1 , investorsf degree of risk avoidance , and
risk-free rate rf , as shown below 19, 20:
2
i 2.4
rtim 2 t1 Wt1 rf .
2.2.1. Stock Price Forecasting Method

The fundamental value is estimated by using the dividend discount model, which is a well-
known model in the field of finance. Fundamentalists estimate the forecasted stock price
f,i
and forecasted profit from profit for the term yt and the discount rate as Pt 1 yt /,
f,i
yt 1 yt .
Forecasting based on trends involves forecasting the next term stock prices and profit
through extrapolation of the most recent stock value fluctuation trends. The next term stock
price and profit is estimated from the most recent trends of stock price fluctuation at1 from
f,i f,i
time point t 1 as Pt 1 Pt1 1 at1 2 , yt 1 yt 1 at1 .
Forecasting based on past averages involves estimating the next term stock prices and
profit based on the most recent average stock value.
2.2.2. Risk Estimation Method

In this analysis, each investor estimates risk from past price fluctuations. Specifically, stock
i h h
risk is estimated as t1 t1 common to each investor. Here, t1 represents the stock
volatility that is calculated from price fluctuation from the most recent 100 terms.
2.3. Determination of Transaction Prices

Transaction prices are determined as the price where stock supply and demand converge

M i i i
i1 Ft wt /Pt N. In this case, the total asset Ft of investor i is calculated from
transaction price Pt for term t, profit yt and total assets from the term t1, stock investment
i
ratio wt1 , and risk-free rate rf , as Fti Ft1
i i
wt1 i
Pt yt /Pt1 1 wt1 1 rf .
2.4. Rules of Natural Selection in the Market

The rules of natural selection can be identified in this market. The driving force behind these
rules is cumulative excess profit 21. The rules of natural selection go through the following
two stages: 1 the identification of investors who alter their investment strategy, and 2 the
actual alteration of investment strategy 17, 22.
Each investor determines the existence of investment strategy alteration based on the
most recent performance of each 5-term period after 25 terms have passed since the beginning
of market transactions. The higher the profit rate obtained most recently is, the lesser the
possibility of strategy alteration becomes. The lower the profit, the higher the possibility
becomes. Specifically, when an investor could not obtain a positive excess profit for the
benchmark portfolio profitability, they are likely to alter their investment strategy with the
probability below:7
pi min1, max100 r cum , 0. 2.5
Here, however, ricum is the cumulative excess profitability for the most recent benchmark of
investor i. Measurement was conducted for 1 term, 5 terms, and 25 terms, and the cumulative
excess profitability was a profitability of one-term conversion.
Regarding the determination of a new investment strategy, an investment strategy that
has a high cumulative excess profit for the most recent five terms forecasting type is more
likely to be selected. Where the strategy of the investor i is zi and the cumulative excess profit
for the most recent five terms is ricum , the probability pi that zi is selected as a new investment
arjcum 8
strategy is given as pi eari / M
cum
j1 e . Those investors who altered their strategies
make investments based on the new strategies after the next step.
3. Analysis Results
First of all, the case where investors make decisions based on past prices in the market is
analyzed. Specifically, a market where there are investors that make forecasts based on past
price trends and past price averages, as well as fundamentals, is analyzed. Afterwards, the
case where there are investors that conduct passive investment strategy in the market is
analyzed.
3.1. Where There Exist Investors That Forecast Based on

Past Price Fluctuations
First of all, the influence of dierences in the forecast accuracy of fundamentals on the
market was analyzed. Afterwards, the influence of the dierence in the forecast dispersion of
investors other than fundamentalists was analyzed.
3.1.1. Influence of the Dierence in the Forecast Accuracy of Fundamentals

This section analyzes the influence of the dierence in the forecast accuracy of fundamental-
ists on the market where there exist heterogeneous investors in the market. From the start,
there are a similar number of the three types of investors Table 1: Type 13 in the market. In
2500
2000
Stock price
1500
1000
500
0
0 250 500 750 1000 1250 1500 1750 2000
Time step
Market price
Fundmental value
Figure 1: Price transitions n 1%.
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step
Fundamentalist
Trend 10 days
Average 10 days
Figure 2: Transition of the number of investors n 1%.
the beginning, the case where the forecast dispersion of investors n is 1% is described, and
then the case where the forecast error of fundamentalists diers is described.
Where the Forecast Dispersion of Investors Is 1%

