Background and contribution

Complex quantitative models are used extensively in actuarial and financial risk management applications, as well as in wider fields such as environmental risk modelling (Tsanakas and Millossovich 2016; Borgonovo and Plischke 2016; Silvana M. Pesenti, Millossovich, and Tsanakas 2019). The complexity of such models (high dimensionality of inputs; non-linear relationships) motivates the performance of sensitivity analyses, with the aim of providing insight into the ways that model inputs interact and impact upon the model output.

When model inputs are subject to uncertainty, global sensitivity methods are often used, considering the full space of (randomly generated) multivariate scenarios, which represent possible configurations of the model input vector. The particular task of ranking the importance of different model inputs leads to the use of sensitivity measures, which assign a score to each model input. A rich literature on global sensitivity analysis exists, with variance decomposition methods being particularly prominent; see Saltelli et al. (2008) and Borgonovo and Plischke (2016) for wide-ranging reviews. The R package sensitivity (Iooss, Janon, and Pujol 2019) implements a wide range of sensitivity analysis approaches and measures.

We introduce an alternative approach to sensitivity analysis called Scenario Weights for Importance Measurement (SWIM) and present the R package implementing it (Silvana M. Pesenti et al. 2020). The aim of this paper is to provide an accessible introduction to the concepts underlying SWIM and a vignette demonstrating how the package is used. SWIM quantifies how distorting a particular model component (which could be a model input, output, or an intermediate quantity) impacts all other model components. The SWIM approach can be summarised as follows:

1. The starting point is a table of simulated scenarios, each column containing realisations of a different model component. This table forms the baseline model as well as the dataset on which the SWIM bases its calculations.

2. A stress is defined as a particular modification of a model component (or group of components). This could relate to a change in moments, probabilities of events of interest, or risk measures, such as Value-at-Risk or Expected Shortfall (e.g. McNeil, Frey, and Embrechts (2015)).

3. SWIM calculates a set of scenario weights, acting upon the simulated scenarios and thus modifying the relative probabilities of scenarios occurring. Scenario weights are derived such that the defined stress on model components is fulfilled, while keeping the distortion to the baseline model to a minimum, as quantified by the Kullback-Leibler divergence (relative entropy).

4. Given the calculated scenario weights, the impact of the stress on the distributions of all model components is worked out and sensitivity measures, useful for ranking model components, are evaluated.

A key benefit of SWIM are that it provides a sensitivity analysis framework that is economical both computationally and in terms of the information needed to perform the analysis. Specifically, sensitivity analysis is performed using only one set of simulated scenarios. No further simulations are needed, thus eliminating the need for repeated evaluation of the model, which could be numerically expensive. Furthermore, the user of SWIM needs to know neither the explicit form of the joint distribution of model components nor the exact form of functional relations between them. Hence, SWIM is appropriate for the analysis of black box models, thus having a wide scope of applications.

The SWIM approach is largely based on Silvana M. Pesenti, Millossovich, and Tsanakas (2019) and uses theoretical results on risk measures and sensitivity measures developed in that paper. An early sensitivity analysis approach based on scenario weighting was proposed by Beckman and McKay (1987). The Kullback-Leibler divergence has been used extensively in the financial risk management literature – papers that are conceptually close to SWIM include Weber (2007); Breuer and Csiszár (2013); and Cambou and Filipović (2017). Some foundational results related to the minimisation of the Kullback-Leibler divergence are provided in Csiszár (1975).

Installation

The SWIM package can be installed from CRAN or through GitHub:

r} # directly from CRAN install.packages("SWIM") # and the development version from GitHub devtools::install_github("spesenti/SWIM")

Structure of the paper

Section @ref(Sec:Intro) provides an introduction to SWIM, illustrating key concepts and basic functionalities of the package on a simple example. Section @ref(Sec:Scope) contains technical background on the optimisations that underlay the SWIM package implementation. Furthermore, Section @ref(Sec:Scope) includes a brief reference guide, providing an overview of implemented R functions, objects, and graphical/analysis tools. Finally, a detailed case study of a credit risk portfolio is presented in Section @ref(Sec:CreditModel). Through this case study, advanced capabilities of SWIM for sensitivity analysis are demonstrated.

Sensitivity testing and scenario weights

The purpose of SWIM is to enable sensitivity analysis of models implemented in a Monte Carlo simulation framework, by distorting (stressing) some of the models’ components and monitoring the resulting impact on quantities of interest. To clarify this idea and explain how SWIM works, we first define the terms used. By a model, we mean a set of n (typically simulated) realisations from a vector of random variables (X₁, …, X_d), along with scenario weights W assigned to individual realisations, as shown in the table below. Hence each of of the columns 1 to d corresponds to a random variable, called a model component, while each row corresponds to a scenario, that is, a state of the world.

Each scenario has a scenario weight, shown in the last column, such that, scenario i has probability $\frac{w_i}{n}$ of occurring. Scenario weights are always greater or equal than zero and have an average of 1. When all scenario weights are equal to 1, such that the probability of each scenario is $\frac 1 n$ (the standard Monte Carlo framework), we call the model a baseline model – consequently weights of a baseline model will never be explicitly mentioned. When scenario weights are not identically equal to 1, such that some scenarios are more weighted than others, we say that we have a stressed model.

