Eri WP Predicting Decomposing Risk Data Driven Portfolios 0 PDF
Eri WP Predicting Decomposing Risk Data Driven Portfolios 0 PDF
Predicting and
Decomposing the Risk
of Data-Driven Portfolios
February 2020
————————
1. Introduction
5
3. Implementation
15
4. The unreliability of the
square-root-of-time rule
18
5. Empirical results
22
6. Concluding remarks
30
Appendix
32
References
36
›2 ›2 EDHEC 2020.
Printed in France, February 2020. Copyright
The opinions expressed in this study are those of the authors and do not necessarily reflect those of EDHEC Business School.
Abstract
————————
Sophisticated algorithmic techniques are complementing human judgement across the fund industry. Whatever
the type of rebalancing that occurs in the course of a longer horizon, it probably violates the buy-and-hold
assumption. In this article, we develop the methodology to predict, dissect and interpret the h-day financial risk
in data-driven portfolios. Our risk budgeting approach is based on a flexible risk factor model that accommodates
the dynamics in portfolio composition directly within the risk factors. Once these factors are defined, we
cast portfolio risk measures, such as value-at-risk, into an additive mean-variance-skewness-kurtosis format.
The simulation study confirms the gains in accuracy compared to the widespread square-root-of-time rule. Our
main empirical findings rely on the two-decade performance of a portfolio insurance investment
strategy. Rather than looking at total portfolio risk, we conclude that it is more informative to look inside
the portfolio.
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›3
ABOUT THE AUTHORS
————————
Nabil Bouamara is a PhD candidate in Applied Economics at Katholieke Universiteit Leuven and
Vrije Universiteit Brussel (Belgium). He is a fellow of the Flemish Research Foundation (FWO).
His research revolves around the use of econometric methods to assess the risk profile of highly
dynamic portfolios. Specifically, he investigates the risk of algorithmic portfolios, sluggish prices
in financial time series and the real-time incorporation of news. In 2019, he has been awarded
the Gustave Boël - Sofina scientific prize for a one-year research stay at the Aix-Marseille School
of Economics.
Kris Boudt is professor of finance and econometrics at Ghent University, Vrije Universiteit Brussel
and Vrije Universiteit Amsterdam. He teaches the online courses "Introduction to portfolio
analysis in R" and "GARCH models in R" at DataCamp. Kris Boudt obtained his PhD in 2008 for his
developments in the modelling and estimation of financial risk under non-normal distribution.
He has published his research in the Journal of Econometrics, Journal of Portfolio Management,
Journal of Financial Econometrics, and the Review of Finance, among others. Kris Boudt received
several awards for outstanding research and refereeing and is an active contributor to the open
source community.
Jürgen Vandenbroucke PhD is expert general manager and head of innovation at KBC Asset
Management, Belgium. He is guest lecturer at University of Antwerp on Financial Engineering,
guest lecturer at Ehsal Management School Brussels on Financial Securities and research associate
at Edhec-Risk Institute in the program “technology, big data and artificial intelligence for
investment solutions”. He has published in both academic and practitioners-oriented journals
such as Journal of Behavioral Finance, European Journal of Law and Economics, Journal of
Investment Management, Journal of Asset Management, Journal of Wealth Management and
Journal of Investing.
Acknowledgements:
We are grateful to KBC Asset Management’s R2 team for generously sharing their expertise on portfolio insurance strategies. We have received helpful comments and suggestions from
David Ardia, Leopoldo Catania, Serge Darolles, Hans Degryse, Jean-Yves Gnabo, Geert Huyghe, Olivier Scaillet, Piet Sercu, James Thewissen, Tim Vanvaerenbergh, Jens Verbrugge as well
as the participants at the Mathematical and Statistical Methods for Actuarial Sciences and Finance Conference (Madrid, 2018), the European Actuarial Journal Conference (Leuven,
2018), the Conference in International Macroeconomics and Financial Econometrics (Nanterre, 2019), the KU Leuven seminars, the Quantitative Finance and Financial Econometrics
Conference (Marseille, 2019) and the Society of Financial Econometrics Summer School (Brussels, 2018; Evanston, 2019). This work was supported by the Research Foundation –
Flanders [PhD fellowship 11F8419N].
›4 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
————————
1. Introduction
————————
One of the fastest growing innovations in the asset propose a single option-like factor that captures the
management industry uses algorithms and machine- characteristics of the entire family of trend-following
learning techniques to find profitable patterns in hedge funds. Our approach is distinctly different in that
market data.1 A tension arises linked to the feature that we incorporate the specific aspects of the investment
the portfolio composition tends to change frequently strategy directly into the assets. In doing so, we offer
across the risk evaluation horizon, naturally suggesting a multi-period portfolio analysis tool for individual
the use of risk signalling tools on each single trade. On investors that relates each observable portfolio
the other hand, it is more realistic to rely on aggregated component (e.g. Dow 30 constituents) directly to a
information on lower-frequency intervals. This turns manageable set of risk factors. Second, we contribute to
out to be even more problematic because the classical the interpretability of risk measure. Much of the recent
approaches to understand risk over a longer-horizon literature has focused the sensitivity of tail risk to
are predicated on the buy-and-hold assumption, or portfolio moments Stoyanov et al. (2013a,b), dynamic
equivalently, that the timing of risk evaluation and portfolio risk (Francq and Zakoïan, 2018), dynamic asset
portfolio allocation coincide. allocation models (Basstürk et al., 2019) and portfolio
optimization with non-Gaussian objectives (see e.g.
In this article, we focus on the insights afforded by Boudt et al., 2020a,b). The idea that active rebalancing
Litterman’s (1996) “hot spots” – a traditional tool to does not merely result in the shift of the marginal
reveal how much of the total portfolio risk can be risk impacts but completely transforms the structural
attributed to underlying building blocks, such as relationship between individual companies, requires us
sectors, asset classes, geographical regions, etc. Our to look inside the portfolio. In particular, we offer a
main methodological contribution is that we account complementary picture of risk in a portfolio, treating
for the mismatch between the risk evaluation horizon portfolio risk as the result of different combinations
and the frequency of investment decisions. We do so by of factors and their interactions.
accommodating the random nature of the intra-
window composition directly within the risk factors, In the empirical application, we investigate the
allowing for a direct description of non-buy-and- expected shortfall of active positions in the 30 Dow
hold portfolios. We impose some structure on the Jones Industrial Average constituents. Specifically, we
time-varying position vector by defining it within a examine a Constant-Proportion Portfolio Insurance
flexible class of data-driven investment strategies, (CPPI) investment algorithm, similar to the CPPI
which includes any investment that is mechanically products sold to European retail investors.2 Previous
driven by past price movements. Furthermore, to research by Ardia et al. (2016) has concluded that, due
dissect and interpret financial risk, we consider to the non-linear and multi-horizon nature of this
mean-variance-skewness-kurtosis (MVSK) portfolio investment, its path dependence and the asymmetric
risk measures, i.e. closed form expressions of the impact of volatility, it is not immediately obvious how
exposure vector and the portfolio co-moments (i.e., the risk characteristics of the underlying components
co-variance, co-skewness and co-kurtosis). An intuitive, affect the performance of the investment. In response,
two-dimensional companion table is meant to illustrate we find that our portfolio analysis tool overcomes
how, vertically, the risks of the factors add up to create these problems and allows for a better judgement as
the risk of the portfolio and, horizontally, portfolio to how the shape of the left tail is affected through
and factor risk can be decomposed into Gaussian and each component.
non-Gaussian attributes.
