Argo

Sight Deposit and

Non-Maturing Liability
Modelling

in a funding mix they contribute to abate the total

In this article we present a review of cost of funding.
the most significant approaches provided The modelling of deposits and non-maturing liabil-
by the literature and the market prac- ities is a crucial task for the liquidity management
tice for the modeling of non-maturing de- of a financial institution. It has become even much
more crucial in the current environment after the
posits accounts. We describe the bond
liquidity crisis that affected the money market in
portfolio replication approach and then 2008/2009. Typically ALM departments of banks,
move to the class of stochastic factor involved in the management of interest rate and
models, showing how the latter are ca- liquidity risks, face the task of forecasting deposit
pable of provide more effective tools for volumes, so as to design and implement consequent
the interest rate and liquidity risk man- liquidity strategies. Moreover deposit accounts rep-
resent the main source of funding for the bank,
agement of these balance-sheet items.
primarily for those institutions focused on the retail
business, and they heavily contribute to the fund-
Antonio CASTAGNA ing available in every period for lending activity.
Francesco MANENTI Amongst the different funding sources, deposits
have lower costs, so that in a funding mix they con-
tribute to abate the total cost of funding. Deposit
he modelling of deposits and non-maturing

T
contracts indeed have the peculiar feature of not-
liabilities is a crucial task for the liquidity having a predetermined maturity, since the holder
management of a financial institution. It has is free to withdraw the whole amount in every time.
become even much more crucial in the current en- The liquidity risk for the bank thus arises from the
vironment after the liquidity crisis that affected the mismatch between the term structures of assets and
money market in 2008/2009. Typically ALM de- liabilities of the banks balance sheet, being the li-
partments of banks, involved in the management abilities mostly made up by non-maturing items
of interest rate and liquidity risks, face the task of and the asset by long term investments (such as
forecasting deposit volumes, so as to design and mortgage loans).
implement consequent liquidity strategies. More- The optionality embedded in non-maturing prod-
over deposit accounts represent the main source ucts, related to the possibility for the customer to
of funding for the bank, primarily for those insti- arbitrarily choose any desired schedule of principal
tutions focused on the retail business, and they cash-flows, has to be understood and accounted
heavily contribute to the funding available in every for when performing liabilities valuation and hedg-
period for lending activity. Amongst the different ing the market and liquidity risk. Thus a sound
funding sources, deposits have lower costs, so that

Spring 2013
1
 ASSET LIABILITY MANAGEMENT

model is essential to deal with nested optionality [8].

for liquidity risk management purposes. Thirdly, the volatile part is invested in very short
term assets, typically in overnight deposits, and it
Modelling Approaches represents a liquidity buffer to cope with the daily
withdrawals by depositors.
There are two different approaches in the financial The critical point of this approach stands in the
literature and the market practice for the modelling estimation of the amortisation schedule of non-
of deposits balance evolution: maturing accounts, that is performed on statistical
- Bond portfolio replication, bases and have to be reconsidered periodically. One
of the flaws of the bond replica approach is that
- OAS models. the risk-factors affecting the evolution of the de-
posits are not modelled as stochastic variables. As
The bond portfolio replication, probably the most
such, once the statistical analysis is performed, the
common approach adopted by banks, can be shortly
weights are applied by considering the current mar-
described as follows. First, the total deposits
ket value of the relevant factors (basically, market
amount is split in two components:
and deposit rates) without considering their future
- a core part that is assumed to be not sensible evolution. This flaw is removed, at least partially,
to market variable evolution, such as interest by the so called Option Adjusted Spread (OAS)
rates and deposit rates. This fraction of the to- approach, which we prefer to define as Stochas-
tal volume of deposits is supposed to decline tic Factor (SF) approach.1 The approach is not in
gradually on a medium-long term period (say, principle different from the Bond Portfolio Replica
10 or 15 years) and to amortise completely at approach: one tries and identify statistically how
the end of it. the evolution of the deposits volume is linked to
risk factors (typically market and interest rates) and
- a volatile part that is assumed to be withdrawn then set up a hedge portfolio that covers the expo-
by depositors over a short horizon. This frac- sures them.
tion basically refers to the component of the The main difference lies in that, differently from
total volume of deposits that is normally used the Bond Portfolio Replica, in the SF approach the
by depositors to match their liquidity needs. weights of the hedging instruments are computed
Secondly, the core part is hedged with a portfolio considering the future random evolution of the risk
of vanilla bonds and money market instruments, factors, so that the hedging activity resembles the
whose weights are computed by solving an opti- dynamic replication of derivatives contracts. The
misation problem that could be set according to hedging portfolio is revised based on the market
different rules. Typically, the portfolio weights are movements of the risk factors, depending on the
chosen so as to replicate the amortisation schedule stochastic process adopted to model them. We pre-
of deposits or, equivalently said, their duration. In fer to work with a SF approach to model deposit
this way the replication portfolio tries and preserve volumes for several reasons. First, we think that
the economic value of the deposits (as defined later the SF approach is more advanced under a mod-
on) against the market interest rates movements. elling perspective, taking into account explicitly the
Another constraint, usually imposed in the choice stochastic nature of the risk factors. Secondly, if
of portfolio weights, is the target return expressed the Bond Portfolio Replica can be deemed adequate
as a certain margin over the market rates. Since for hedging the interest rate margin and the eco-
deposit rates are updated, within a relatively large nomic value of the deposits, under a liquidity risk
freedom of action, by banks to align them to market management point of view the SF approach is su-
rates, the replication portfolio can comprise fixed perior, for the very fact that it is possible to jointly
rate bonds, to match the inelastic part of the de- evaluate within a unified consistent framework the
posit rates that is not reactive to changes of market effects of the risk factors both on the economic
rates, and floating rate bonds, to match the elastic value and on the future inflows and outflows of the
part of the deposits rates. The process to re-balance deposits. Thirdly, it is easier to include in the SF ap-
the bond portfolio, although simple in theory, is proach complex behavioural functions linking the
quite convoluted in practice. For an more detailed evolution of the volumes to the risk-factors. Finally,
explanation of the mechanism, see Bardenhewer in bank-run events can be also considered and prop-
1 We think that OAS is misleading for a number of reasons: the approach does not explicitly model any optionality and does not