The scenario weights make the joint distribution of model components under the stressed model different, compared to the baseline model. For example, under the baseline model, the expected value of X₁ and the cumulative distribution function of X₁, at threshold t, are respectively given by:

$$ E(X_1)=\frac 1 n \sum_{i=1}^nx_{1i},\quad F_{X_1}(t)= P(X_1\leq t)=\frac 1 n \sum_{i=1}^n \mathbf 1 _{x_{1i}\leq t}, $$

where 1_{x1i ≤ t} = 1 if x_1i ≤ t and 0 otherwise. For a stressed model with scenario weights W, the expected value E^W and cumulative distribution function F^W become:

$$ E^W(X_1)=\frac 1 n \sum_{i=1}^n w_i x_{1i},\quad F_{X_1}^W(t)=P^W(X_1\leq t)=\frac 1 n \sum_{i=1}^n w_i \mathbf 1 _{x_{1i}\leq t}. $$

Similar expressions can be derived for more involved quantities, such as higher (joint) moments and quantiles.

The logic of stressing a model with SWIM then proceeds as follows. An analyst or modeller is supplied with a baseline model, in the form of a matrix of equiprobable simulated scenarios of model components. The modeller wants to investigate the impact of a change in the distribution of, say, X₁. To this effect, she chooses a stress on the distribution of X_i, for example requiring that E^W(X₁) = m; we then say that she is stressing X₁ and, by extension, the model. Subsequently, SWIM calculates the scenario weights such that the stress is fulfilled and the distortion to the baseline model induced by the stress is as small as possible; specifically the Kullback-Leibler divergence (or relative entropy) between the baseline and stressed models is minimised. (See Section @ref(Rfunctions) for more detail on the different types of possible stresses and the corresponding optimisation problems). Once scenario weights are obtained, they can be used to determine the stressed distribution of any model component or function of model components. For example, for scenario weights W obtained through a stress on X₁, we may calculate

$$ E^W(X_2)=\frac 1 n\sum_{i=1}^n w_i x_{2i},\quad E^W(X_1^2+X_2^2)=\frac 1 n \sum_{i=1}^n w_i \left(x_{1i}^2+ x_{2i}^2 \right). $$

Through this process, the modeller can monitor the impact of the stress on X₁ on any other random variable of interest. It is notable that this approach does not necessitate generating new simulations from a stochastic model. As the SWIM approach requires a single set of simulated scenarios (the baseline model) it offers a clear computational benefit.

An introductory example

Here, through an example, we illustrate the basic concepts and usage of SWIM for sensitivity analysis. More advanced usage of SWIM and options for constructing stresses are demonstrated in Sections @ref(Sec:Scope) and @ref(Sec:CreditModel). We consider a simple portfolio model, with the portfolio loss defined by Y = Z₁ + Z₂ + Z₃. The random variables Z₁, Z₂, Z₃ represent normally distributed losses, with Z₁ ∼ N(100, 40²), Z₂ ∼ Z₃ ∼ N(100, 20²). Z₁ and Z₂ are correlated, while Z₃ is independent of (Z₁, Z₂). Our purpose in this example is to investigate how a stress on the loss Z₁ impacts on the overall portfolio loss Y. First we derive simulated data from the random vector (Z₁, Z₂, Z₃, Y), forming our baseline model.

set.seed(0)
# number of simulated scenarios
n.sim <- 10 ^ 5
# correlation between Z1 and Z2
r <- 0.5
# simulation of Z1  and Z2
# constructed as a combination of independent standard normals U1, U2
U1 <- rnorm(n.sim)
U2 <- rnorm(n.sim)
Z1 <- 100 + 40 * U1
Z2 <- 100 + 20 * (r * U1 + sqrt(1 - r ^ 2) * U2)
# simulation of Z3
Z3 <- rnorm(n.sim, 100, 20)
# portfolio loss Y
Y <- Z1 + Z2 + Z3
# data of baseline model
dat <- data.frame(Z1, Z2, Z3, Y)

Now we introduce a stress to our baseline model. For our first stress, we require that the mean of Z₁ is increased from 100 to 110. This is done using the stress function, which generates as output the SWIM object str.mean. This object stores the stressed model, i.e. the realisations of the model components and the scenario weights. In the function call, the argument k = 1 indicates that the stress is applied on the first column of dat, that is, on the realisations of the random variable Z₁.

library(SWIM)
str.mean <- stress(type = "mean", x = dat, k = 1, new_means = 110)

##   cols required_moment achieved_moment abs_error rel_error
## 1    1             110             110     1e-08   9.2e-11

summary(str.mean, base = TRUE)

## $base
##                   Z1       Z2       Z3        Y
## mean         1.0e+02  99.9404  99.9843 299.9811
## sd           4.0e+01  19.9969  19.9818  56.6386
## skewness    -6.1e-04   0.0012  -0.0025  -0.0023
## ex kurtosis -1.1e-02  -0.0089  -0.0125  -0.0093
## 1st Qu.      7.3e+01  86.4745  86.4816 261.6121
## Median       1.0e+02  99.9866 100.0091 300.0548
## 3rd Qu.      1.3e+02 113.3957 113.4934 338.2670
## 
## $`stress 1`
##                   Z1       Z2       Z3        Y
## mean        110.0000 102.4437  99.9828 312.4265
## sd           40.0331  19.9953  19.9761  56.6170
## skewness     -0.0024  -0.0015  -0.0049  -0.0037
## ex kurtosis  -0.0050  -0.0031  -0.0154  -0.0011
## 1st Qu.      82.9984  88.9771  86.4815 274.2200
## Median      110.0759 102.4810  99.9954 312.5039
## 3rd Qu.     136.9310 115.8744 113.5019 350.6120

The summary function, applied to the SWIM object str.mean, shows how the distributional characteristics of all random variables change from the baseline to the stressed model. In particular, we see that the mean of Z₁ changes to its required value, while the mean of Y also increases. Furthermore there is a small impact on Z₂, due to its positive correlation to Z₁.

