Problem-driven scenario generation: an analytical approach for stochastic programs with tail risk measure

Stochastic programming is a tool for making decisions under uncertainty. Under this modeling paradigm, uncertain parameters are modeled as a random vector, and one attempts to minimize (or maximize) the expectation or risk measure of some loss function which depends on the initial decision. However, what distinguishes stochastic programming from other stochastic modeling approaches is its ability to explicitly model future decisions based on outcomes of stochastic parameters and initial decisions, and the associated costs of these future decisions. The power and flexibility of the stochastic programming approach comes at a price: stochastic programs are usually analytically intractable, and often not susceptible to solution techniques for deterministic programs.

Typically, a stochastic program can only be solved when it is scenario-based, that is when the random vector for the problem has a finite discrete distribution. For example, stochastic linear programs become large-scale linear programs when the underlying random vector is discrete. In the stochastic programming literature, the mass points of this random vector are referred to as scenarios, the discrete distribution as the scenario set and the construction of this set as scenario generation. Scenario generation can consist of discretizing a continuous probability distribution, or directly modeling the uncertain quantities as discrete random variables. The more scenarios in a set, the more computational power that is required to solve the problem. The key issue of scenario generation is therefore how to represent the uncertainty to ensure that the solution to the problem is reliable, while keeping the number of scenarios low so that the problem is computationally tractable.

A common approach to scenario generation is to fit a statistical model to the uncertain problem parameters and then generate a random sample from this for the scenario set. This has desirable asymptotic properties [22, 33], but may require large sample sizes to ensure the reliability of the solutions it yields. This can be mitigated somewhat by using variance reduction techniques such as stratified sampling and importance sampling [24]. Sampling also has the advantage that it can be used to construct confidence intervals on the true solution value [25]. Another approach is to construct a scenario set whose distance from the true distribution, with respect to some probability metric, is small [12, 19, 28]. These approaches tend to yield better and much more stable solutions to stochastic programs than does sampling.

A characteristic of these approaches to scenario generation is that they are distribution-driven; that is, they only aim to approximate a distribution and are divorced from the stochastic program for which they are producing scenarios. By exploiting the structure of a problem, it may be possible to find a more parsimonious representation of the uncertainty. Note that such a problem-driven approach may not yield a discrete distribution which is close to the true distribution in a probabilistic sense; the aim is only to find a discrete distribution which yields a high quality solution to our problem.

Stochastic programs often have the objective of minimizing the expectation of a loss function. This is particularly appropriate when the initial decision represents a strategic decision that is going to be used again and again, and individual large losses do not matter in the long term. For example, in a stochastic facility location problem (e.g. see [5]) the locations of several facilities must be chosen subject to the unknown demands of customers in a way which minimizes fixed investment costs, and future distribution costs. In other cases, the decision may be only used once or a few times, and the occurrence of large losses may have serious consequences such as bankruptcy. This is characteristic of the portfolio selection problem [26] studied in detail in the latter part of this paper. In this latter case, minimizing the expectation alone is not appropriate as this does not necessarily mitigate against the possibility of large losses. One possible remedy is to use a risk measure which penalises in some way the likelihood and severity of potential large losses.

In this paper we are interested in stochastic programs which use tail risk measures. A precise definition of a tail-risk measure will be given in Sect. 3 but for now, one can think of a tail risk measure as a function of a random variable which only depends on the upper tail of its distribution function. Tail risk measures are useful as they summarize the extent of potential losses in the worst possible outcomes. Examples of tail risk measures include the Value-at-Risk (VaR) [21] and the Conditional Value-at-Risk (CVaR) [29], both of which are commonly used in financial contexts. Although the methodology developed in this paper can in principle be applied to any loss function, in this work we are mainly interested in loss functions which arise in one and two-stage stochastic programs.

Distribution-driven scenario generation methods are particularly problematic for stochastic programs involving tail risk measures. This is because these methods tend to spread their scenarios evenly across the support of distribution and so struggle to adequately represent the tail risk without using a potentially prohibitively large number of scenarios.

In this paper, we propose an analytic problem-driven approach to scenario generation applicable to stochastic programs which use tail risk measures of a form made precise in Sect. 3. We observe that the value of a tail risk measure depends only on scenarios confined to an area of the distribution that we call the risk region. This means that all scenarios that are not in the risk region can be aggregated into a single point. By concentrating scenarios in the risk region, we can calculate the value of a tail risk measure more accurately.

Given a risk region for a problem, we propose a simple algorithm for generating scenarios which we call aggregation sampling. This algorithm takes samples from the random vector until a specified number of samples in the risk region have been produced, and all other scenarios are aggregated into a single scenario. We provide and give proofs of conditions under which this method is asymptotically consistent with standard Monte Carlo sampling.