Figures 1 and 2 show the transitions of transaction prices and the number of investors. The
typical price transitions obtained in this analysis are shown with respect to the transition
of transaction prices. With regard to the transition of the numbers of investors, the average
value obtained by conducting the analysis 50 times was used the same being true in the
following analysis. Figure 1 shows that transaction prices are consistent with fundamental
values throughout the entire transaction period. Figure 2 confirms that the number of
fundamentalists increases as time goes on.
2500
2000
Stock price
1500
1000
500
0
0 250 500 750 1000 1250 1500 1750 2000
Time step
Market price
Fundmental value
Where the Forecast Error of Fundamentalists Is 2%

Figures 3 and 4 show the transaction prices and the transition of the number of
investors where the forecast error of fundamentalists increases. In this analysis, the forecast
error of fundamentalists n is 2%, and the forecast dispersion of investors other than
fundamentalists n is 1%.
With respect to the transitions of transaction prices, as shown in Figure 3, it can be
confirmed that transaction prices are consistent with fundamental values. Regarding the
transition of the number of investors, similarly, the number of fundamentalists increases
as time passes. When Figures 2 and 4 are compared regarding the rate of increase in the
number of fundamentalists, the rate of increase is significant where fundamentalists have
high accuracy in their forecasting. These results show that the better the forecast accuracy of
fundamentalists becomes, the greater the rate of increase in the number of fundamentalists.
The next section analyzes the case where the forecast error is 0%.9

Figure 5 shows the transition of the number of investors where the forecast error of
fundamentalists n is 0%.10 The forecast dispersion of investors other than fundamentalists
n is constant at 1%. Where Figure 5 is compared with Figures 2 and 4, the rate of increase
of the number of fundamentalists is fastest in the case of Figure 5. When it comes to the rate
of increase in the number of fundamentalists, the rate of increase goes up as the forecast
accuracy improves. These results show that the better the forecast accuracy becomes, the
more likely it is that fundamentalists can survive in the market. This result is consistent
with traditional financial theory. The influence of the dierence in the forecast accuracy of
fundamentalists was analyzed in this analysis. Whether or not fundamentalists can survive
in the market likely depends on the influence of the forecast dispersion of other investors.
To confirm the influence of other investors, the next section analyzes the influence of forecast
dispersion of investors other than fundamentalists.
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step
Fundamentalist
Trend 10 days
Average 10 days
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step
Fundamentalist
Trend 10 days
Average 10 days
3.1.2. Influence of Forecast Dispersion of Other Investors

Here, the influence of forecast dispersion of other investors on the rate of increase in the
number of fundamentalists is analyzed. The forecast accuracy of fundamentalists in this
analysis n is consistent at 1%.
Figures 6 and 7 show the transitions of the number of investors where the forecast
dispersion of investors other than fundamentalists is 2% and 3%. As the same with the
previous section with respect to the transition of the number of investors, the number of
fundamentalists increases as time passes. Furthermore, the rate of increase becomes faster as
the dispersion of investors other than fundamentalists becomes significant.11 These results
show that interaction with other investors should be taken into consideration in order to
clarify the mechanism of financial markets. In this sense, these results are highly suggestive.12
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 25 50 75 100 125 150 175
Time step
Fundamentalist
Trend 10 days
Average 10 days
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 25 50 75 100 125 150 175
Time step
Fundamentalist
Trend 10 days
Average 10 days
3.2. Where There Exist Investors That Conduct Passive Investment Strategy
This section analyzes the case where there exist investors that conduct passive investment
strategy. First of all, the influence of the dierence in the forecast accuracy of fundamentalists
is analyzed, and then another analysis considers more practical conditions.
3.2.1. Influence of the Dierence in Forecast Accuracy of Fundamentalists

This section analyzes the influence of dierences in the forecast accuracy of fundamentalists
on the market where there exist investors that conduct passive management. In the early
stages, there are a similar number of the four types of investors Table 1: Type 14 in the
market. First of all, the case where the forecast dispersion of investors n is 1% is described.
Afterwards, a case where the forecast error of fundamentalists is 0% is described.
2500
2000
Stock price
1500
1000
500
0
0 250 500 750 1000 1250 1500 1750 2000
Time step
Market price
Fundmental value
Where the Forecast Dispersion of Investors Is 1%