Beyond considering the standard statistics evaluated via the summary function, stressed probability distributions can be plotted. In Figure @ref(fig:example1-cdfs-mean) we show the impact of the stress on the cumulative distribution functions (cdf) of Z₁ and Y. It is seen how the stressed cdfs are lower than the original (baseline) ones. Loosely speaking, this demonstrates that the stress has increased (in a stochastic sense) both random variables Z₁ and Y. While the stress was on Z₁, the impact on the distribution of the portfolio Y is clearly visible.

# refer to variable of interest by name...
plot_cdf(str.mean, xCol = "Z1", base = TRUE)
# ... or column number
plot_cdf(str.mean, xCol = 4, base = TRUE)

Baseline and stressed empirical distribution functions of model components Z₁ (left) and Y (right), subject to a stress on the mean of Z₁.

The scenario weights, given their central role, can be extracted from a SWIM object. In Figure @ref(fig:example1-weights-mean), the scenario weights from str.mean are plotted against realisations from Z₁ and Y respectively. It is seen how the weights are increasing in the realisations from Z₁. This is a consequence of the weights’ derivation via a stress on the model component Z₁. The increasingness shows that those scenarios for which Z₁ is largest are assigned a higher weight. The relation between scenario weights and Y is still increasing (reflecting that high outcomes of Y tend to receive higher weights), but no longer deterministic (showing that Y is not completely driven by changes in Z₁).

# parameter n specifies the number of scenario weights plotted
plot_weights(str.mean, xCol = "Z1", n = 1000)
# specifying the limits of the x-axis
plot_weights(str.mean, xCol = "Y", x_limits = c(90, 550), n = 1000)

Scenario weights against observations of model components Z₁ (left) and Y (right), subject to a stress on the mean of Z₁.

The stress to the mean of Z₁ did not impact the volatility of either Z₁ or Y, as can be seen by the practically unchanged standard deviations in the output of summary(str.mean). Thus, we introduce an alternative stress that keeps the mean of Z₁ fixed at 100, but increases its standard deviation from 40 to 50. This new stress is seen to impact the standard deviation of the portfolio loss Y.

str.sd <- stress(type = "mean sd", x = dat, k = 1, new_means = 100, new_sd = 50)

##   cols required_moment achieved_moment abs_error rel_error
## 1    1             100             100   3.0e-07   3.0e-09
## 2    1           12500           12500   5.2e-05   4.2e-09

summary(str.sd, base = FALSE)

## $`stress 1`
##                   Z1      Z2       Z3        Y
## mean        100.0000  99.941  99.9782 299.9187
## sd           50.0002  21.349  19.9799  67.9230
## skewness     -0.0027   0.007  -0.0034   0.0049
## ex kurtosis  -0.0556  -0.033  -0.0061  -0.0426
## 1st Qu.      66.0964  85.495  86.4822 253.7496
## Median      100.1290  99.974 100.0455 299.9766
## 3rd Qu.     133.7733 114.301 113.4701 345.9159

Furthermore, in Figure @ref(fig:example1-cdfs-sd), we compare the baseline and stressed cdfs of Z₁ and Y, under the new stress on Z₁. The crossing of probability distribution reflects the increase in volatility.

plot_cdf(str.sd, xCol = "Z1", base = TRUE)
plot_cdf(str.sd, xCol = 4, base = TRUE)

Baseline and stressed empirical distribution functions of model components Z₁ (left) and Y (right), subject to a stress on the standard deviation of Z₁.

The different way in which a stress on the standard deviation of Z₁ impacts on the model, compared to a stress on the mean, is reflected by the scenario weights. Figure @ref(fig:example1-weights-sd) shows the pattern of the scenario weights and how, when stressing standard deviations, higher weight is placed on scenarios where Z₁ is extreme, either much lower or much higher than its mean of 100.

plot_weights(str.sd, xCol = "Z1", n = 2000)
plot_weights(str.sd, xCol = "Y", n = 2000)

Scenario weights against observations of model components Z₁ (left) and Y (right), subject to a stress on the standard deviation of Z₁.