The remainder of this article is structured as follows.
We contribute to the literature in two ways. First, we Section 2 sets out a general framework for the
introduce risk factors for active investment strategies. decomposition of MVSK portfolio risk, which nests any
A similar motivation for the use of strategy-based data-driven portfolio strategy and investment horizon.
returns can be found in Fung and Hsieh (2001), who Section 3 is concerned with practical guidelines and
1 - See for example, CFA institute, 2019; AI pioneers in investment management; Economist, October 2019; The stockmarket is now run by computers, algorithms and passive
managers; Financial Times, October 2019, Meet The Buffet bot: quant funds try to crack the value code; Financial Times, July 2019, Quants seek human touch in reboot of investing
strategy; Financial Times, October 2019, Why hedge fund managers are happy to let the machines take over; Financial Times, October 2019, Will bots replace humans in active
equity investment.
2 - Examples include Candriam Dynamic Global, KBC Privileged Portfolio Pro (95, 90 and 85), BNP Paribas B Control, DWS Fund Global Protect (80 and 90), Swiss Life Multi Asset
Protected Fund 80 and Fideuram Equity Smart Beta Dynamic Pro 80.
›6 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
sketches a simple step-by-step procedure for risk
budgeting a portfolio. The simulated results in Section
4 confirm our intuition that the naive square-root-of-
time rule is inappropriate for dynamic portfolios.
Section 5 discusses the empirical results of MVSK
risk measures to gauge the risk of active Dow 30
portfolios over the period 1987 to 2009. We conclude in
Section 6.
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›7
————————
2. Methodology for
long-horizon risk
decomposition
————————
Figure 1 sets out the main idea. The investor has the formally, the intra-window weights are defined by
responsibility to gauge ex-ante risk over the dash (2)
multi-period horizon, while being aware that the
investment rule may actively rebalance the portfolio in which the u(·) function is an updating equation. The
at a set of potential rebalancing points, indicated by static parameter vector defines the characteristics
t = 0, 1, 2, …, h. of the investment strategy and thus defines the type
of updating used for wt . The recursive nature of the
This section introduces the methodology for formulation implies that is path
long-horizon risk decomposition in all its generality. We -dependent – it is a function of the multivariate
begin in Section 2.1 by making the explicit distinction history of p-dimensional of asset returns ,
between longer-horizon risk evaluation and weights and possibly other exogenous variables.
shorterhorizon portfolio rebalancing intervals. We
also propose a model that will be central to disentangle Notable examples that fit into this data-driven
longer-horizon portfolio risk across factors. In Section 2.2 framework include momentum or relative-strength
we introduce our MVSK risk measure. In Section 2.3, we strategies that buy past winners and sell past
map the risk measure into factors and MVSK contributions. losers, contrarian strategies that exploit reversal
Finally, in Section 2.4, we include an illustration trades, fundamental value portfolios, the family of
of the MVSK decomposition using a stylistic risk-based portfolios (inverse volatility, minimum
example. variance, maximum diversification, equally weighted,
etc.), technical trading rules, so-called “smart beta”
strategies, portfolio insurance techniques, and the more
2.1. Notation recent techniques leveraging articifical intelligence
Let ri,t be the intra-window log-returns of the i–th criteria (see e.g. De Prado, 2018, for a broad treatment).
asset at time t, respectively for i = 1, …, p and t = 1, …,
h. We normalize the horizon h to be a risk evaluation The return across the risk evaluation window, or the
interval. The portfolio log-returns yt(wt-1) are equal h-period portfolio log-return, is the sum of one-step
to log-returns
, (3)
for 1 ≤ t ≤ h and i = 1, 2,…, p: (1)
Now we turn to the linear connection between the
The time-subscript in the component weights longer-horizon portfolio return yh and the underlying
highlights the portfolio is rebalanced using past building blocks, which is central in the subsequent
information. equations. A convenient way to describe the multi-
period portfolio return of an active p-asset portfolio
We impose that the investment portfolio log-returns is a factor model
y t( ) adhere to a data-driven system. More (4)
Note: We visualize the low-frequency risk evaluation interval vs. higher-frequency portfolio decision interval. The investor gauges risk over the h-period horizon
(top panel), while being aware that the investment rule may actively rebalance the portfolio at a set of potential rebalancing points, indicated by t = 0, 1, 2,…, h
(bottom panel).
›9 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
in which a is the intercept, is the in calculating left-tail risk has been well-studied in the
exposure vector consisting of factor loadings and literature. For example, Zoia et al. (2018) make use of
is the q-dimensional factor variable. Gram-Charlier expansions to calculate value-at-risk
The benefit of this structure is that it is now possible to and expected shortfall. We refer to Appendix B for
decompose portfolio risk for horizons longer than a comprehensive definition of these risk measures.
the frequency at which portfolios are rebalanced. As
we will show later in this paper, additional granularity is
possible by casting portfolio returns in a meanvariance- 2.3. Two-dimensional mean-variance-
skewness-kurtosis format. Therefore, is assumed to skewness-kurtosis risk attribution
have finite fourth-order moments. We focus on the part of the overall risk that should
be the object of concern in risk management or asset
Several approaches exist to select the risk factors allocation, along the lines of Litterman (1996, p. 92),
. In financial applications, the assumption of i.e. “the practice of risk managers needs to reflect that
macroeconomic, fundamental and statistical models what matters is the marginal impact on the risk of
in multi-factor models is common (see, for example, the portfolio, not the risk of the individual securities.”
Connor, 1995). Without loss of generality, we use Since the portfolio return is defined as a linear function
synthetic risk factors that mimic and accommodate the of factors (see Eq. 4), it follows that the portfolio
portfolio fluctuations of the corresponding investment risk measure is a one-homogeneous function of the
strategy across a longer horizon. We refer to Section portfolio exposure vector . Risk factor decomposition
3.1 for further elaboration. is then straightforward via Euler’s Homogeneous
Function Theorem (see Gourieroux et al., 2000; Scaillet,
2004, for examples).
2.2. Portfolio risk
To dissect and interpret financial risk, we require a The Euler Theorem for dynamic portfolios is formulated
multivariate risk function r that can be expressed as in the following equation. For portfolio returns yh
a closed form of the exposure vector and the described by exposure vector a and the q-dimensional
co-moments of the risk factors (i.e. the mean vector risk factor set , we have a top-down attribution of
, the co-variance matrix , the co-skewness matrix total portfolio risk to the part that can be attributed
and the excess3 co-kurtosis matrix ) to risk factors
(5) (6)
in which is a one-homogeneous
function of a.4 We refer to Appendix A for the necessary in which is the partial
formulas to calculate the co-moments. derivative of the portfolio risk with respect to the
exposure to the risk factor j. The contribution of
Examples of risk functions nested in Eq. (5) include factor j over the same risk evaluation window is
the portfolio standard deviation, portfolio value-at- equal to . Similarly,
risk and portfolio expected shortfall. Non-normal the percentage contribution is defined a
features in financial returns can be accounted for . We use the dots
via Cornish-Fisher value-at-risk (Zangari, 1996) and shorthand to save space.