Figures 8 and 9 show the transitions of transaction prices and the numbers of investors.
Figure 9 shows that the number of investors that conduct passive investment strategy
increases as time passes, and all the investors are conducting passive investment strategy
after a certain period of time.13 These results support the eectiveness of conducting passive
investment strategy from the viewpoint of investment performance, which is consistent with
traditional asset pricing theories 2. Transaction prices, however, do not show fundamental
values from around the middle of the transaction period. The same trend can also be
confirmed in the case where the forecast accuracy of fundamentalists is 2%. Under the present
conditions, the investment behavior of fundamentalists and passive management is the same
on average. However, among fundamentalists, the number of investors that conduct passive
investment strategy increases due to forecast error of fundamentalists.14 Where the forecast
accuracy of fundamentalists is good, n 0, coexistence of fundamentalists and passive
management can be predicted. The next section analyzes the case where the forecast error of
fundamentalists n is 0%.

Figures 10 and 11 show the transitions of transaction prices and the number of investors
where the forecast error of fundamentalists n is 0%. In this analysis, the forecast dispersion
of investors other than fundamentalists n is constant at 1%. The price history shows
that traded prices are consistent with fundamental values throughout the entire transaction
period. Additionally, the transitions of the number of investors show that fundamentalists
coexist with those investors that conduct passive investment strategy in the market. Just as
with the present conditions, where there exist only two types of investors in the market,
those investors that conduct passive investment strategy and fundamentalists, and where
the forecast accuracy of fundamentalists is good n 0, as a result, investment behavior
of both investors becomes equal. Given this, both types of investors are likely to exist in the
market.
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step
Fundamentalist Average 10 days

Trend 10 days Passive investor
1200
1000
800
Stock price
600
400
200
0
0 250 500 750 1000 1250 1500 1750 2000
Time step
Market price
Fundmental value
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step


1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step

Figure 12: Transition of the number of investors mutation: 1%, n 1%.
The analysis conducted in this section confirmed that dierent situations could be
generated. For example, fundamentalists and investors who employ passive investment
strategy coexisted in the market according to the forecast accuracy level of fundamentalists
or fundamentalists could die out in the market. These results suggest that the dierence in
estimation accuracy of fundamentalists should have a significant impact on the market. Thus,
the results obtained in this analysis are very interesting.
3.2.2. Analysis That Considers the Actual Investment Environment

This section conducts analysis under conditions close to that of actual market conditions. In
real markets, investors do not always determine their investment strategy based only on past
investment performance. This section focuses on how to change investment strategy in order
to analyze the case where some investors randomly change investment strategy.15
First of all, the case where 1% of investors changes investment strategy in a random
manner is analyzed. Afterwards, another case, where there is an increase in the rate
of investors in changing investment strategy randomly, is analyzed. In the early stage,
there are a similar number of the four types of investors Table 1: Type 14 in the
market. The forecast dispersion of investors other than fundamentalists n is constant at
1%.
Where a 1% of Investors Randomly Changes Investment Strategy

Figure 12 shows the transitions of the number of investors where the forecast accuracy of
fundamentalists n is 1%. This shows that fundamentalists and investors that conduct
passive investment strategy coexist together in the market.16 Since investors who randomly
change investment strategy exist, fundamentalists and passive investors coexist in the market
when fundamentalists forecasts are not entirely accurate. The existence of a wide variety of
investors can make it possible for fundamentalists to obtain more excess earnings through
market transactions. As a result, the number of fundamentalists probably increases.17
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step

1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step

Figures 13 and 14 show the transitions of the number of investors where the forecast
accuracy of fundamentalists is 2% and 0%. These results can confirm that fundamentalists
and investors that conduct passive investment strategy coexist. In addition, a comparison
of Figures 13 and 14 shows that as the estimation accuracy of fundamentalists increases,
there is a corresponding increase in the number of fundamentalists over time. These
results are interesting. They show that the number of fundamentalists who can survive
in the market is significantly influenced by the estimation accuracy level of fundamental-
ists.
Where the Rate of Investors That Randomly Change Investment Strategy Increases
Figure 15 shows the transitions of the number of investors where the rate of investors
that randomly change investment strategy is 2%.18 If there is an increase in the rate of
1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step