Finally we ought to note that not all stresses that one may wish to apply are feasible. Assume for example that we want to increase the mean of Z₁ from 100 to 300, which exceeds the maximum realisation of Z₁ in the baseline model. Then, clearly, no set of scenario weights can be found that produce a stress that yields the required mean for Z₁; consequently an error message is produced.

stress(type = "mean", x = dat, k = 1, new_means = 300)

## Error in stress_moment(x = x, f = means, k = as.list(k), m = new_means, :
Values in m must be in the range of f(x)

max(Z1)

## [1] 273

Stressing a model

We briefly introduce key concepts, using slightly more technical language compared to Section @ref(Sec:Intro). A model consists of a random vector of model components X = (X₁, …, X_d) and a probability measure; we denote the probability measure of a baseline model by P and that of a stressed model by P^W, where $W= \frac{dP^W}{dP}$, satisfying E(W) = 1 and W ≥ 0, is a Radon-Nikodym derivative. In a Monte Carlo simulation context, the probability space is discrete with n states Ω = {ω₁, …, ω_n}, each of which corresponds to a simulated scenario. To reconcile this formulation with the notation of Section @ref(Sec:Intro), we denote, for i = 1, …, n, j = 1, …, d, the realisations X_j(ω_i) := x_ji and W(ω_i) := w_i; the latter are the scenario weights. Under the baseline model, each scenario has the same probability P(ω_i) = 1/n, while under a stressed model it is P^W(ω_i) = W(ω_i)/n = w_i/n.

The stressed model thus arises from a change of measure from P to P^W, which entails the application of scenario weights w₁, …, w_n on individual simulations. SWIM calculates scenario weights such that model components fulfil specific stresses, while the distortion to the baseline model is as small as possible when measured by the Kullback-Leibler divergence (or relative entropy). Mathematically, a stressed model is derived by solving

In what follows, we denote by a superscript W quantities of interest under the stressed model, such as F^W, E^W for the probability distribution and expectation under the stressed model, respectively. We refer to Silvana M. Pesenti, Millossovich, and Tsanakas (2019) and references therein for further mathematical details and derivations of solutions to @ref(eq:optimisation).

Table @ref(tab:Rfunstress) provides a collection of all implemented types of stresses in the SWIM package. The precise constraints of @ref(eq:optimisation) are explained below.

The solutions to the optimisations @ref(eq:optimisationprob) and @ref(eq:optimisationVaR) are worked out fully analytically (Silvana M. Pesenti, Millossovich, and Tsanakas 2019), whereas problems @ref(eq:optimisationmean), @ref(eq:optimisationmeansd), @ref(eq:optimisationmoment) and @ref(eq:optimisationVaRES) require some root-finding. Specifically, problems @ref(eq:optimisationmean), @ref(eq:optimisationmeansd) and @ref(eq:optimisationmoment) rely on the package nleqslv, whereas @ref(eq:optimisationVaRES) uses the uniroot function.

Analysis of stressed models

Table @ref(tab:Ranalysis) provides a complete list of all implemented R functions in SWIM for analysing stressed models, which are described below in detail.

A credit risk portfolio

The credit model in this section is a conditionally binomial loan portfolio model including systematic and specific portfolio risk. We refer to the Appendix @ref(AppendixCM) for details about the model and the generation of the simulated data. A key variable of interest is the total aggregate portfolio loss L = L₁ + L₂ + L₃, where L₁, L₂, L₃ are homogeneous subportfolios on a comparable scale (say, thousands of $). The dataset contains 100,000 simulations of the portfolio L, the subportfolios L₁, L₂, L₃ as well as the random default probabilities within each subportfolio, H₁, H₂, H₃. These default probabilities represent the systematic risk within each subportfolio, while their dependence structure represents a systematic risk effect between the subportfolios. The simulated data of the credit risk portfolio are included in the SWIM package and can be accessed via data("credit_data"). A snippet of the dataset looks as follows:

data("credit_data")
head(credit_data)

##         L L1    L2  L3       H1      H2     H3
## [1,]  692  0 346.9 345 1.24e-04 0.00780 0.0294
## [2,] 1006 60 515.6 430 1.16e-03 0.01085 0.0316
## [3,] 1661  0 806.2 855 5.24e-04 0.01490 0.0662
## [4,] 1708  0 937.5 770 2.58e-04 0.02063 0.0646
## [5,]  807  0  46.9 760 8.06e-05 0.00128 0.0632
## [6,] 1159 20 393.8 745 2.73e-04 0.00934 0.0721

Stressing the portfolio loss

In this section, following a reverse sensitivity approach, we study the effects that stresses on (the tail of) the aggregate portfolio loss L have on the three subportfolios; thus assessing their comparative importance. First, we impose a 20% increase on the VaR at level 90% of the portfolio loss.

stress.credit <- stress(type = "VaR", x = credit_data, k = "L", alpha = 0.9, q_ratio = 1.2)

## Stressed VaR specified was 2174.25 , stressed VaR achieved is 2173.75

The 20% increase was specified by setting the q_ratio argument to 1.2 – alternatively the argument q can be set to the actual value of the stressed VaR.

Using the function VaR_stressed, we can quantify how tail quantiles of the aggregate portfolio loss change, when moving from the baseline to the stressed model. We observe that the increase in the VaR of the portfolio loss changes more broadly its tail quantiles; thus the stress on VaR also induces an increase in ES. The implemented functions VaR_stressed and ES_stressed calculate respectively VaR and ES; the argument alpha specifies the levels of VaR and ES, respectively, while the stressed model under which the risk measures are calculated can be chosen using wCol (by default equal to 1).