In general, finding a risk region is difficult as it is determined by the loss function, problem constraints and the distribution of the uncertain parameters. Therefore, we derive risk regions for two classes of problem as a proof-of-concept of our methodology. The first class of problems are those with monotonic loss functions which, as will be shown, occur naturally in the context of network design. The second class are portfolio selection problems. For both types of risk regions we run numerical tests which demonstrate that our methodology yields better quality solutions and with greater reliability than standard Monte Carlo sampling.

This paper is organized as follows: in Sect. 2 we discuss related work; in Sect. 3 we define tail risk measures and their associated risk regions; in Sect. 4 we discuss how these risk regions can be exploited for the purposes of scenario generation; in Sect. 5 we prove that our scenario generation method is consistent with standard Monte Carlo sampling; in Sects. 6 and 7 we derive risk regions for the two classes of problems described above; in Sect. 8 we present numerical tests; finally in Sect. 9 we summarize our results and make some concluding remarks.

Notation Throughout this paper random variables and vectors are represented by bold (mainly Greek) letters: \(\varvec{\theta },\ \varvec{\xi },\ \varvec{\zeta }\) and outcomes of these are represented by the corresponding non-bold letters: \(\theta ,\ \xi ,\ \zeta \). Inequalities used with vectors and matrices always apply component-wise. \(\left\| \cdot \right\| \) represents the standard Euclidean norm.

There are relatively few cases of problem-driven scenario generation in the literature. The earliest example of which we are aware is the importance sampling approach of [8] which constructs a sampler from the loss function. Importance sampling has been used more recently for scenario generation for problems which, like our own, concern rare events. In [23] an importance sampling scheme is used for a multistage problem involving the CVaR risk measure. In [4], an importance sampling approach is proposed for chance-constrained stochastic programs where the permitted probabilities of constraint violation are very small.

There is an interesting connection between problem-driven scenario generation and distributionally robust optimization [11, 38, 39]. In distributionally robust optimization, the distribution of the random variables in a stochastic program is itself uncertain, and one must optimize for the worst-case distribution. Solving a distributionally robust optimization problem thus involves finding, at least implicitly, the worst-case distribution or scenario set for given objective and constraints. In this sense, distributionally robust optimization could be considered as a problem-driven scenario generation method. Of particular relevance for this work, the paper [9] solves a distributionally robust portfolio selection problem involving the CVaR risk measure where the distribution of asset returns has specified discrete marginals, but unknown joint distribution.

The idea that in stochastic programs with tail risk measures some scenarios do not contribute to the calculation of the tail-risk measure was also exploited in [17]. However, they propose a solution algorithm rather than a method of scenario generation. Their approach is to iteratively solve the problem with a subset of scenarios, identify the scenarios which have loss in the tail, update their scenario set appropriately and resolve, until the true solution has been found. Their method has the benefit that it is not distribution dependent. On the other hand, their method works for only the \(\beta {\text {-CVaR}}\) risk measure, while our approach works in principle for any tail risk measure.

In this section we present the core theory to our scenario generation methodology. Specifically, in Sect. 3.1 we formally define tail-risk measures of random variables and in Sect. 3.2 we define risk regions and present some key results related to these.

In the previous section, we showed that under mild conditions the value of a tail risk measure only depends on the distribution of outcomes in the risk region. In this section we demonstrate how this feature may be exploited for the purposes of scenario generation.

We assume throughout this section that our scenario sets are constructed from some underlying probabilistic model from which we can draw independent identically distributed samples. We also assume we have a set \({\mathcal {R}}_{{\mathcal {X}}}\subseteq {\mathcal {R}}\subset \varXi \) which satisfies the aggregation condition for the problem under consideration, and for which we can easily test membership. The set \({\mathcal {R}}\) may be an exact risk region, that is \({\mathcal {R}}={\mathcal {R}}_{{\mathcal {X}}}\), or it could a conservative risk region, that is \({\mathcal {R}}\supset {\mathcal {R}}_{{\mathcal {X}}}\). To avoid repeating cumbersome terminology, we simply refer to \({\mathcal {R}}\) as a risk region, differentiating between the conservative and exact cases only where necessary. The complement \({\mathcal {R}}^{c}\) will be referred to as the aggregation region for reasons which will become clear. Our general approach to scenario generation is to prioritize the construction of scenarios in the risk region \({\mathcal {R}}\).

In Sect. 4.1 we present and analyse a scenario generation method which we call aggregation sampling. In Sect. 4.2 we briefly discuss alternative ways of exploiting risk regions for scenario generation.

4.1 Aggregation sampling

In aggregation sampling the user specifies a number of risk scenarios, that is, the number of scenarios to represent the risk region. The algorithm then draws samples from the distribution, storing those samples which lie in the risk region and aggregating those in the aggregation region into a single point. In particular, the samples in the aggregation region are aggregated into their mean. The algorithm terminates when the specified number of risk scenarios has been reached. This is detailed in Algorithm 1.