1000
900
800
Number of investors
700
600
500
400
300
200
100
0
0 250 500 750 1000 1250 1500 1750
Time step

investors who randomly change investment strategy, there is a corresponding increase

in the number of fundamentalists.19 Figure 16 shows the transitions of the number
of investors where the rate of investors who randomly change investment strategy is
3%. This shows that the number of fundamentalists increases even further, exceeding
the number of investors who conduct passive investment strategy.20 In real markets,
the eectiveness of passive investment strategy has widely been recognized from the
viewpoint of practical business aairs as well as the academic standpoint. However,
when the entire market is focused on, the rate of investors that adopt passive invest-
ment strategy is not always high. These results suggest that the existence of various
sources of excess earnings should be included as one of the factors for the low ratio
of passive investors. These are interesting results from both business and academic
viewpoints.21
4. Summary
Using analyses of agent-based model, this research looked at the influence of the dierence
in the forecast accuracy of fundamental values on financial markets. As a result of this
computer-based market analysis, the following findings were made: 1 improvements in
the forecast accuracy of fundamentalists can contribute to an increase in the number of
fundamentalists; 2 certain situations might occur, according to the level of the forecast
accuracy of fundamentalists, in which fundamentalists and passive management coexist,
or in which fundamentalists die out of the market, and furthermore; 3 where a variety of
investors exist in the market, improvements in forecast accuracy could increase the number of
fundamentalists more than the number of investors that conduct passive investment strategy.
These results contribute to clarifying the mechanism of price fluctuations in financial markets
and also indicate one of the factors for the low ratio of passive investors in real financial
markets. At the same time, they indicate that agent-based modeling is eective in conducting
analyses in the field of financial studies. The results obtained in this analysis have significant
meaning from both an academic and a practical business viewpoint. A more detailed analysis
that considers the actual investment environment should be included in future research.
5. List of Parameters
This section lists the major parameters of the financial market designed for this paper. The
explanation and value for each parameter is described.
Parameters Abbreviations
M: Number of investors 1000
N: Number of shares 1000
Fti : Total asset value of investor i for term t F0i 2000: common
Wt : Ratio of stock in benchmark for term t W0 0.5
wti : Stock investment rate of investor i for term t w0i 0.5: common
yt : Profits generated during term t y0 0.5
y : Standard deviation of profit fluctuation 0.2/ 200
: Discount rate for stock 0.1/200
: Degree of investor risk aversion 1.25
n : Standard deviation of dispersion from short-term expected rate of return on shares
0.010.03
a: Degree of selection pressure 10
c: Adjustment coecient 0.01
rtim : Expected rate of share return as estimated from risk etc.
ts : Assessed value of standard deviation of share fluctuation
th : Historical volatility of shares
Pt : Transaction prices for term t
f,i
Pt : Forecast value of transaction prices of investor i for term t
f,i
yt : Forecast value of profits of investor i for term t
r f,i : Short-term expected rate of return on shares of investor i
at : Price trend on stock until term t
ricum : Cumulative excess return of investor i for the most recent five terms
pi : Probability that investorsf who alter their strategy will adopt investor ifs strategy.
Endnotes
1. CAPM is also applied frequently to evaluate the enterprise value in Mergers and
Acquisitions M&A 23.
2. Passive investment strategy has been well-known in the actual asset management
businesses. On the other hand, active investment strategy that tries to obtain excess
earnings using investments has been widely prevalent. There also exist investment trust
funds that look for their basis of conducting active management in behavioral finance.
3. In the case of a financial market, investors represent agents and a stock market represents
a multiagent system 17, 24.
4. Buy-and-hold method is an investment method to hold shares for medium to long term.
5. This analysis covered major types of investor behavior as the analysis object 9.
6. The value of objective function fwti depends on the investment ratio wti . The investor
decision-making model here is based on the Black/Litterman model that is used in the
practice of securities investment 19, 25.
7. In the actual market, evaluation tends to be conducted according to baseline profit and
loss.
8. Selection pressures on an investment strategy become higher as the coecients value
increases.
9. This is one of the characteristics of agent-based modeling where such an analysis can be
conducted.
10. Where the forecast accuracy n is 0%, there are no forecast errors by fundamentalists.
11. In other words, these results show that the rate of increase in the number of
fundamentalists is influenced by the forecast dispersion of other investors.
12. A detailed analysis of the forecast accuracy of fundamentalists and the forecast