Edgeworth expected shortfall (Boudt et al., 2008). The
former relies on Cornish-Fisher expansions around the Aside from a decomposition into factors, the
standard Gaussian quantile funtion and the latter on ability to decompose the portfolio risk into its
Edgeworth expansions around the standard Gaussian Gaussian and non-Gaussian characteristics is also
distribution function. The use of asymptotic expansions useful for explaining the sources of risk. In this
3 - We denote * as the co-kurtosis matrix. The matrix is the co-kurtosis matrix under multivariate normality. For example, for dimensions equal to 2, we have that the
bivariate normal co-skewness matrix N and co-kurtosis matrix are equal to,
(26)
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›10
vein Andreas Steiner proposes in a white paper to zero matrix, the subscript of cmr, with moment
disentangle portfolio modified value-at-risk into . Each subequation corresponds to the
additive mean, standard deviation, skewness and mean, volatility, skewness and kurtosis contributions to
kurtosis contributions5 total risk, respectively. For example, the generalization
naturally extends to the portfolio modified expected
shortfall originally proposed by Boudt et al. (2008),
which involves complex products of density
(7a) functions.
with, (7b)
(7c) Finally, we wish to see how each of the factors affects
(7d) the portfolio-level moment budgets in Eqs. (8a –8e).
We simply apply Euler’s Theorem on each of the
(7e) one-homogeneous subequations.
(7e)
(7f) Table 1 depicts the concept of two-dimensional risk
attribution to factors and moments. The elements
in which, za is the standard Gaussian quantile, m is ; for j = 1,…, q
the portfolio mean, s is the portfolio volatility, s is and (9)
the portfolio skewness and k is the portfolio kurtosis.
Each subequation corresponds to the mean, volatility, represent the contribution of factor j to the conditional
skewness and kurtosis contributions to total modified moment contribution cmr (viz., Eqs. (8b) through (8e)).
value-at-risk, respectively. The partial derivative is the
moment contribution of the partial derivative with
We find that similar results hold for many other risk respect to the exposure to a particular risk factor.
functions of the form defined in Eq. (5). More precisely, Note that
we propose a generalized decomposition in which
portfolio risk is mapped into a mean-varianceskewness-
kurtosis framework by calculating the marginal
effect of each moment contribution in which .x. is a It can be seen that the companion table reflects three
levels of diagnostic information: portfolio, factor and
(8a) moment. In the bottom-right, we report the total
risk of a portfolio. Vertically, total risk is decomposed
into the risk contribution of each risk factor position.
Horizontally, total risk and each factor contribution are
(8b)
disentangled into the effects of traditional moments.
By construction, these contributions add up to total
portfolio risk.
(8c)
›11 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
Table 1: Mean-variance-skewness-kurtosis risk budgeting table
Moment Contributions
Factor Mean St.Dev. Skewness Kurtosis Total
Factor1 r1
Factor2 r2
Factor3 r3
Factorj rj
Factorq rq
Total rµ rσ rs rk r
Note: We propose a tabular form for the decomposition of mean-variance-skewness-kurtosis risk r := rµ + rσ + rs + rk, itself made up of factor contributions
defined , with j = 1,…, q and . We make use of a simple rectangle tabular structure to illustrate how, vertically, the risks of the components
add up to create the risk of the portfolio and, horizontally, portfolio and component risk can be decomposed into normal and non-normal attributes.
characteristics (i.e. mean, volatility, skewness and different factor contributions. In fact, every expected
kurtosis). We make sure that these portfolios have an shortfall can be endlessly expressed as the sum of four
identical 5% portfolio modified expected shortfall traditional portfolio moment contributions, themselves
(mES) equal to 1.459 by rescaling the volatility (see the made up of additive factor contributions (see Eq.(9)).
location-scale representation in Eq. (23)). In addition, the multivariate interactions among
the portfolios factors are equally endless. After all,
We call these stylized examples “seemingly iso-risk different entries in the co-variance, co-skewness or
portfolios” to emphasize that, while these portfolios co-kurtosis matrices can lead to similar levels of total
have exactly the same aggregate portfolio risk, their portfolio risk.
inherent risk profiles are very different. This can also
be observed visually through the markedly different For simplicity of exposition, we consider a bivariate
formations of the left tail. For example, in investment portfolio with independent factor components. For
Strategy B, large negative observations are more likely instance, the co-moments that were used to obtain
to appear than large positive realizations (s = —0.500). Investment strategy D is equal to,
In addition, the return distribution has a tail that may
not vanish as quickly as those expected from a normal
distribution (k = 1.000).
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›12
Figure 2: Mean-variance-skewness-kurtosis factor attribution for seemingly iso-risk portfolios
(a) Iso-risk in different portfolio return density functions
›13 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
The other investment strategies are characterized
by different combinations of skewness and excess
kurtosis. For example, the univariate return distribution
of Strategy D has a portfolio return distribution that
is described by positive skewness (s = 0.500) and
positive excess kurtosis (k = 1.000). Correspondingly,
in the bottom panel, we observe that portfolio risk
is affected by a negative skewness contribution
(rs =0.235), lowering portfolio risk, and a positive
kurtosis contribution (rk = 0.021), increasing portfolio
risk.
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›14
————————
3. Implementation
————————
The implementation of the proposed risk budgeting Both portfolio returns and style factor returns are
methodology requires to estimate the exposure vector compounded over the risk evaluation interval for
( ), multivariate moments ( , , , ), and portfolio each simulation path, leading to
risk (r). We propose the Simulation-based Multiperiod and with being a sample
Portfolio Attribution (SiMPA) algorithm, which is a of p-dimensional simulated observations of the
three-step procedure for risk budgeting a class of cumulative factor returns. For example, in the empirical
investment strategies that may trade within the risk application, we use the cumulative simulated h-period
prediction horizon. returns of a dynamic strategy that invests in each
subperiod in each separate constituent of the Dow
30 index.
3.1. Construction of the data matrices
In this subsection we elaborate on the practical Since we repeat this procedure for n simulation paths,
considerations of associating the weight updating it follows by construction that these p prespecified
equation defined in Eq. (2) proportionally with the synthetic factors combine the information on out-of-
underlying components across a longer-horizon risk sample returns and daily changing weights.
evaluation interval.
We assume to have a fully specified simulation model 3.2. Estimation of factor model
for the daily asset returns. An example of such a model We maintain a linear relationship between multi-
is given in Appendix C. Denote the simulated daily period portfolio return and our pre-specified synthetic
data by with a sample factors (see Eq. (4)). However, the exposure vector is
of p-dimensional observations of the asset log-return typically unknown and has to be estimated. In low
variable at time t. dimensions, the exposure vector is easily obtained
using an ordinary least squares regression. Since the
The portfolio composition is path-dependent, meaning number of potential risk factors is large and/or some of
that it is conditional on the multivariate lagged series the risk factors are highly linearly dependent (through,
of assed returns and lagged asset weights (see Eq. for example, a shared weight updating equation), it is
(2)). For each time index t in the prediction horizon, important to regularize the influence of the factors.
we apply a data-driven investment rule (described by We recommend to use the elastic net regularization of
the updating equation) on asset return draws. After Zou and Hastie (2005) to uniquely identify the vector
applying Eq. (1), this leads to rebalanced daily portfolio of factor loadings.
returns described by a set of intra-
window active weights , The data matrix is the standardized version of
in which is a sample of p-dimensional the risk factor data matrix , with components
observations of component weights. , where avi and stdi are
the sample mean and standard deviation of
In order to construct style risk factors, we correct . The standardization is crucial in
the simulated asset returns for the intra-window penalized regressions, since the penalty depends on
composition fluctuations via pairwise multiplication the scale of the components.
between component weights and simple asset
returns , leading to , To obtain the vector of slope coefficients, we minimize
with a sample of p-dimensional observations an objective function of the following form
of the intra-evaluation-window factor returns. In other
words, we incorporate the weight signal directly and
proportionally in each of the asset returns. (10)
›16 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
in which ||·||v is the norm and l ≥ 0 sets the level
of regularization.6
6 - The implementation of the elastic net regularization in Eq. (10) requires the calibration of the penalty parameters l and g. We follow the idea of Zou et al. [2007] and minimize
a BIC-like criterion (where BIC stands for Bayesian Information Criterion) over a grid of (l, g) candidate pairs. The standard deviation d of the unexplained variation is calculated
in accordance with Reid et al. [2016].