adjust any spread, as it will be clear from what we will show below. The name is likely derived from a contagion from the (bad)
practice, in the fixed income market, to use an effective discount rate to price assets taking into account embedded optionalities
(whence the name).

iasonltd.com
2
Argo
erly taken into account in SF approach, whereas it simulation-based framework like Montecarlo meth-
seems quite difficult their inclusion within a Bond ods. Since closed-form formulae for deposits value
Portfolio Replica approach. are expressed as risk-neutral expectations, the sce-
nario generation process has to be accomplished
The Stochastic Factor Approach with respect to the equivalent martingale probabil-
ity measure. For liquidity management purposes,
The first attempt to apply the SF approach, within it is more appropriate to use real-world parameter
an arbitrage-free derivatives pricing framework, to processes. In what follows we will not make any dif-
deposit accounts was made by Jarrow and van De- ference, though: as we have assumed also in other
venter [1]. They derived a valuation framework for parts of this book, assuming a risk-aversion param-
deposits based on the analogy between these liabil- eter equal to zero, real-world process for interest
ities and an exotic swap whose principal depends rates clash with risk-neutral ones. We propose a
on the past history of market rates. They provide a specification of the SF approach that we think is
linear specification for deposit volumes evolution enough parsimonious, yet effective.
applied to U.S. federal data.
Other similar models have been proposed,2 within Modelling of Market Interest Rates
the SF approach: it is possible to identify three
building blocks common to all of them: The dynamics for the market interest rates can be
chosen rather arbitrarily in the class of short rate
1. a stochastic process for the interest rates: in models. In the our specification we adopted a one-
the above mentioned Jarrow and van Deven- factor CIR++ model: we know that such model is
ter, for example, it is the Vasicek model; capable to to perfectly match the current observed
term structure of risk-free zero rates. The market
2. a stochastic model for the deposit rates: typi- instantaneous risk-free rate is thus given by
cally these are linked to the interest rates by
means of a more or less complex function; rt = xt + t
where xt has dynamics
3. a model for the evolution of the volume of de-
posits: since this is linked by some functional dxt = k ( xt ) dt + xt dWt
forms to the two risk factors at points 1 and
and t is a deterministic function of time.
2, it is a stochastic process as well.