VaR_stressed(object = stress.credit, alpha =  c(0.75, 0.9, 0.95, 0.99), xCol = "L", wCol = 1, base = TRUE)

##        L base L
## 75% 1506   1399
## 90% 2174   1812
## 95% 2426   2085
## 99% 2997   2671

ES_stressed(object = stress.credit, alpha = 0.9, xCol = "L", wCol = 1, base = TRUE)

##        L base L
## 90% 2535   2191

As a second stress, we consider, additionally to the 20% increase in the VaR_0.9, an increase in ES_0.9 of the portfolio loss L. When stressing VaR and ES together via stress_VaR_ES, both VaR and ES need to be stressed at the same level, here alpha = 0.9. We observe that when stressing the VaR alone, ES increases to 2535. For the second stress we want a greater impact on ES, thus we require that the stressed ES be equal to 3500. This can be achieved by specifying the argument s, which is the stressed value of ES (rather than s_ratio, the proportional increase).

stress.credit <- stress(type = "VaR ES", x = stress.credit, k = "L", alpha = 0.9, q_ratio = 1.2, s = 3500)

## Stressed VaR specified was 2174.25 , stressed VaR achieved is 2173.75

When applying the stress function or one of its alternative versions to a SWIM object rather than to a data frame (via x = stress.credit in the example above), the result will be a new SWIM object with the new stress “appended” to existing stresses. This is convenient when large datasets are involved, as the stress function returns an object containing the original simulated data and the scenario weights. Note however, that this only works if the underlying data are exactly the same.

Analysing stressed models

The summary function provides a statistical summary of the stressed models. Choosing base = TRUE compares the stressed models with the the baseline model.

summary(stress.credit, base = TRUE)

## $base
##                    L    L1     L2      L3       H1      H2     H3
## mean        1102.914 19.96 454.04 628.912 0.000401 0.00968 0.0503
## sd           526.536 28.19 310.99 319.713 0.000400 0.00649 0.0252
## skewness       0.942  2.10   1.31   0.945 1.969569 1.30836 0.9501
## ex kurtosis    1.326  6.21   2.52   1.256 5.616080 2.49803 1.2709
## 1st Qu.      718.750  0.00 225.00 395.000 0.000115 0.00490 0.0318
## Median      1020.625  0.00 384.38 580.000 0.000279 0.00829 0.0464
## 3rd Qu.     1398.750 20.00 609.38 810.000 0.000555 0.01296 0.0643
## 
## $`stress 1`
##                   L    L1     L2     L3       H1      H2     H3
## mean        1193.39 20.83 501.10 671.46 0.000417 0.01066 0.0536
## sd           623.48 29.09 363.57 361.21 0.000415 0.00756 0.0285
## skewness       1.01  2.09   1.36   1.02 1.973367 1.35077 1.0284
## ex kurtosis    0.94  6.14   2.24   1.22 5.630325 2.23363 1.2383
## 1st Qu.      739.38  0.00 234.38 405.00 0.000120 0.00512 0.0328
## Median      1065.62 20.00 412.50 605.00 0.000290 0.00878 0.0483
## 3rd Qu.     1505.62 40.00 675.00 865.00 0.000578 0.01422 0.0688
## 
## $`stress 2`
##                   L    L1     L2     L3       H1      H2     H3
## mean        1289.90 21.70 558.27 709.93 0.000437 0.01180 0.0566
## sd           875.89 30.57 507.78 447.30 0.000448 0.01045 0.0351
## skewness       1.90  2.17   2.10   1.57 2.090457 2.10131 1.5385
## ex kurtosis    3.67  6.74   4.79   2.80 6.203613 4.97016 2.6143
## 1st Qu.      739.38  0.00 234.38 405.00 0.000123 0.00512 0.0328
## Median      1065.62 20.00 412.50 605.00 0.000297 0.00879 0.0484
## 3rd Qu.     1505.62 40.00 675.00 875.00 0.000594 0.01439 0.0697

Note that stress 1 is the summary output corresponding to the 20% increase in the VaR, while stress 2 corresponds to the stress in both VaR and ES. The information on individual stresses can be recovered through the get_specs function, and the actual scenario weights using get_weights. Since the SWIM object stress.credit contains two stresses, the scenario weights that are returned by get_weights form a data frame consisting of two columns, corresponding to stress 1 and to stress 2, respectively.

get_specs(stress.credit)

##            type k alpha       q    s
## stress 1    VaR L   0.9 2173.75 <NA>
## stress 2 VaR ES L   0.9 2173.75 3500

Next, we produce a scatter plot of the scenario weights against the portfolio loss L. As the number of scenario weights is large, we only 5000 data points. This can be achieved via the parameter n in the function plot_weights, that has a default of n = 5000.

plot_weights(stress.credit, xCol = "L", wCol = 1, n = 2000)
# parameter `wCol` specifies the stresses, whose scenario weights are plotted.
plot_weights(stress.credit, xCol = "L", wCol = 2, n = 7000)

Scenario weights against the portfolio loss L for stressing VaR (left) and stressing both VaR and ES (right).

It is seen in Figure @ref(fig:credit-weights) that the weights generated to stress VaR, and VaR and ES together, follow different patterns to the weights used to stress means and standard deviations, as shown in Section @ref(Sec:Intro). Recall that SWIM calculates the scenario weights such that under the stressed model the given constraints are fulfilled. Thus, an increase in the VaR and/or ES of the portfolio loss L results in large positive realisations of L being assigned higher weight. On the other hand, when the standard deviation is stressed, scenario weights are calculated that inflate the probabilities of both large positive and negative values.