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›17
————————
4. The unreliability of the
square-root-of-time rule
————————
In Figure 3 we use a simulated example to show are the log asset prices recorded at time t. We
the improvements of our SiMPA approach over the assume a fully-invested portfolio. All processes, with
traditional square-root-of-time rule in the presence the exception of the passive buy-and-hold strategy,
of dynamic weights. are based solely on the time index t, and not the
previous price path. The constant(-mix), step and
Consider a monthly risk evaluation horizon with out- ramp functions are chosen because of rapid changes,
of-sample rebalancing points t = 1, 2,…, 20 and let gradual changes and periods of constancy across the
asset i = (1, 2) make up a two-asset portfolio with risk evaluation interval.
asset returns drawn from a multivariate normal
distribution. We consider a homogeneous asset mix in It is interesting to note the apparent resemblance
which we posit similar risk contributions, defined by a of these patterns to actual practice in the financial
null mean vector and volatilities (2.00, 2.00)'. Means industry. The constant-mix process is conceptually
and volatilities are expressed in percentages. For consistent with a balanced equally weighted portfolio.
the sake of simplicity, the multivariate dependence This strategy aims at diversification and, in doing so,
parameter is set to zero. The portfolios start off with downweighs winners and overweighs losers. The ramp
an equal allocation, wi,t=0 = (0.5, 0.5)'. profile mirrors a market timing strategy, which builds
up capital steadily in a particular asset and disinvests
The standard workhorse to assess multi-horizon risk after a distressing market signal. The step function
is to scale risk estimates using the square-root-of- appears like the flight-to-safety phenomenon, in
time rule. This approach embodies the buy-and- which funds are suddenly shifted from an equally
hold-assumption, which is implemented analytically weighted portfolio to complete concentration in a
by means of the true mean vector, the covariance safe haven asset.
matrix and the weights observed at the beginning of
the risk evaluation horizon, In the utmost left column we report analytical results
based on initial weights consistent with the buy-and-
hold assumption. Absent any intra-window weight
(11) changes, the risk budgets behave as expected
from their static portfolio composition and return
The key point that we stress in this paper is that it characteristics. More precisely, in this setting, risk
is indispensable to sort out the effects of dynamic scaling characterizes a portfolio in which each
weight adjustments when identifying the sources of component contributes equally to total portfolio risk.
multi-period risk. While this problem was pinpointed In columns (2) through (5), we apply the proposed
in Diebold et al. (1998), it is still unresolved in the SiMPA approach and use deterministic weight
context of risk allocation. In light of this problem, we processes to replicate a dynamic investment strategy.
consider the implications of four stylized weighting SiMPA accounts for both the randomness in returns
profiles on the evaluation of risk using the proposed and randomness in weights and, thus, describes
SiMPA approach. The deterministic weight processes effective risk contributions.
are defined as
buy-and-hold: w1,t = p1,t /(p1,t + p2,t) We inspect the allocation of the exposure vector
with p1,0 = p2,0 (12a) and the percentage risk contributions to portfolio
constant(-mix): w1,t = 0.5, (12b) volatility. The slope coefficients reported in panels
ramp: w1,t = mod [(t +1)/4]; (12c) (1) and (2) load on the transformed risk factors
step: w1,t = 0.5 — 0.5(t > 2); (12d) (associated to the corresponding assets) with a slope
coefficient almost equal to 1. The residual risk factor
in which w1,t indicates the capital to be invested in is almost negligible. A similar multi-period allocation
asset 1 at the intra-window time t and pt = (p1,t , p2,t) description is not possible for the square-root-of-
›19 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
Figure 3: Inaccuracy of square-root-of-time using one-month risk attribution
(a) Intra-month weight evolution
time rule, since scaling does not have an obvious (Column 1 vs. Column 2) with sufficient accuracy
relation to the proposed active risk factor model. in terms of portfolio volatility and percentage
factor contributions. The most remarkable aspect is
The risk contributions in Panel 2 confirm the the misclassification that results from the square-
adequacy of the SiMPA approach. As expected, we root-of-time rule. In particular, we observe that
observe that the SiMPA estimation strategy is able more frequent and/or abrupt movements lead
to capture the analytical buy-and-hold assumption to a misleading bias in percentage component
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›20
contributions. The severest mismatch can be seen in
the step function (Column 5), in which percentage
contributions go from an expected 50/50 percentage
risk budget judgement to an approximate 16/84
realized attribution.
›21 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
————————
5. Empirical results
————————
5.1. Data and strategy description in which are the asset weights within the risky
We study a representative data set composed of reference portfolio, the exposure Et = max{min{pCt
the Dow Jones Industrial Average constituents. The ,Vt }, 0} is the nominal amount allocated to a risky
sample period is from March 17, 1987 to February portfolio of assets and Vt is the portfolio value, with
3, 2009 corresponding to 5,520 trading days and V0 =100. This exposure is defined by the multiplier
includes several episodes of exceptional risk (the coefficient p and a cushion Ct. The latter indicates
LTCM collapse, the September 11 attacks, the dot- the difference between the portfolio value Vt and the
com bubble, the financial crisis of 07-08, etc.).7 constant floor F = 90 up to the portfolio’s anniversary.
For the out-of-sample evaluations of a data-driven The exposure is bounded to Vt so that no leverage is
strategy in sections 5.2, we impose a minimum allowed. The floor level is yearly reset to avoid a cash
estimation window of length T = 240 days, starting lock-in.
from February 26, 1988. This sample window
corresponds to 1,056 five-day risk evaluation The risky reference pool of assets is typically an index.
windows, consisting of 5,280 potential rebalancing We consider three types of prior-return-weighted
points at which the investment rule can change the portfolios that are rebalanced end-of-month: an
portfolio composition. equally weighted index, a low-risk index (computed
via inverse volatility in the previous 240 trading days)
Table 2 describes the variation in return distributions and a price-weighted index. The application is chosen
and non-normality in the observed financial returns such that there are two types of dynamic weight
across weekly risk management horizons. We report adjustments within the risk evaluation horizon,
the averaged summary statistics by sectors for each namely, (i) a prior-return-weighted monthly
of the constituents of the Dow index. rebalancing in the reference indices and (ii) a
dynamic floor-protection which allocates funds to
The companies belong to the following sectors: the reference portfolio on a daily basis.
Energy (2), Financials (5), Industrials (6), Consumer
Cyclicals (4), Technology (3), Computers & Technology Table 3 describes the variation in return distributions
(1), Healthcare (3), Consumer Non-Cyclicals, Telecom and non-normality in the observed financial returns
(2) and Oil & Gas (1). The number of assets operating across weekly risk management horizons of the
under each sector label are reported in parentheses, prior-return-weighted indices (Panel 1) and the
and each ticker is reported in Column (2). corresponding CPPI-rebalanced portfolios (Panel 2).