The specification of deposit volumes dynamics is Modelling of Deposit Rates

the crucial feature distinguishing the different SF The deposit rate evolution is linked to the pricing
models: looking things under a micro-economic policy of the banks, providing a tool that can be
perspective, volumes depend on the liquidity pref- exploited to drive deposits volume across time. It is
erence and risk-aversion of depositors, whose be- reasonable to think that an increase of the deposit
haviour is driven by the opportunity costs between rate will work as an incentive for existing deposi-
alternative allocations. When market rates rise, de- tors not to withdraw from their accounts or to even
positors have a greater convenience to withdraw increase the amount deposited. The rate paid by
money from sight deposits and invest it in other as- the bank on deposit accounts can be determined ac-
sets offered in the market. SF models can be defined cording to different rules. Here are some examples:
behavioural in the sense that they try to capture the
dynamics of depositorss behaviour with respect 1. constant spread below market rates:
to market rates and deposit rates movements. In dt = max [rt , 0]
doing this, these models exploit the option pricing
to avoid having negative rates on the deposit,
technology, developed since 1970s, and depend on
there is a floor at zero.
stochastic variables, in contrast with the previously
mentioned class on simpler statistical models, as 2. a proportion of market rates:
mentioned above. Depositors behaviour is synthe-
dt = rt
sized in a behavioural function that depends on
risk factors and determines their choice in terms of 3. a function similar to the two above but depen-
amount allocated in deposits. This function could dent also on the amount deposited:
be specified in various forms, allowing for different
m
degrees of complexity. Given their stochastic nature,
those models are suitable to be implemented in
dt = i j (rt ) 1{Dtj ,Dtj+1 } Dt
j =1
2 See, amongst others, Frauendorfer and Schrle in [2], Dewachter et al. , Kalkbrener and Willing [9] and Blchlinger [4]

Spring 2013
3
 ASSET LIABILITY MANAGEMENT

where D j and D j+1 are range of the deposit Modelling of Deposit Volumes: Non-Linear Be-
volume D producing different level of the de- havioural Models
posits rate.
The behavioural function linking the evolutions of
We adopt a rule slightly more general than the pro- deposits volume to the risk factors can be also non-
portional one, i.e.: a linear affine relation between linear, possibly involving complex forms. In recent
the deposit rate and the market short rate years some efforts have been made to formulate this
dt = + rt + ut (1) relation according to more sophisticated functions,
trying to describe peculiar features of depositss dy-
where E (ut ) = 0 t. As it will be manifest in
namics. The main contribution in this direction was
what follows, the deposit volumes evolution de-
provided by Nystrm [6], who introduced in the val-
pends on the deposit rate, so in this framework
uation SF framework we are discussing a non-linear
the pricing policy function, that is obviously dis-
dependency of the deposit volumes dynamics from
cretionary for the bank, represents a tool to drive
the interest rates. The formalization of such dynam-
deposit volumes and consequently it can be used
ical behaviour is not trivial and we propose a model
to define liquidity strategies.
specification, inspired to the cited Nystrms work.
The main reason why non-linear behavioural func-
Modelling of Deposit Volumes: Linear Be- tions have been proposed is a drawback of equation
havioural Functions 2: it does not allow to fully capture the empirically
We can model the evolution of the total deposit vol- observed depositors reactions to market and de-
ume by establishing a linear relationship between posit rates movements. Actual behaviour exhibits
its log-variations and the risk factors, i.e.: the mar- high non-linearity with respect to these, in the sense
ket interest and deposit rates: this is the simplest be- that it depends not only on variations, as implied
havioural functional form we can devise. Moreover by equation 2, but also on the levels of market and
we add an autoregressive component, by imposing deposit rates.
that the log-variation of the volume at a given time The main idea in modelling the non-linear be-
is linked to the log-variation of the previous period haviour is based on the micro-economic liquidity
with a given factor and finally we include also a re- preference theory: depositors (and, generally speak-
lationship with the time, so as to detect time-trends. ing, investors) prefer to keep their investments in
The volumes evolution is in this case given by the liquidity for low levels of market rates. As mar-
equation: ket rates increase, the preference for liquidity is
counterbalanced so that depositors transfer higher
log Dt = 0 + 1 log Dt1 + 2 t fractions of their income and wealth to less liquid
(2)
+ 3 rt + 4 dt + et investments. In more detail, the first variable to con-
with being the first-order difference operator and sider is the total depositors income, I, growing at
et the idiosyncratic error term with zero mean. This an annual rate : on an aggregated base we could
formula is in practice the same as in Jarrow and see it as the growth rate of the economy (GDP)
van Deventer. or simply the growth rate of the income for each
The model in 2 is convenient because parame- depositor (customer). Secondly, the allocation of
ters can be easily estimated on historical data via the income between deposits and other (less liquid)
standard OLS algorithm. The presence of a time investments hinges on the following assumptions:
component in equation 2 is justified by empirical
evidence on deposit series, that exhibit a trend com- - each depositor modifies his balance on the de-
ponent. This factor could be modelled in alternative posit account targeting a given fraction of
ways, substituting the linear trend with a quadratic their income I. This level can be interpreted
or exponential one. For interest rate risk manage- as the amount they need to cover his short-
ment purposes, we can however be interested in time liquidity needs. At any time t, given the
understanding how deposit evolution explained current fraction t of the income invested in
only by market and deposit rates movements. To deposit, the adjustment toward the target
this end, we can introduce a reduced version of occurs at a speed ;
the model that is estimated excluding the trend
component, i.e.: - there is an interest rates strike level E, specific
to the customer, such that, when the market
log Dt = 0 + 1 log Dt + 2 rt + 3 dt (3)
rate is above it, then they reconsider the target
Empirical analysis of both the models form will be level and redirects a higher amount to other
presented below. investments, by a fraction of their income;