Visualising stressed distributions

The change in the distributions of the portfolio and subportfolio losses, when moving from the baseline to the stressed models, can be visualised through the functions plot_hist and plot_cdf. The following figure displays the histogram of the aggregate portfolio loss under the baseline and the two stressed models. It is seen how stressing VaR and ES has a higher impact on the right tail of L, compared to stressing VaR only. This is consistent with the tail-sensitive nature of the Expected Shortfall risk measure (McNeil, Frey, and Embrechts 2015).

plot_hist(object = stress.credit, xCol = "L", base = TRUE)

Histogram of the portfolio loss L under the baseline and the two stressed models.

The arguments xCol and wCol (with default to plot all stresses) define the columns of the data and the columns of the scenario weights, respectively, that are used for plotting. Next, we analyse the impact that stressing the aggregate loss L has on the subportfolios L₁, L₂ L₃. Again, we use the function plot_hist and plot_cdf for visual comparison, but this time placing the distribution plots and histograms of subportfolio losses along each other via the function ggarrange (from the package ggpubr). The plots obtained from plot_hist and plot_cdf can be further personalised when specifying the argument displ = FALSE, as then the graphical functions plot_hist and plot_cdf return data frames compatible with the package ggplot2.

pL1.cdf <- plot_cdf(object = stress.credit, xCol = 2, wCol = "all", base = TRUE)
pL2.cdf <- plot_cdf(object = stress.credit, xCol = 3, wCol = "all", base = TRUE)
pL3.cdf <- plot_cdf(object = stress.credit, xCol = 4, wCol = "all", base = TRUE)

pL1.hist <- plot_hist(object = stress.credit, xCol = 2, wCol = "all", base = TRUE)
pL2.hist <- plot_hist(object = stress.credit, xCol = 3, wCol = "all", base = TRUE)
pL3.hist <- plot_hist(object = stress.credit, xCol = 4, wCol = "all", base = TRUE)

ggarrange(pL1.cdf, pL1.hist, pL2.cdf, pL2.hist, pL3.cdf, pL3.hist, ncol = 2, nrow = 3,
          common.legend = TRUE)

Distribution functions and histograms of the subportfolios L₁, L₂, L₃ for the stresses on the VaR (stress 1) and on both the VaR and ES (stress 2) of the portfolio loss L.

It is seen from both the distribution plots and the histograms in Figures @ref(fig:CM-histL) and @ref(fig:CM-plot1) that the stresses have no substantial impact on L₁, while L₂ and L₃ are more affected, indicating a higher sensitivity. The higher impact on the tails of stress 2 (on both VaR and ES) is also visible. Sensitivity measures quantifying these effects are introduced in the following subsection.

Sensitivity measures

The impact of the stressed models on the model components can be quantified through sensitivity measures. The function sensitivity includes the Kolmogorov distance, the Wasserstein distance, and the sensitivity measure Gamma; the choice of measure is by the argument type. We refer to Section @ref(Sec:analysis) for the definitions of those sensitivity measures. The Kolmogorov distance is useful for comparing different stressed models. Calculating the Kolmogorov distance, we observe that stress 2 produces a larger Kolmogorov distance compared to stress 1, which reflects the additional stress on the ES for the stressed model stress 2.

sensitivity(object = stress.credit, xCol = 1, wCol = "all", type = "Kolmogorov")

##     stress       type      L
## 1 stress 1 Kolmogorov 0.0607
## 2 stress 2 Kolmogorov 0.0748

We now rank the sensitivities of model components by the measure Gamma, for each stressed model. Consistently with what the distribution plots showed, L₂ is the most sensitive subportfolio, followed by L₃ and L₁. The respective default probabilities H₁, H₂, H₃ are similarly ranked.

sensitivity(object = stress.credit, xCol = c(2 : 7), wCol = "all", type = "Gamma")

##     stress  type    L1    L2    L3    H1    H2    H3
## 1 stress 1 Gamma 0.150 0.819 0.772 0.196 0.811 0.767
## 2 stress 2 Gamma 0.113 0.734 0.639 0.171 0.708 0.636

Using the sensitivity function we can analyse whether the sensitivity of the joint subportfolio L₁ + L₃ exceeds the sensitivity of the (most sensitive) subportfolio L₂. This can be accomplished by specifying, through the argument f, a list of functions applicable to the columns k of the dataset. By setting xCol = NULL only the transformed data is considered. The sensitivity measure of functions of columns of the data is particularly useful when high dimensional models are considered, providing a way to compare the sensitivity of blocks of model components.

sensitivity(object = stress.credit, type = "Gamma", f = sum, k = c(2, 4), wCol = 1, xCol = NULL)

##     stress  type    f1
## 1 stress 1 Gamma 0.783

We observe that the sensitivity of L₁ + L₃ is larger than the sensitivity to either L₁ and L₃, reflecting the positive dependence structure of the credit risk portfolio. Nonetheless, subportfolio L₂ has not only the largest sensitivity compared to L₁ and L₃ but also a higher sensitivity than the combined subportfolios L₁ + L₃.