The data-driven investment algorithm that we Unsurprisingly, for low values of the multiplier, the
investigate aims at floor protection by using a CPPI standard deviations (Column 4), minima (Column 7),
strategy. This means that the portfolio investment maxima (Column 8) and modified expected shortfall
rule dynamically allocates capital between a risk- (Column 9) are considerably lower compared to the
free fixed income asset (which provides protection assets and Dow 30 indices (cfr. Table 2). Note that
against market losses) and a risky reference pool of when the multiplier increases the CPPI portfolio
assets (with upside potential). We refer to Black and returns become increasingly similar to the equity
Perold (1992) for a detailed treatment of portfolio index returns. Hence, the choice of multiplier thus
insurance. determines the trade-off of a reward vs. risk.
The weights on the asset returns at rebalancing date The values for skewness (Column 5) and excess
t are defined as kurtosis (Column 6) are in all cases high enough to
(13) reject the p-value of the Jarque-Bera statistic for
normality at less than 1% (not tabulated). Because of
7 - The daily log-returns are available in the R package rmgarch [Ghalanos, 2019].
›23 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
Table 2: Full-sample summary statistics of weekly log-returns of Dow 30 constituents (grouped by sector)
Sector/Portfolio Tickers (%) (%) mini(%) maxi(%) mESi(%)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Table 3: Full-sample summary statistics of weekly log-returns of active portfolios on Dow 30 constituents
Portfolio (%) (%) mini(%) maxi(%) mESi(%)
(1) (2) (3) (4) (5) (6) (7) (8)
Panel 1: Prior-return-weighted Dow 30 indices
Equally weighted 0.254 2.592 —0.555 5.375 —17.448 13.071 8.060
Low-risk 0.251 2.161 —0.469 4.600 —12.637 12.080 6.307
Price-weighted 0.259 2.769 —0.328 3.652 —15.608 12.649 7.453
Panel 2: CPPI portfolios
2a: Reference index: Equally weighted index
CPPI (p = 2) 0.054 0.526 —0.465 2.648 —3.284 2.125 1.405
CPPI (p = 4) 0.122 1.158 —0.334 3.435 —7.425 5.982 3.084
CPPI (p = 6) 0.171 1.597 —0.535 4.532 —11.136 6.606 4.776
CPPI (p = 8) 0.188 1.797 —0.780 6.104 —14.680 7.089 6.149
CPPI (p = 10) 0.197 1.934 —1.459 15.124 —21.390 8.495 9.230
2b: Reference index: Low-risk index
CPPI (p = 2) 0.052 0.460 —0.586 4.055 —3.284 1.729 1.365
CPPI (p = 4) 0.112 1.011 —0.456 4.792 —7.425 4.849 2.959
CPPI (p = 6) 0.157 1.441 —0.687 6.188 —11.136 5.587 4.797
CPPI (p = 8) 0.173 1.621 —0.938 8.509 —14.680 5.587 6.208
CPPI (p = 10) 0.175 1.757 —1.866 21.763 —21.390 5.587 8.732
2c: Reference index: Price-weighted index
CPPI (p = 2) 0.059 0.586 —0.242 3.398 —3.419 2.843 1.503
CPPI (p = 4) 0.138 1.324 —0.195 5.868 —9.616 7.239 3.436
CPPI (p = 6) 0.182 1.739 —0.161 5.766 —11.167 10.944 4.412
CPPI (p = 8) 0.190 1.929 —0.426 6.384 —14.752 11.075 5.777
CPPI (p = 10) 0.188 2.081 —1.131 13.436 —21.850 11.075 8.845
Note: We report descriptive statistics for the in-sample weekly returns (non-annualized) of the monthly-rebalanced indices of the Dow 30 index (an equally
weighted index, a low-risk index and a price-weighted index) and dailyrebalanced CPPI portfolios (with multipliers p = 2, 4, 6, 8, 10). The full in-sample window
spans the horizon March 17, 1987 to February 3, 2009. We report the Portfolio (Column 1) the respective values for mean (Column 2), standard deviation (Column 3),
skewness (Column 4), excess kurtosis (Column 5), minimum percentage (Column 6), maximum percentage (Column 7) and modified expected shortfall (Column 8).
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›24
the non-normality of these returns, they constitute expected shortfall (mES) as a risk measure, which
an excellent case for advocating the use of MVSK is of the form proposed in Eqs. (8a) – (8e). This
portfolio risk measures. Note that these strategies risk measure accounts for the asymmetry and the
deliberately affect the portfolio moments through peakedness of the portfolio return distribution.
floor protection (as they cover the left tail and aim
for high upside potential). We take an ex-ante approach and simulate asset
returns for each out-of-sample day. We rely on
In Figure 4 we graph the effect of floor-protection on the Conditionally Heteroscedastic Independent
the performance and weights of an activelymanaged, Components Analysis Generalized Orthogonal
equally weighted Dow 30 index. (CHICAGO) GARCH model of Broda and Paolella
(2009) to model the temporal dependence up to the
In the top panel, we observe that the investment fourth-order moment of the multivariate asset return
algorithm leads to more concentrated returns as series, as detailed in Appendix C. The simulated out-
compared to the risky asset mix, indicative of the of-sample asset return distribution consists of 500
lower standard deviation and dampened minima- draws for each day in the five-step prediction interval.
maxima range. More interestingly, the bottom panel The model for the asset returns is re-estimated using
contains a visual depiction of the typical sharp and all the data up to that point in time.
abrupt movements in allocation for a strategy with
year-end resets. The fluctuations of the portfolio Table 4 provides a direct view of concentration risk
weights summarize the history of the CPPI’s exposure in floor-protected portfolios using the proposed
to the equally weighted index. In the allocation view companion table. The estimators – averaged across
of Figure 4, we observe that across 1987–1999 the 1,056 risk evaluation horizons – cover the average
investment strategy was strongly invested in the five-day risk of active CPPI fluctuations over the
market, showing many spikes in exposure to the risky period 1988-2009. Emphasis is placed on percentage
reference portfolio. The flat lines correspond to the sector-budgets, meaning that the budgets sum to
bounded portfolio value. As of 2000, we observe more 100 percent. The last column represents the position
turbulence in the strategy. We see that the allocation size with respect to these factors.
strongly moves away from the boundaries of 0 and
100 percent. Needless to say, such large swings in We observe similar levels of portfolio expected
the composition can have a significant impact on shortfall across the three CPPI portfolios (3.578, 3.937
the operations of the investor in gauging the risk at and 3.616 percent, respectively), rendering these
longer horizons. The portfolio at the beginning of the portfolios observationally equivalent using univariate
risk evaluation interval shares little relation to the tools. A meaningful practice of risk assessment
variety of positions that could be taken in the course recognizes that these similar levels of portfolio risk
of the next few days. are caused by different drivers of downside risk.
›25 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
Figure 4: The effect of daily floor-protection on the weights and performance of a CPPI invested in the equally weighted Dow 30 index
Note: We plot in-sample rebalancing results of a CPPI strategy (multiplier p = 6) on daily equally weighted (EW) Dow 30 index returns (risky reference portfolio). The
full in-sample window spans the horizon March 17, 1987 to February 3, 2009, consisting of 5;520 observations. We graph the weights signalled by the CPPI strategy
to be invested in the risky reference index (top), daily returns of an equally weighted Dow 30 portfolio (bottom-left) and CPPI-rebalanced returns (bottom-right).