iasonltd.com
4
Argo
40

35

30

25
Disitribution 1
20
Distribution 2
15

10

0.02
0.04
0.06
0.08

0.12
0.14
0.16
0.18

0.22
0.24
0.25
0.27
0.29
0.31
0.33
0.35
0.37
0.1

0.2
0

FIGURE 1: Two possible distributions produced by the Gamma function. On the x-axis: the interest rate level.

- there is an deposit rates strike level F, spe- We can alternatively use the equivalent func-
cific to the customer, such that, when the rate tional form of the Gamma distribution written as:
received on deposits is above it, then they
1 t
are more reluctant to withdraw money by a h ( x; k, ) := t k 1 e
fraction of their income; k (k)

Under these assumptions, the evolution of the frac- This is actually what we will use in the estimation
tion t of the income allocated in sight deposit is: of the parameters from historical data we will show
below.
t+t t = ( t )t + 1[ E,) (rt )
(4)
+ 1[ F,) (rt ) The evolution of total volume of deposits can be
written by modifying equation 4 and considering
where 1[ E,) is the indicator function equal to 1
the distributions of the strike rates instead of the
when the condition in the subscript is verified. The
single strike rates for each depositor:
income I grows as follows:
It+t It = It t (5) t+t t = ( t )t
(7)
+ H (rt , k1 , 1 ) + H (dt , k2 , 2 )
and the deposit volume at time t is:
Rx
Dt = t It (6) where H ( x, k, ) = 0 h(u; k, )du is the Gamma
cumulative distribution function.
In reality, since each depositor has different level To make the econometric estimation of the pa-
of strike rates E and F, due to their preferences for rameters easier, we rewrite equation 7 in the follow-
liquidity, on an aggregated basis, considering all ing way:
banks customers, there is a distribution of strike
rates, reflecting their heterogeneity in behaviour. t = + t1 + H (rt , k1 , 1 ) + H (dt , k2 , 2 ) (8)
So, when we pass from the evolution of the single
deposit, to the evolution of the total volume of de- where = t and = t. Equation 8 can
posits on banks balance sheet, strike rates can be be applied by the bank to the average customer.
thought to be distributed according to any suitable Given the heterogeneity of the behaviours, given
probability function h( x ): in the specification we current market and deposit rates, the incentive to
present here we choose a Gamma function, i.e.: change the income allocation by increasing less liq-
uid investments, balanced by the incentive to keep
( x/) 1 exp( x/)
h( x; , ) = the investment in deposits provided by the deposit
()
rates, is synthesized in the Gamma distribution
functions, so that H ( x, k, ) turns out to be the cu-
EXAMPLE 1 The Gamma function is very flex- mulative density of the average customers strike.
ible and it allows for a wide range of possible
shapes of the distribution. If we set = 1.5 and Economic Evaluation and Risk Manage-
= 0.05, or example, we have a distribution la- ment of Deposits
beled as 1 in Figure 1. If = 30 and = 0.002
we have a distribution 2. It is possible to model The three building blocks to model the deposits can
the aggregated customers behaviour, making it be used to compute the economic value to the bank
more or less concentrated around specific levels. of the total amount held on balance sheet. At time

Spring 2013
5
 ASSET LIABILITY MANAGEMENT

FIGURE 2: On the left: time series of 1-month Eonia swap rates for the period 3/1999:4/2012; on the right: actual time series of
deposit rates vs. fitted values.