The importance_rank function, having the same structure as the sensitivity function, returns the ranks of the sensitivity measures. This function is particularly useful when several risk factors are involved.

importance_rank(object = stress.credit, xCol = c(2 : 7), wCol = 1, type = "Gamma")

##     stress  type L1 L2 L3 H1 H2 H3
## 1 stress 1 Gamma  6  1  3  5  2  4

Constructing more advanced stresses

# 90% VaR of L under the baseline model
VaR.L <- quantile(x = credit_data[, "L"], prob = 0.9, type = 1) 
# 75th quantile of H2 under the baseline model
q.H2 <- quantile(x = credit_data[, "H2"], prob = 0.75, type = 1) 
# columns to be stressed (L, H2, H2)
k.stressH2 = list(1, 6, 6) 
# functions to be applied to columns
f.stressH2 <- list(
                 # indicator function for L, for stress on VaR
                 function(x)1 * (x <= VaR.L * 1.2), 
                 # mean of H2
                 function(x)x, 
                 # indicator function for 75th quaantile of H2
                 function(x)1 * (x <= q.H2)) 
# new values for the 90% VaR of L, mean of H2, 75th quantile of H2
m.stressH2 = c(0.9, mean(credit_data[, "H2"]), 0.75) 
stress.credit <- stress_moment(x = stress.credit, f = f.stressH2, k = k.stressH2, 
                               m = m.stressH2)

##   cols required_moment achieved_moment abs_error rel_error
## 1    1         0.90000         0.90000  7.62e-09  8.47e-09
## 2    6         0.00968         0.00968  7.59e-11  7.84e-09
## 3    6         0.75000         0.75000  3.80e-10  5.06e-10

summary(stress.credit, wCol = 3, xCol = 6, base = TRUE)

## $base
##                  H2
## mean        0.00968
## sd          0.00649
## skewness    1.30836
## ex kurtosis 2.49803
## 1st Qu.     0.00490
## Median      0.00829
## 3rd Qu.     0.01296
## 
## $`stress 3`
##                  H2
## mean        0.00968
## sd          0.00706
## skewness    1.39137
## ex kurtosis 2.26516
## 1st Qu.     0.00453
## Median      0.00786
## 3rd Qu.     0.01296

sensitivity(object = stress.credit, xCol = c(2:4), type = "Gamma")

##     stress  type     L1     L2    L3
## 1 stress 1 Gamma 0.1501 0.8195 0.772
## 2 stress 2 Gamma 0.1131 0.7336 0.639
## 3 stress 3 Gamma 0.0102 0.0203 0.366

Stressing tails of subportfolios

Up to now, we have considered the impact of stressing the aggregate portfolio loss on subportfolios. Now, following a forward sensitivity approach, we consider the opposite situation: stressing the subportfolio losses and monitoring the impact on the aggregate portfolio loss L. First, we impose a stress requiring a simultaneous 20% increase in the 90^th quantile of the losses in subportfolios L₂ and L₃. Note that the function stress_VaR (and stress_VaR_ES) allow to stress the VaR and/or the ES of only one model component. Thus, to induce a stress on the 90^th quantiles of L₂ and L₃, we use the function stress_moments and interpret the quantile constraints as moment constraints, via E(1_{L2 ≤ VaR^W(L2)}) and E(1_{L3 ≤ VaR^W(L3)}), respectively, where VaR^W = VaR ⋅ 1.2 denotes the VaRs in the stressed model.

# VaR of L2 and L3, respectively
VaR.L2 <- quantile(x = credit_data[, "L2"], prob = 0.9, type = 1) 
VaR.L3 <- quantile(x = credit_data[, "L3"], prob = 0.9, type = 1) 
#stressing VaR of L2 and L3
f.stress <- list(function(x)1 * (x <= VaR.L2 * 1.2), 
                 function(x)1 * (x <= VaR.L3 * 1.2)) 
stress.credit.L2L3 <- stress_moment(x = credit_data, f = f.stress, k = list(3, 4), 
                                    m = c(0.9, 0.9))

##   cols required_moment achieved_moment abs_error rel_error
## 1    3             0.9             0.9 -1.13e-09 -1.26e-09
## 2    4             0.9             0.9 -1.70e-10 -1.89e-10

#impact on portfolio tail
VaR_stressed(stress.credit.L2L3, alpha = c(0.75, 0.9, 0.95, 0.99), xCol = "L", 
             base = TRUE)

##        L base L
## 75% 1556   1399
## 90% 2086   1812
## 95% 2423   2085
## 99% 3072   2671

It is seen how the stressing of subportfolios L₂ and L₃ has a substantial impact on the portfolio loss. Given the importance of dependence for the distribution of the aggregate loss of the portfolio, we strengthen this stress further, by additionally requiring that the frequency of joint high losses from L₂ and L₃ is increased. Specifically, we require the joint exceedance probability to be P^W(L₂ > VaR^W(L₂), L₃ > VaR^W(L₃)) = 0.06, which is almost doubling the corresponding probability in the last stressed model, which was equal to 0.0308.