dispersion of other investors needs to be carried out in the future.
13. See Takahashi et al. 6 for a detailed analysis of the influence of passive investment
strategy on stock markets.
14. Under the present conditions, where the estimation accuracy of fundamentalists n is
1% and 2%, all the investors conduct passive investment strategy in either case. In this
sense, under the present conditions, the estimation accuracy of fundamentalists n does
not have any impact on whether or not fundamentalists can survive in the market.
15. Such a mechanism that works can make it possible for investors other than fundamen-
talists to always exist in the market.
16. In this case, transaction prices are consistent with fundamental values.
17. In the case under discussion here, market transactions consist of funds transferred
between fundamentalists and other kinds of investors. The existence of various investors
serves to provide the source of excess earnings for fundamentalists. These transactions
serve to determine transaction prices and therefore conform to fundamental values.
18. The increase in the rate of investors who randomly change investment strategy means
that there are more investors whose investments are based on trends and past averages
in the market.
19. In this case, transaction prices are consistent with fundamental values.
20. This paper analyzes the relationship between microbehavior and macrobehavior under
conditions where market prices are consistent with fundamental values. Analyzing the
market under other conditions, such as when fundamentalists are eliminated from the
market, will form part of our future work 17.
21. These results provide a significant suggestion with regard to the meaning of conducting
active investment strategy.
References
1 H. Markowitz, Portfolio Selection, Journal of Finance, vol. 7, pp. 7791, 1952.
2 W. F. Sharpe, Capital asset prices: a theory of market equilibrium under condition of risk, The Journal
of Finance, vol. 19, pp. 425442, 1964.
3 F. Black and M. Scholes, Pricing of options and corporate liabilities, Bell Journal of Economics and
Management Science, vol. 4, pp. 141183, 1973.
4 J. E. Ingersoll, Theory of Financial Decision Making, Rowman & Littlefield, Lanham, Md, USA, 1987.
5 E. Fama, Ecient capital markets: a review of theory and empirical work, Journal of Finance, vol. 25,
pp. 383417, 1970.
6 H. Takahashi, S. Takahashi, and T. Terano, Agent-based modeling to investigate the eects of
passive investment strategies in financial markets, in Social Simulation Technologies: Advances and
New Discoveries, B. Edmonds, C. Hernandes, and K. Troitzsch, Eds., pp. 224238, Idea Group, 2007,
Representing the best of the European Social Simulation Association conferences.
7 D. Kahneman and A. Tversky, Prospect theory of decisions under risk, Econometrica, vol. 47, pp.
263291, 1979.
9 A. Shleifer, Inecient Markets, Oxford University Press, Oxford, UK, 2000.
10 R. J. Shiller, Irrational Exuberance, Princeton University Press, Princeton, NJ, USA, 2000.
11 R. Axtell, Why Agents? On the Varied Motivation for Agent Computing in the Social Sciences, Working
Paper no.17, Brookings Institution Center on Social and Economic Dynamics, 2000.
12 S. J. Russel and P. Norvig, Artificial Intelligence, Prentice-Hall, Upper Saddle River, NJ, USA, 1995.
13 J. M. Epstein and R. Axtell, Growing Artificial Societies Social Science from the Bottom Up, MIT Press,
Cambridge, Mass, USA, 1996.
14 W. B. Arthur, J. H. Holland, B. LeBaron, R. G. Palmer, and P. Taylor, Asset pricing under endogenous
expectations in an artificial stock market, in The Economy as an Evolving Complex System II, pp. 1544,
Addison-Wesley, Reading, Mass, USA, 1997.
15 R. Axelrod, The Complexity of Cooperation -Agent-Based Model of Competition and Collaboration, Princeton
University Press, Upper Saddle River, NJ, USA, 1997.
16 L. Tesfatsion, Agent-Based Computational Economics, Economics Working Paper, No.1, Iowa Sate
University, Iowa, Miss, USA, 2002.
17 H. Takahashi and T. Terano, Agent-based approach to investors behavior and asset price fluctuation
in financial markets, Journal of Artificial Societies and Social Simulation, vol. 6, no. 3, 2003.
18 P. OBrien, Analysts forecasts as earnings expectations, Journal of Accounting and Economics, pp.
5383, 1988.
19 F. Black and R. Litterman, Global portfolio optimization, Financial Analysts Journal, pp. 2843, 1992.
20 W. F. Sharpe, Integrated asset allocation, Financial Analysts Journal, pp. 2532, 1987.
21 D. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-Wesley,
Reading, Mass, USA, 1989.
22 H. Takahashi and T. Terano, Analysis of micro-macro structure of financial markets via agent-based
model: risk management and dynamics of asset pricing, Electronics and Communications in Japan, Part
II, vol. 87, no. 7, pp. 3848, 2004.
23 R. Brealey, S. Myers, and F. Allen, Principles of Corporate Finance, McGraw-Hill, New York, NY, USA,
8th edition, 2006.
24 M. Levy, H. Levy, and S. Solomon, Microscopic Simulation of Financial Markets, Academic Press, Boston,
Mass, USA, 2000.
25 L. Martellini and V. Ziemann, Extending black-litterman analysis beyond the mean-variance
framework: an application to hedge fund style active allocation decisions, Journal of Portfolio
Management, vol. 33, no. 4, pp. 3345, 2007.