FIGURE 3: Actual time series of deposit volumes vs. fitted values for the linear behavioural model.

t = 0, for a time horizon T, the economic value to launch a Montecarlo simulation on the two risk
is the expected margin that can be earned by the factors, i.e.: the risk-free instantaneous interest rate
bank on the present and future volume of deposits. and deposit rate. We operate the following steps:
In fact, the amount of funds raised by the bank in - given a time-horizon T, divide the period
form of deposits can be invested at in short expiry [0, T ] in M steps,;
risk-free investments yielding rt ; on the other hand
the deposits cost to the bank the rate dt that it has - simulate N paths for each risk factor;
to pay to depositors. In formula: - compute the expected level of deposit volume
n Z T h i V (0, Ti ) at each step i {0, 1, ..., M }, by av-
D
EQ rt d j,t D j,t P D (0, t) dt eraging out on the N scenarios, by means of


V (0, T ) =
j =1 0 equation 2 or 8:
(9) M m
m =1 D ( Ti )
where D j,t is the amount deposited in the account j D e ( Ti ) = E [ D ( Ti )] =
M
at time t and n is the number of deposit accounts.
The expectation is taken under the equivalent mar- - compute the stressed level of deposit volume
tingale risk-neutral measure Q. Equation 9 is the ex- at a given confidence level p, V p (0, Ti ) at each
pected Net Interest Margin to the bank over the pe- step at each step i {0, 1, ..., M }, based in
riod [0, T], for all the deposits accounts, discounted the M scenarios. Since for liquidity risk man-
in 0 by the risk-free discount factor P D . As sug- agement purposes the bank is interested at
gested by Jarrow and van Deventer [1], the value of the minimum levels of the deposit volume at
deposits can be seen as the value of an exotic swap, a given time Ti , then we define the stressed
paying the floating rate d j,t and receiving the float- level at p confidence level as:
ing rate rt , on the stochastic principal D j,t for the D p ( Ti ) = inf { D ( Ti ) : Pr[ D ( Ti ) < D p ( Ti )] p}
period between 0 and T. The approach we outlined
above is also a good tool for liquidity risk manage- Banks can be interested in computing the minimum
ment, since it can be used to predict expected or level of deposits during the entire period included
stressed (at a given confidence level) evolution of between the reference time (say, 0) and a given time
deposit volume. To compute these metrics, we need Ti : this is actually the value that corresponds to

iasonltd.com
6
Argo
the actual available liquidity that can be used for functions we have presented above, based on pub-
investments expiring in Ti . To this end it is useful to lic aggregated data for sight deposits in Italy.
introduce the process of the minima of the deposit We considered a sample of monthly observa-
volume, defined as: tions in the period 3/1999 : 4/2012 for sight
deposits total volume and average deposit rates
Dmin ( Ti ) = min D (s) paid by the bank. Data for deposits are pub-
0s Ti
lished by Bank of Italy (Bollettino Statistico).3
Basically the process exclude all the growth of the We considered the euro 1-month overnight in-
volume of deposits due to new deposits or to an dex average (Eonia swap) rate as a proxy for the
increase of the amount of the exiting ones, but it market short risk-free rate: values for the anal-
considers only the abating effects that the risk fac- ysis period are plotted in Figure 2 on the left.
tors produce. The metric is also consistent with the The CIR model for the market rate was calibrated
factual truth that in any case the bank can never on the time series of Eonia rates via Kalman filter,
invest more than the existing amount of deposits it and the resulting values for the parameters are:
has on its balance sheet. The SF approach can be
= 0.053, = 7.3, = 8.8%
used also for interest rate management purposes.
Once we have the computed the economic value of For the second building block (deposit rates), the
deposits, it is straightforward to compute its sensi- linear relation between market rates and deposit
tivities to risk-factors to set up hedging strategies rates in Equation 1 has been estimated via stan-
with liquid market instruments such as swaps. To dard OLS, results are shown in Table 1. Figure
this end, we can calculate the sensitivities of de- 2 plots on the right the actual time series of de-
posits economic value to perturbations in the mar- posit rates and fitted values from the estimated
ket zero-rate curve. The sensitivity to the forward regression. The model shows a good fitting of
rate F (0; ti , ti+1 ) = Fi (0) is obtained numerically by the time series and we can observe that the linear
means of the following: affine relation is strongly consistent with the data.