# probability of joint exceendance under the baseline model
mean(1 * (credit_data[, "L2"] > VaR.L2 * 1.2) * (credit_data[, "L3"] > VaR.L3 * 1.2))

## [1] 0.00865

# probability of joint exceendance under the stressed model
mean(get_weights(stress.credit.L2L3) * (credit_data[, "L2"] > VaR.L2 * 1.2) * (credit_data[, "L3"] > VaR.L3 * 1.2))

## [1] 0.0308

# additionally stress joint exceedance probability of L2 and L3
f.stress.joint <- c(f.stress, function(x)1 * (x[1] > VaR.L2 * 1.2) * (x[2] > VaR.L3 * 1.2)) 
stress.credit.L2L3 <- stress_moment(x = stress.credit.L2L3, f = f.stress.joint, k = list(3,4,c(3, 4)), m = c(0.9, 0.9, 0.06))

##      cols required_moment achieved_moment abs_error rel_error
## 1       3            0.90            0.90  1.96e-11  2.18e-11
## 2       4            0.90            0.90  2.72e-11  3.03e-11
## 3 c(3, 4)            0.06            0.06 -2.55e-10 -4.26e-09

We analyse the impact the stresses of the tail of the subportfolios L₂ and L₃ have on the aggregate portfolio L. For this, we plot in Figure @ref(fig:CM-joint-stress-VaR-2-effect) the quantile of the aggregate portfolio under the baseline model (blue), under the stress on the tail of L₂ and L₃ (red), and under the additional stress on the joint tail of L₂ and L₃ (green).

plot_quantile(stress.credit.L2L3, xCol = "L", wCol = "all", base = TRUE, x_limits = c(0.75, 1))

Quantiles of the aggregate loss L under the baseline (blue), the stress on the tails of L₂ and L₃ (red), and the additional stress on the joint tail of L₂ and L₃ (green).

The results and the plots indicate that the additional stress on joint exceedances of subportfolios, increases the tail quantiles of L even further, demonstrating the importance of (tail-)dependence in portfolio risk management.

Credit Model assumptions

The credit risk portfolio of Section @ref(Sec:CreditModel) is based on the conditionally binomial credit model described in Section 11.2 of McNeil, Frey, and Embrechts (2015) which belongs to the family of mixture models. Specifically, we consider a portfolio that consists of three homogeneous subportfolios and denote the aggregate portfolio loss by L = L₁ + L₂ + L₃, with L₁, L₂, L₃ the losses of each subportfolio, given by where e_i and M_i are the exposure and number of defaults of the i^th subportfolio, respectively, and LGD_i is the loss given default of subportfolio i. M_i is Binomially distributed, conditional on H_i, a random common default probability. Specifically M_i|H_i ∼ Binomial(m_i, H_i), where m_i is the portfolio size. The H_is follow a Beta distributions with parameters chosen so as to match given overall unconditional default probabilities p_i and default correlations ρ_i, that is, the correlation between (the indicators of) two default events within a subportfolio, see McNeil, Frey, and Embrechts (2015). The dependence structure of (H₁, H₂, H₃) is modelled via a Gaussian copula with correlation matrix

Table @ref(tab:credit) summarises the parameter values used in the simulation.

Code for generating the data

  set.seed(1)
  library(copula)
  nsim <- 100000
  
# counterparties subportfolio 1, 2 and 3
  m1 <- 2500
  m2 <- 5000
  m3 <- 2500
  
  # prob of default for subportfolios 1, 2 and 3
  p1 <- 0.0004 
  p2 <- 0.0097 
  p3 <- 0.0503  
  
  # correlation between default probabilities
  rho1 <- 0.0004
  rho2 <- 0.0044
  rho3 <- 0.01328
  
# exposures
  e1 <- 80
  e2 <- 25 
  e3 <- 10 
  
# loss given default
  LGD1 <- 0.25
  LGD2 <- 0.375
  LGD3 <- 0.5
  
# beta parameters: matching subportfolios default probabilities and correlation
  alpha1 <- p1 * (1 / rho1 - 1)
  beta1 <- alpha1 * (1 / p1 - 1)
  
  alpha2 <- p2 * (1 / rho2 - 1)
  beta2 <- alpha2 * (1 / p2 - 1)
  
  alpha3 <- p3 * (1 / rho3 - 1)
  beta3 <- alpha3 * (1 / p3 - 1)
  
# correlations between subportfolios
  cor12 <- 0.3
  cor13 <- 0.1
  cor23 <- 0.4
  
# Gaussian copula structure
  myCop <- normalCopula(param = c(cor12, cor13, cor23), dim = 3, dispstr = "un")
  
# multivariate beta with given copula
  myMvd <- copula::mvdc(copula = myCop,
                margins = c("beta", "beta", "beta"),
                paramMargins = list(list(alpha1, beta1),
                                    list(alpha2, beta2),
                                    list(alpha3, beta3)))

# simulation from the chosen copula
  H <- copula::rMvdc(nsim, myMvd)
  
# simulate number of default per subportfolios (binomial distributions)
  M1 <- rbinom(n = nsim, size = m1, prob = H[, 1])
  M2 <- rbinom(n = nsim, size = m2, prob = H[, 2])
  M3 <- rbinom(n = nsim, size = m3, prob = H[, 3])
  
# total loss per subportfolio
  L1 <- M1 * e1 * LGD1
  L2 <- M2 * e2 * LGD2
  L3 <- M3 * e3 * LGD3
  
# aggregate portfolio loss
  L <- L1 + L2 + L3
  
# the credit data included in SWIM
  credit_data <- cbind(L, L1, L2, L3, H)
  colnames(credit_data) <- c("L", "L1", "L2", "L3", "H1", "H2", "H3")