835618

Uploaded by

Document Informationclick to expand document information-

Document Informationclick to expand document information

Copyright:

Available Formats

835618

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

835618

Uploaded by

Copyright:

Available Formats

Journal of Probability and Statistics

Actuarial and Financial Risks: Models,

Guest Editors: Ricardas Zitikis, Edward Furman, Abdelhakim Necir,

Actuarial and Financial Risks: Models,

The Foundations of Probability with Black Swans, Graciela Chichilnisky

Asteroids: Assessing Catastrophic Risks, Graciela Chichilnisky and Peter Eisenberger

Continuous Time Portfolio Selection under Conditional Capital at Risk,

On Some Layer-Based Risk Measures with Applications to Exponential Dispersion Models,

Pricing Equity-Indexed Annuities under Stochastic Interest Rates Using Copulas,

Estimating L-Functionals for Heavy-Tailed Distributions and Application, Abdelhakim Necir

POT-Based Estimation of the Renewal Function of Interoccurrence Times of Heavy-Tailed

An Analysis of the Influence of Fundamental Values Estimation Accuracy on Financial

Ricardas Zitikis,1 Edward Furman,2 Abdelhakim Necir,3

Correspondence should be addressed to Ricardas Zitikis, [email protected]

Received 10 June 2010; Accepted 10 June 2010

W. John Braun,1 Bruce L. Jones,1 Jonathan S. W. Lee,1

Correspondence should be addressed to W. John Braun, [email protected]

Received 2 October 2009; Revised 6 March 2010; Accepted 27 April 2010

Academic Editor: Ricardas Zitikis

Figure 1: Location of the District of Muskoka within the Province of Ontario.

2. The Data and Study Area

2.2. Description of Historic Fire Data

3. Ignition and Burn Probability Modelling Using Generalized

Figure 4: Map of human ignition locations in study area 19802007.

Figure 5: Map of lightning ignition locations in study area 19802007.

3.2. Simple Burn Probability Map

0.000050.00043 0.002340.00271 0.004230.0046

0.000440.00081 0.002720.00309 0.004610.00498

0.00120.00157 0.003470.00384 0.005370.00574

4. Burn Probability Modelling Using a Fire Growth Model

00.00043 0.002560.00298 0.004690.0051

0.000440.00085 0.002990.0034 0.005110.00553

0.001290.0017 0.003840.00425 0.005960.00638

4.2. Canadian Fire Behaviour Prediction System and Fire

4.3. Burn-P3 Simulation Model

C-1 C-6 S-1

C-2 D-1 Water

Figure 10: Corrected vector fuel map.

4.4.2. Verification in the Field

Table 1: Fire Behavior Prediction System fuel types 10.

Group/Identifier Descriptive name

4.4.3. Unmapped Private Properties

Figure 12: Map of the 20 randomly selected roads to be sampled.

Name of Road Number of observations Locations along road km

Table 3: Summary of fuel map accuracy sampling procedure.

Tolerance Margin % Misclassified

Figure 13: A structure embedded in a region of continuous fuel.

4.4.4. Fuel Breaks

C-1 C-6 S-1

C-2 D-1 Water

Figure 14: Rasterized fuel map at 25 m resolution.

4.4.6. Historic Weather

4.4.7. Seasons and Causes

4.4.8. Ignition Grids

4.4.9. Estimation of Spread Event Days

Human caused fires

5. Analysis and Results

Residential Railway Recreational

Table 1: Fire Behavior Prediction System fuel types 10.

Proof. This is immediate from the definitions of both 3, 12.

Proof. It follows from Propositions 2.1 and 3.2 9, 12.

y y a.e. on Sc R, x x a.e. on Sc R, where S < . 6.3

For a proof see the theorem in Chichilnisky 6.

For a proof see 4.