V (0, T; Fi (0)) = V 0, T; Fi (0) V (0, T; Fi (0))

(10) Coefficient Significance (p-values)

Intercept 1.05 0.042
where V ()is provided by (9) and Fei (0) is the rele- Market rate rt 0.92 1.86E-51
vant forward rate bumped by a given amount (e.g.: R2 0.92
10 bps). We have assumed that the instantaneous F statistics 1773
short rate follows a one-factor CIR++ dynamics. F significance 4.23E-87
Assuming now that the initial zero-rate curve gen-
n TABLE 1: Regression results for the deposit rates equation 1.
erated by the model, i.e.: the series P D (0, Ti ) i=1 ,


perfectly matches the market observed term struc-

ture, we have to modify the short rate dynamics Coefficient Significance (p-values)
Intercept 1.05 0.042
in a way that produces the desired bump on the Lagged Dt1 0.92 1.86E-51
forward rates time 0, by suitably modifying the Time t 0.4E-3 0.093
the deterministic time dependent term (t) of the Market rate variations rt -3.45 0.001
Deposit rate variation dt 7.54 0.009
CIR++ process. This is easily done: let bmp be the R2 0.99
size of the bump to the term structure of starting F statistics 4518
forward rate Fi (0); in the CIR++ the tilted forward F significance 1.17E-157

Fi (0) is obtained by modifying the integrated time TABLE 2: Regression results for the linear behavioural equa-
dependent function (t) as: tion 2.
Z T Z T
i +1 i +1 ln(bmp) Finally, we need to adopt a behavioural function.
(s)ds (s) + ds
Ti Ti i We start with the linear model for deposit volumes
where i = Ti+1 Ti . We present below some prac- in Equation 2: estimation results are shown in Table
tical applications to the approach sketched above. 2 and also in this case the model proves to be a good
explanation of the data. We note that the signs of
coefficients multiplying, respectively, the variations
EXAMPLE 2 in the market rate and the variations in the de-
We perform an empirical estimation and test of posit rate, are opposite as expected. Figure 3 plots
the the SF approach, with the two behavioural actual and fitted time series of deposit volumes.
3 Data are available also at the web site www.bancaditalia.it.

Spring 2013
7
 ASSET LIABILITY MANAGEMENT

FIGURE 4: From the top on the left clockwise: simulated paths for 1-month Eonia swap rate, deposit rate and deposit volume, term
structure of expected and minimum (99% c.l) future volumes.

FIGURE 5: On the left, simulated paths for deposit volume; on the right: term structure of expected and minimum (99% c.l) future
volumes.

FIGURE 6: On the left: time series of Italian nominal GDP for the sample 3/1999:4/2012; quarterly data are linearly interpolated to
obtain the monthly time series. On the right: Actual time series of deposit volumes vs. fitted values for the non-linear behavioural
model.

iasonltd.com
8
Argo
We can now use estimated parameters to com- Coefficient Significance (p-values)
Intercept 0.25 1.09e-167
pute the economic value of deposits via Monte-
Lagged t1 0.53 1.20e-158
carlo simulations of the Formula 9. The stan- Gamma market rates H (rt ) -0.09 4.12e-083
dard approach requires to generate a number of Gamma market rates 1 18.77 1.13e-077
simulated paths for the risk factors by means of Gamma market rates k1 0.001 3.86e-081
the estimated dynamics, following these steps: Gamma deposit rates H (dt ) 0.14 0.0054
Gamma deposit rates 2 24.26 1.67e-066
Gamma deposit rates k2 0.001 3.01e-056
- compute 10,000 paths for the market rate R2 0.97
evolution, simulated with the CIR dynamics; F statistics 4518
F significance 1.1767E-157