To help understand what drives these portfolio- approach is that it yields both performance (proxied
level mean, variance, skewness and kurtosis budgets, by mean contributions) and risk contributions. It is
we rely on factor attribution analysis, offering us a insightful to jointly consider them in a scatter plot
picture where risk is coming from. We shade the three reporting the performance contribution to risk on
most likely ex-ante hot spots that merit the investor’s the vertical axis and the factor risk contributions on
attention. Despite the fact that these index portfolios the horizontal axis. In case the portfolio achieves the
contain the same Dow 30 constituents, we clearly highest relative performance, as is defined by the
recover a different structure of risk contributions ratio between portfolio performance and portfolio
under different investment strategies. The low-risk risk, all dots are on a line with a slope equal to the
CPPI portfolio, seems to shy away from Technology relative performance of the portfolio (see e.g., Ardia
and Computers & Technology. The diagnostic table et al., 2018). Components that are above (resp. below)
also points toward an excessive kurtosis contribution the line contribute relatively more (resp. less) to
in the Industrials category. In the price-weighted performance than to risk. We show this performance-
Dow 30 index we find a different factor mix: the risk allocation chart in Figure 5.
highest contributor is now the Technology sector.
These results show that it is possible to mix weight The relative performance of the portfolio can be
processes, while still distinguishing between factor marginally improved by decreasing (resp. increasing)
and moment contributions, which is of course the allocation to a position below (resp. above) the
essential from a diversification perspective. line. For example, we observe the largest absolute
deviation from maximum relative performance in
Once we have identified the hot spots, the next step the low-risk CPPI portfolio. Industrials stand out
in managing portfolio risk is to decide how to change as an obvious candidate for attention: the sector
the composition. An additional advantage of our contributes a substantial portion to risk without
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›26
Table 4: Sector-risk percentage risk allocation
Modified expected shortfall perc. contributions (%)
Mean St.Dev. Skew Kurtosis Total Avg. alloc.
(1) (2) (3) (4) (5) (6)
Panel 1: CPPI with equally weighted index (3:499%)
Energy (%) —0.787 11.445 —0.720 0.147 10.085 7.077
Financials (%) —0.990 7.190 —0.341 —0.099 5.760 9.024
Industrials (%) —2.556 34.378 —2.234 0.057 29.646 23.768
Consumer Cyclicals (%) —1.448 16.878 —1.101 0.054 14.383 14.534
Technology (%) —1.208 9.364 —0.427 —0.026 7.703 10.500
Computers & Technology (%) —0.562 3.395 —0.228 —0.036 2.569 4.069
Healthcare (%) —1.254 14.470 —0.983 0.064 12.296 10.011
Consumer Non-Cyclicals (%) —0.903 6.698 —0.324 —0.093 5.377 8.058
Telecom (%) —1.129 11.186 —0.519 0.092 9.629 8.872
Oil & Gas (%) —0.387 3.100 —0.145 —0.017 2.552 4.083
Residual (%) 0.000 0.001 0.001 0.000 0.001 0.004
Total —11.226 118.105 —7.021 0.142 100.000 100.000
Panel 2: CPPI with low-risk index (3:967%)
Energy (%) —0.767 12.891 0.944 0.022 11.202 8.265
Financials (%) —1.016 7.803 —0.681 0.118 6.224 11.416
Industrials (%) —2.591 35.037 —4.456 14.578 42.568 24.206
Consumer Cyclicals (%) —0.633 7.939 —0.672 —0.504 6.129 7.771
Technology (%) —0.040 0.326 —0.032 0.009 0.263 0.525
Computers & Technology (%) —0.456 3.380 —0.318 —0.160 2.446 4.142
Healthcare (%) —1.113 13.982 1.033 —6.531 7.370 10.491
Consumer Non-Cyclicals (%) —1.421 13.665 4.578 —11.510 5.313 14.319
Telecom (%) —1.326 13.304 —0.773 0.096 11.302 11.033
Oil & Gas (%) —0.801 8.437 —0.468 0.005 7.173 7.822
Residual (%) 0.000 0.001 —0.001 0.011 0.010 0.010
Total —10.164 116.765 —2.735 —3.866 100.000 100.000
Panel 3: CPPI with price-weighted index (3:517%)
Energy (%) —0.410 5.651 —0.422 0.159 4.978 3.794
Financials (%) —0.675 4.861 —0.051 V0.040 4.095 5.995
Industrials (%) —1.575 21.021 —3.156 1.156 17.446 16.593
Consumer Cyclicals (%) —2.112 27.097 —2.651 0.577 22.911 43.784
Technology (%) —3.256 36.334 4.195 —3.153 34.120 6.564
Computers & Technology (%) —0.216 1.070 —0.172 0.017 0.699 3.952
Healthcare (%) —0.858 7.870 —0.362 0.138 6.788 5.241
Consumer Non-Cyclicals (%) —0.866 5.427 1.413 —0.594 5.380 5.258
Telecom (%) —0.652 6.655 —7.481 3.133 1.654 5.869
Oil & Gas (%) —0.259 1.835 0.576 —0.225 1.927 2.948
Residual (%) 0.000 0.001 0.000 0.001 0.002 0.003
Total V10.879 117.820 —8.110 1.169 100.000 100.000
Note: We report average sector/moment percentage risk contributions to total modified expected shortfall for three CPPI portfolios; equally weighted (Panel 1), low-
risk (Panel 2) and price-weighted index portfolio (Panel 3). The out-of-sample window spans the horizon February 26, 1988 to March 3, 2009. The total risk levels are
reported in the headings of each panel. For each index, we report the risk budgeting table, consisting of the percentage mean contribution to risk (Column 1), the
percentage volatility contribution to risk (Column 2), the percentage skewness contribution to risk (Column 3), the percentage kurtosis contribution to risk (Column
4), the total modified expected shortfall (Column 5) and the average exposure vector (Column 6). Each column contribution is made up of sector contributions. For
each portfolio budget, we highlight the largest relative positive contributor and, in lightgray, the two subsequent highest contributors.
›27 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
Figure 5: Average performance-risk contributions of a CPPI invested in prior-return-weighted Dow 30 indices
(a) Equally weighted index
Note: We plot the performance-risk allocation chart for the different CPPI portfolios under consideration (i.e. equally weighted, low-risk and price). The vertical axis
shows the performance contribution to risk (proxied by mean contributions) and the horizontal axis shows the factor risk contributions of each of the sectors. The
slope of the line indicates the maximum relative performance, as defined by the ratio between portfolio performance and portfolio risk.
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›28
a corresponding contribution to performance. As
is clear in Table 4, the Industrials sector also has a
significant overweight position, partly explaining the
excessive risk contribution. It is likely that lowering
the exposure, at least at the margin, will reduce risk
and improve the relative performance.
›29 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
————————
6. Concluding remarks
————————
Sophisticated algorithmic techniques are
complementing human judgement across the fund
industry. Nonetheless, asset class strategists, risk
managers, or anyone who gauges the performance
of the investment strategy on an intermittent basis,
rarely acknowledges the feature that the composition
tends to change regularly, both within and across
days.