- for each path, compute the corresponding TABLE 4: Regression results for the non-linear behavioural
path for the deposit rate and the deposit equation 8.
volume according to estimated regressions
(equations 1 and 2); We now estimate the parameters of the non-linear
behavioural model in Equation 8, via a Non-Linear
Least Squares algorithm; we still use the same
- compute deposits value at each time steps in
dataset as above, i.e. the sample 3/1999 : 4/2012
the simulation period;
of monthly data for non-maturing deposits vol-
umes, 1-month Eonia swap rates and deposit rates.
- sum discounted values path by path and In this case, what we actually model is the evolu-
average them to obtain the present value of tion of the proportion of the depositors income
the total amount of deposits. held in a sight deposit. At an aggregated level,
we approximated the total income with the nom-
Figure 4 shows simulated paths for state variables, inal GDP, so that the fraction will be referred
using the CIR process with the estimated param- to this quantity.Since we are working with Italian
eters, starting from the first date after the end of deposits, we take the Italian GDP data that are pub-
the sample period, the average path of deposits lished quarterly and we operate a linear interpola-
volume and the minimum amount computed at tion to obtain monthly values.4 The reconstructed
the 99% c.l. With an initial deposits total vol- nominal GDP time series, for the estimation pe-
ume of 834, 468 bln euros and a simulation pe- riod we consider, is shown in Figure 6 on the left.
riod of 10 years, the estimated economic value Estimated coefficients and their significance are
to the bank of holding deposits is 121, 030 bln. shown in Table 4. Figure 7 plots the pdf of the strike
We provide empirical results also for the reduced respectively for the market (E) and the deposit (F)
version of the linear behavioural model given rates. We can see that the cumulative density func-
in Equation 3. Table 3 reports regression pa- tions reach their maximum when the market rate ex-
rameters for this model, and simulated paths ceeds 3.55% and deposit rate exceeds 4.25%. These
are plotted in Figure 5. As expected, exclud- should be considered the levels for market interest
ing the time trend, deposits volume forecast rates and deposit rates when most of customers
is much more conservative, and the minimum consider re-allocating the fraction of income held in
volume at the 99% c.l. rapidly decreases. deposits on other investments. The regression has
an R2 value lower than the linear model tested be-
fore: this is also confirmed by the plot of the actual
vs fitted deposits volumes in Figure 6 on the right.
Coefficient Significance (p-values) As already done for the linear model, we can
Intercept Intercept 0.19 0.048 compute the economic value of deposits with a
Lagged Dt1 0.98 2.34E-159
Market rates variations rt -4.1 2.57E-04 Montecarlo simulation. Figure 8 shows simu-
Deposit rates variation dt 6.52 0.04 lated paths deposits volumes and the term struc-
R2 0.99
ture of expected and minimum volumes. With
F statistics 5837
F significance 1.66E-157 a simulation period of 10 years and an initial
volume of 834, 467 bln Eur, the estimated de-
TABLE 3: Regression results for the reduced version of the lin- posits economic value is 88, 614, so the non-linear
ear behavioural equation 3.
model is more conservative than the linear one.
Also for the non-linear model, we can run Mon-
tecarlo simulations after freezing the time-trend
4 We are aware this is likely not the most sound way to interpolate GDP data, but we think it is reasonably good for the limited

purpose of our analysis.

Spring 2013
9
 ASSET LIABILITY MANAGEMENT

FIGURE 7: Gamma probability density function of the strike level for market interest rates (upper graph) and for deposit rates
(lower graph), given the estimated parameters.

FIGURE 8: Simulated paths (upper graph) and term structure of expected and minimum (99% c.l) future volumes (lower graph)
derived with the estimated non-linear model.

(which in this case means keeping the GDP con- of deposits volume are concerned, is shown in
stant to the initial level) and the deposit rate. Figure 11. It is quite clear that the non-linear
In this way we isolate the effect produced by model seems to be much more conservative in
the market interest rates on the deposits volume. terms of expected and minimum level of volumes.
The results for the case when only the time- We present a comparison also of the market
trend is frozen, are shown in Figure 9. It is rate sensitivities of the economic value of the
worth noticing that without time-trend, the frac- deposits obtained by the linear and non-linear
tion of income held in deposits rapidly reach the model. In Table 5 sensitivities to the 1-year for-
minimum and then, given the autoregressive na- ward (risk-free) Eonia rates, fixed every year up
ture of the model in Equation 8, it keeps con- to 10 years, are shown. Sensitivities are referred
stant at this level. Figure 10 shows the results to a bump of the relevant forward rate of 10
when both the time trend and the deposit rate basis points. The linear model has bigger sen-
are frozen: qualitatively they are the same as sitivities due to the higher volumes, and hence
in the case when only the time-trend is frozen. higher economic value, expected in the future.

Sensitivity Sensitivity Inclusion of Bank-Runs

Years Linear Model Non-linear Model
1 730 524 It can be interesting to include the possibility of a
2 770 683 bank-run in the future, due to a lack of confidence
3 810 663
of the depositors in the creditworthiness and the
4 860 657
5 910 659 accountability of the bank. If this occurs, it is rea-
6 960 664 sonable to expect a sharp and sudden decline in the
7 1000 678 deposits volumes. To take into account a bank-run,
8 1050 689 one needs to consider some variable that is linked
9 1090 703
10 1100 723
to the banks credit robustness (or the lack of it).
One possible solution could be the credit spread
TABLE 5: Sensitivities to 1Y1Y forward Eonia swap rates for of the bank, on short or long-term debt: it can be
10 bps up to 10 years, for the linear and non-linear model. either extracted from market quotes of the bonds
issued by the bank, of from banks CDS quotes. As
A comparison between the linear and non-linear for the model, for the very nature of the bank-run,
model, as far as the expected and minimum level the non-linear behavioural model is more suitable

iasonltd.com
10
Argo

FIGURE 9: Simulated paths (upper graph) and term structure of expected and minimum (99% c.l) future volumes (lower graph)
derived with the estimated non-linear model, when the time-trend is frozen.

FIGURE 10: Simulated paths (upper graph) and term structure of expected and minimum (99% c.l) future volumes (lower graph)
derived with the estimated non-linear model, when the time-trend and the deposit rate are frozen.

FIGURE 11: Term structure of expected and minimum (99% c.l) future volumes (lower graph) derived with the linear (equation 2)
and the non-linear (equation 8) model.

FIGURE 12: Simulated paths (upper graph) and term structure of expected and minimum (99% c.l) future volumes in the case of
bank-run inclusion.

Spring 2013
11
 ASSET LIABILITY MANAGEMENT

to accommodate for it. In fact, it is possible to add shown in Table 4. The parameters of the ad-
an additional behavioural function, related to the ditional behavioural function are set as follows:
banks credit spread, which will likely be densely
concentrated around a high level (denoting an id- = 0.2, k3 = 32, 3 = 0.002
iosyncratic critical condition). The inclusion of the
bank-run is operated by extending formula (8) as Given the parameters of the Gamma function k3
follows: and 3 , when the credit spread of the bank reaches
a level above 800 bps, then a drop of 20% in the level
t = + t + H (rt ; k1 , 1 ) of the deposits is experience in each period (we re-
(11)
+ H (it , k2 , 2 ) + H (stB ; k3 , 3 ) call we use monthly steps in our examples). To
model the credit spread and simulate its evolution
The new behavioural function H (stB ; k3 , 3 ) is still in the future, we assume that the default intensity of
a Gamma function taking as an input the banks the bank is given by a CIR process, with parameters:
spread s B . It is quite difficult to estimate the param-
eters of this function, since it is quite unlikely that 0 = 0.2, = 0.5, = 5%, = 12%
the bank experienced many bank-runs. One can
resort to bank-runs occurred to comparable banks, Besides we assume a = 60% upon banks default.
but also in this case not to many events can be ob- We assume that the spread entering in the be-
served for a robust estimation of the parameters. havioural function is the 1-month one, for short-
Nonetheless, the bank can include the bank-run on term debt. Figure 12 shows the simulated paths
a judgmental basis, by assigning given values to the and the term structure of the expected and min-
behavioural function according to its hypothesis of imum volume of deposits: when compared with
stressed scenarios. Figure 8 it is evident the lower levels projected by
the model.
EXAMPLE 3
We extend the non-linear model we estimated ABOUT THE AUTHORS
in Example to include the possibility of a
Antonio Castagna is Senior Consultant at Iason.
bank-run. To compute the term structure of Email address: [email protected]
expected and minimum volume of deposits, Francesco Manenti is Consultant at Iason and corresponding
we use Equation (11), with parameters set as author. Email address: [email protected]

References

[1] Jarrow, R.A. and D.R. van Deven- maturing deposits Risk 19(11):82-87, liabilities. Journal of Economic Dy-
ter. The arbitrage-free valuation and 2006. namics & Control, 32:709-756, 2008.
hedging of demand deposits and
credit card loans. Journal of Bank- [4] Blchlinger A. Management frame- [7] Boekel P. et al. Replicating port-
ing and Finance 22:249-272, 1998. work for non-maturing account: folios. Milliman Research Report,
from the marketing to the hedging 2009.
strategy 2011.
[2] Frauendorfer, K. and M. Schrle.
[8] Matz, L. and P. Neu. Liquidity risk
Dynamic modelling and optimiza- [5] Janosi, T., R.A. Jarrow and F. Zullo.
An empirical analysis of the Jarrow measurement and management. Wi-
tion of non-maturing accounts.
- van Deventer model for valuing ley Finance, 2006.
Working Paper. University of
St.Gallen, 2006. non-matury demand deposits. The
Journal of Derivatives, 7:8-31, 1999. [9] Kalkbrener P. and J. Willing. Risk
management of non-maturing liabil-
[3] Elkenbracht, M. and B. Nauta. Man- [6] Nystrm K. On deposit volumes ities. Journal of Banking and Fi-
aging interest rate risk for non- and the valuation of non-maturing nance 28(7):1547-1568, 2004.

iasonltd.com
12