›31 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
————————
Appendix
————————
A. Multivariate factor co-moments and partial derivatives
The multivariate factor co-moments used in Section 2.2 are defined as
(14a)
(14b)
(14c)
in which denotes the q-dimensional of multi-period risk factor returns, is the vector of factor means and
⊗ denotes the Kronecker product. Then, for a given exposure vector , the second- to fourth-order moments
of the portfolio return are defined as
(15a)
(15b)
(15c)
(16a)
(16b)
For a given portfolio exposure vector , the partial derivatives to component j = (1, …, q) of the second-to
fourth-order moments of the portfolio return are defined as
(17a)
(17b)
(17c)
The corresponding partial derivatives of the portfolio skewness s and excess portfolio kurtosis k are given
by
(18a)
(18b)
The class of risk functions consists of density and quantile expressions and are typically expressed for the
random variable Z with mean zero and unit variance. We may write that the standardized innovations Z
follow a conditional distribution g with vector of shape parameters h, Z∼g(Z h). The shape parameter h
of the conditional distribution involves parameters that capture the asymmetry and fat-tailedness of the
distribution.
›33 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
When it is believed that the true pdf of the random variable Z is rather close to a normal one, most expansions
of the density are based on the form
(19)
where j(Z) is the density of a standard normal random variable and the function pd, a polynomial of degree
d, is such that the first moments of Z are equal to the moments of the approximating g(Z h). For example,
the Edgeworth expansion is a typical representation of this form,
. (20)
in which Hv denotes the vth Hermite polynomial and h = (g1, g2)'. It can be shown that the two parameters
g1 and g2 are directly related to the skewness s and excess kurtosis k of g(Z h).
Quantile approximations are also popular in the literature. The Cornish-Fisher equation is based on inverting
Eq. (19), in cumulative form, to obtain, for a level q ∈(0,1)
(21)
Using previously mentioned insights, Zangari (1996) defines the q % modified value-at-risk as
(22)
with the second-order Cornish-Fisher expansion around the Gaussian quantile function and
being the univariate portfolio volatility. Similarly, Boudt et al. (2008) and Martin and Arora (2017)
define q % modified expected shortfall as
with (23)
with
This equation involves the density of the Cornish-Fisher quantile f(gq) and powers of the same quantile
function gq.
Consider the random vector t = (r1,t,…, rn,t)', with being the (n,1) vector of conditional asset
means and being the information set consisting of past realizations t. The GO-GARCH model then maps
innovations, ut = t — t, onto an unobserved independent factor framework. The joint dynamics are given by
(24a)
(24b)
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›34
in which the innovation vector ut is model as a linear combination of d unobserved factors Lt or so-called
structural errors, A is an time-invariant, invertible mixing matrix. The authors impose the identifying condition
that factors are assumed to be independent and have unit variance. Via a polar decomposition, matrix A is
factorized into a symmetric, de-whitened square root of the unconditional covariance matrix S1/2 and an
orthogonal rotation matrix U:
A = S1/2 U (25)
The mixing matrix therefore represents the independent factor weights assigned to each asset. The factors
have a GARCH-type dynamics.
Originally, this model was estimated via a 1-step maximum likelihood approach to jointly estimate the rotation
matrix and factor dynamics. Differently, in the Conditionally Heteroscedastic Independent Components
Analysis Generalized Orthogonal (CHICAGO) GARCH model of Broda and Paolella (2009), the authors propose
a two-step approach which exploits the decomposition in Eq. (25). The unconditional covariance matrix is
obtained from the ordinary least squares residuals, the matrix U is estimated in a separate step by means
of Independent Component Analysis (ICA), which allow univariate factor dynamics and the innovations
distribution to be estimated separately from the covariance specification. The factor dynamics conditionally
follow the multivariate affine Generalized Hyperbolic (maGH) distribution of Schmidt et al. (2006), in which
each component is expressed as a linear combination of independent univariate General Hyperbolic variables.
These independent margins that allow to take separate values of skewness and kurtosis. Note that the factor
dynamics have non-time varying higher moments within the Generalized Hyperbolic distribution. For further
elaboration on the model interpretation and fitting routine, we refer to Boudt et al. (2019). For a practical
implementation, we refer to the R package rmgarch (Ghalanos, 2019).
›35 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
————————
References
————————
• Ardia, D., Boudt, K., and Nguyen, G. (2018). Beyond risk-based portfolios: balancing performance and risk
contributions in asset allocation. Quantitative Finance, 18(8):1249–1259.
• Ardia, D., Boudt, K., and Wauters, M. (2016). Smart beta and CPPI performance. Finance, 37(3):31–65.
• Bastürk,
¸ N., Borowska, A., Grassi, S., Hoogerheide, L., and van Dijk, H. K. (2019). Forecast density combinations
of dynamic models and data driven portfolio strategies. Journal of Econometrics, 210(1):170–186.
• Black, F. and Perold, A. F. (1992). Theory of constant proportion portfolio insurance. Journal of Economic
Dynamics and Control, 16(3-4):403–426.
• Boudt, K., Cornilly, D., and Verdonck, T. (2020a). A coskewness shrinkage approach for estimating the skewness
of linear combinations of random variables. Journal of Financial Econometrics, 18(1):1–23.
• Boudt, K., Cornilly, D., and Verdonck, T. (2020b). Nearest comoment estimation with unobserved factors.
Journal of Econometrics (In Press).
• Boudt, K., Galanos, A., Payseur, S., and Zivot, E. (2019). Multivariate GARCH models for large-scale applications:
A survey. In Vincod, H. D. and Rao, C. R., editors, Handbook of Statistics, Volume 41.
• Boudt, K., Peterson, B. G., and Croux, C. (2008). Estimation and decomposition of downside risk for portfolios
with non-normal returns. Journal of Risk, 11(2):79–103.
• Broda, S. A. and Paolella, M. S. (2009). CHICAGO: A fast and accurate method for portfolio risk calculation.
Journal of Financial Econometrics, 7(4):412–436.
• Connor, G. (1995). The three types of factor models: A comparison of their explanatory power. Financial
Analysts Journal, 51(3):42–46.
• De Prado, M. L. (2018). Advances in Financial Machine Learning. John Wiley & Sons.
• Diebold, F. X., Hickman, A., Inoue, A., and Schuermann, T. (1998). Scale models. Risk, 11:104–107.
• Francq, C. and Zakoïan, J.-M. (2018). Estimation risk for the VaR of portfolios driven by semi-parametric
multivariate models. Journal of Econometrics, 205(2):381–401.
• Fung, W. and Hsieh, D. A. (2001). The risk in hedge fund strategies: Theory and evidence from trend followers.
Review of Financial studies, 14(2):313–341.
• Ghalanos, A. (2019). rmgarch: Multivariate GARCH models. R package version 1.3-7.
• Gourieroux, C., Laurent, J.-P., and Scaillet, O. (2000). Sensitivity analysis of values at risk. Journal of Empirical
Finance, 7(3-4):225–245.
• Litterman, R. (1996). Hot spotsTM and hedges. Journal of Portfolio Management, 23(5):52–75.
• Martin, D. and Arora, R. (2017). Inefficiency of modified VaR and ES. Journal of Risk, 19(6):59–84.
• Scaillet, O. (2004). Nonparametric estimation and sensitivity analysis of expected shortfall. Mathematical
Finance, 14(1):115–129.
• Schmidt, R., Hrycej, T., and Stützle, E. (2006). Multivariate distribution models with generalized hyperbolic
margins. Computational Statistics & Data Analysis, 50(8):2065–2096.
• Scott, R. C. and Horvath, P. A. (1980). On the direction of preference for moments of higher order than the
variance. Journal of Finance, 35(4):915–919.
• Stoyanov, S. V., Rachev, S. T., and Fabozzi, F. J. (2013a). CVaR sensitivity with respect to tail thickness. Journal
of Banking & Finance, 37(3):977–988.
• Stoyanov, S. V., Rachev, S. T., and Fabozzi, F. J. (2013b). Sensitivity of portfolio VaR and CVaR to portfolio
return characteristics. Annals of Operations Research, 205(1):169–187.
• Van der Weide, R. (2002). GO-GARCH: A multivariate generalized orthogonal GARCH model. Journal of
Applied Econometrics, 17(5):549–564.
• Zangari, P. (1996). A VaR methodology for portfolios that include options. RiskMetrics Monitor, First
Quarter:4–12.
›37 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
• Zoia, M. G., Biffi, P., and Nicolussi, F. (2018). Value at risk and expected shortfall based on Gram-Charlier-like
expansions. Journal of Banking & Finance, 93:92–104.
• Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal
Statistical Society: Series B (Statistical Methodology), 67(2):301–320.
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›38
————————
About EDHEC-Risk Institute
————————
About EDHEC-Risk Institute
————————
Founded in 1906, EDHEC is one of the foremost international business schools. Operating from campuses in
Lille, Nice, Paris, London and Singapore, EDHEC is one of the top 15 European business schools. Accredited
by the three main international academic organisations, EQUIS, AACSB, and Association of MBAs, EDHEC has
for a number of years been pursuing a strategy of international excellence that led it to set up EDHEC-Risk
Institute in 2001. This Institute boasts a team of permanent professors, engineers and support staff, and
counts a large number of affiliate professors and research associates from the financial industry among its
ranks.
In this fast-moving environment, EDHEC-Risk Institute positions itself as the leading academic think-tank in the
area of investment solutions, which gives true significance to the investment management practice. Through our
multi-faceted programme of research, outreach, education and industry partnership initiatives, our ambition is
to support industry players, both asset owners and asset managers, in their efforts to transition towards a novel,
welfare-improving, investment management paradigm.
• The EDHEC-Princeton Retirement Goal-Based Investing Index Series, launched in May 2018, which represent
asset allocation benchmarks for innovative mass-customised target-date solutions for individuals preparing for
retirement;
• The EDHEC Bond Risk Premium Monitor, the purpose of which is to offer to investment and academic
communities a tool to quantify and analyse the risk premium associated with Government bonds;
• The EDHEC-Risk Investment Solutions (Serious) Game, which is meant to facilitate engagement with graduate
students or investment professionals enrolled on one of EDHEC-Risk’s various campus-based, blended or fully-
digital educational programmes.
›40 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
About EDHEC-Risk Institute
————————
EDHEC-Risk Institute’s seven research programmes explore interrelated aspects of investment solutions to
advance the frontiers of knowledge and foster industry innovation. They receive the support of a large
number of financial companies. The results of the research programmes are disseminated through the
EDHEC-Risk locations in the City of London (United Kingdom) and Nice, (France).
EDHEC-Risk has developed a close partnership with a small number of sponsors within the framework of
research chairs or major research projects:
• Financial Risk Management as a Source of Performance,
in partnership with the French Asset Management Association (Association Française de la Gestion
financière – AFG);
• ETF, Indexing and Smart Beta Investment Strategies,
in partnership with Amundi;
• Regulation and Institutional Investment,
in partnership with AXA Investment Managers;
• Optimising Bond Portfolios,
in partnership with BDF Gestion;
• Asset-Liability Management and Institutional Investment Management,
in partnership with BNP Paribas Investment Partners;
• New Frontiers in Risk Assessment and Performance Reporting,
in partnership with CACEIS;
• Exploring the Commodity Futures Risk Premium: Implications for Asset Allocation and Regulation,
in partnership with CME Group;
• Asset-Liability Management Techniques for Sovereign Wealth Fund Management,
in partnership with Deutsche Bank;
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›41
About EDHEC-Risk Institute
————————
The philosophy of the Institute is to validate its work by publication in international academic journals,
as well as to make it available to the sector through its position papers, published studies and global
conferences.
To ensure the distribution of its research to the industry, EDHEC-Risk also provides professionals with access
to its website, https://risk.edhec.edu, which is devoted to international risk and investment management
research for the industry. The website is aimed at professionals who wish to benefit from EDHEC-Risk’s
analysis and expertise in the area of investment solutions. Its quarterly newsletter is distributed to more
than 150,000 readers.
›42 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
About EDHEC-Risk Institute
————————
In 2012, EDHEC-Risk Institute signed two strategic partnership agreements. The first was with the Operations
Research and Financial Engineering department of Princeton University to set up a joint research programme
in the area of investment solutions for institutions and individuals. The second was with Yale School of
Management to set up joint certified executive training courses in North America and Europe in the area of risk
and investment management.
As part of its policy of transferring know-how to the industry, in 2013 EDHEC-Risk Institute also set up ERI
Scientific Beta, which is an original initiative that aims to favour the adoption of the latest advances in smart
beta design and implementation by the whole investment industry. Its academic origin provides the foundation
for its strategy: offer, in the best economic conditions possible, the smart beta solutions that are most proven
scientifically with full transparency in both the methods and the associated risks.
EDHEC-Risk Institute also contributed to the 2016 launch of EDHEC Infrastructure Institute (EDHECinfra), a
spin-off dedicated to benchmarking private infrastructure investments. EDHECinfra was created to address the
profound knowledge gap faced by infrastructure investors by collecting and standardising private investment
and cash flow data and running state-of-the-art asset pricing and risk models to create the performance
benchmarks that are needed for asset allocation, prudential regulation and the design of infrastructure
investment solutions.
An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020 ›43
————————
EDHEC-Risk Institute
Publications and Position
Papers (2018-2020)
————————
EDHEC-Risk Institute
Publications and Position Papers (2018-2020)
————————
2020
• Capasso, G., G. Gianfranco and M. Spinelli. Climate Change and Credit Risk
2019
• Martellini, L. and V. Milhau. Factor Investing in Liability-Drivenand Goal-Based Investment Solutions (December).
• Martellini, L. and V. Le Sourd. The EDHEC European ETF, Smart Beta and Factor Investing Survey 2019 (September).
• Maeso, J.M., Martellini, L. and R. Rebonato. Cross-Sectional and Time-Series Momentum in US Sovereign
Bond Markets (May).
• Maeso, J.M., Martellini, L. and R. Rebonato. Defining and Expoiting Value in Bond Market (May).
• Maeso, J.M., Martellini, L. and R. Rebonato. Factor Investing in Sovereign Bond Markets - Time-Series
Perspective (May).
2018
• Goltz, F. and V. Le Sourd. The EDHEC European ETF and Smart Beta and Factor Investing Survey 2018 (August).
• Mantilla-Garcia, D. Maximising the Volatility Return: A Risk-Based Strategy for Homogeneous Groups of
Assets (June).
• Giron, K., L. Martellini, V. Milhau, J. Mulvey and A. Suri. Applying Goal-Based Investing Principles to the
Retirement Problem (May).
• Martellini, L. and V. Milhau. Smart Beta and Beyond: Maximising the Benefits of Factor Investing (February).
›45 An EDHEC-Risk Working Paper — Predicting and Decomposing the Risk of Data-Driven Portfolios — February 2020
For more information, please contact:
Maud Gauchon on +33 (0)4 93 18 78 87
or by e-mail to: [email protected]
EDHEC-Risk Institute
393 promenade des Anglais
BP 3116 - 06202 Nice Cedex 3
France
Tel. +33 (0)4 93 18 78 87
risk.edhec.edu