Open image in new window

Aggregation sampling can be thought of as equivalent to sampling from the aggregated random vector from Definition 6 for large sample sizes. Aggregation sampling is thus consistent with standard Monte Carlo sampling only if \({\mathcal {R}}\) satisfies the aggregation condition and \({{\mathbb {E}}_{} \left[ \varvec{\xi }|\varvec{\xi }\in {\mathcal {R}}^{c} \right] \in {\mathcal {R}}^{c}}\). In Sect. 5, we provide conditions under which we can prove consistency. Note that it is possible that the algorithm could terminate without sampling any scenario in the aggregation region. This could happen in cases where \({\mathbb {P}}\left( \varvec{\xi }\in {\mathcal {R}}^{c}\right) \) is very small, and the number of specified risk scenarios n is relatively small. In this case, to ensure that the algorithm terminates in a reasonable amount of time and that the scenario set which the algorithm outputs always has a consistent number of scenarios, we sample an arbitrary scenario in place of a scenario representing the aggregated scenarios. This situation is irrelevant for the asymptotic analysis of the algorithm.

We now study the performance of our aggregation sampling algorithm. Let \(a={\mathbb {P}}\left( \varvec{\xi }\in {\mathcal {R}}^{c}\right) \) be the probability of the aggregation region, and n the desired number of risk scenarios. Let N(n) denote the effective sample size for aggregation sampling, that is, the number of samples drawn until the algorithm terminates.¹ The aggregation sampling algorithm can be viewed as a sequence of Bernoulli trials where a trial is a success if the corresponding sample lies in the aggregation region, and which terminates once we have reached n failures, that is, once we have sampled n scenarios from the risk region. We can therefore write down the distribution of N(n):

$$\begin{aligned} N(n) \sim n + \mathcal {NB}(n, a), \end{aligned}$$

where \(\mathcal {NB}(n,a)\) denotes a negative binomial random variable whose probability mass function is as follows:

$$\begin{aligned} \left( {\begin{array}{c}k+n-1\\ k\end{array}}\right) (1-a)^{n}a^{k},\qquad k\ge 0. \end{aligned}$$

The expected effective sample size of aggregation sampling is thus:

$$\begin{aligned} {\mathbb {E}}_{} \left[ N(n) \right] = n + n \frac{a}{1-a}. \end{aligned}$$

(12)

The expected effective sample size N(n) can be thought of as the required sample size to construct a scenario set via Monte Carlo sampling with n scenarios in the risk region \({\mathcal {R}}\). Thus, the greater the expected effective sample size, the greater the benefit of using aggregation sampling over standard Monte Carlo sampling. From (12) we can see that the expected effective sample size increases as the probability a of the aggregation region increases. Therefore, when constructing a risk region \({\mathcal {R}}\supseteq {\mathcal {R}}_{{\mathcal {X}}}\) for the purposes of scenario generation, it is important that \({\mathcal {R}}\) is as tight an approximation of the exact risk region \({\mathcal {R}}_{{\mathcal {X}}}\) as possible in order that \(a = {\mathbb {P}}\left( \varvec{\xi }\in {\mathcal {R}}^{c}\right) \) is as large as possible. Also, the fact that the advantage of using aggregation sampling over standard Monte Carlo sample improves as the probability of the risk region increases, also tells us that this methodology will potentially work better for problems with higher values of \(\beta \) and which are more constrained due to the relations (5) and (6).

The reason that aggregation sampling and aggregation reduction work is that, for large sample sizes, they are equivalent to sampling from the aggregated random vector, and if the aggregation condition holds then the aggregated random vector yields the same optimization problem as the original random vector. We only prove consistency for aggregation sampling and not aggregation reduction as the proofs are very similar. Essentially, the only difference is that aggregation sampling has the additional complication of terminating after a random number of samples.

We suppose in this section that we have a sequence of independently identically distributed (i.i.d.) random vectors \(\varvec{\xi }_{1}, \varvec{\xi }_{2}, \ldots \) with the same distribution as \(\varvec{\xi }\), and which are defined on the product probability space \(\varOmega ^{\infty }\).

In order to use risk regions for scenario generation, we need to have a characterization of the risk region which conveniently allows us to test membership. In general this is a difficult as the risk region depends on the loss function, the distribution and the problem constraints. Therefore, as a proof-of-concept, in the following two sections we derive risk regions for two classes of problems. In this section we propose a conservative risk region for problems which have monotonic loss functions.

Monotonic loss functions occur naturally in stochastic linear programming. The following result presents a class of loss functions which arise in the context of network design, and gives conditions under which they are monotonic.

This recourse function arises in stochastic network design, and the problem formulation in the previous proposition was adapted from a model in the paper [31]. In this type of problem, we have a network consisting of suppliers, processing units, and customers, and decisions must be made relating to opening facilities and the capacities of nodes and arcs. The problem which defines the recourse function \(Q(x, \xi )\) depends on the capacity and opening decisions x of the first stage, and the demand of the customers \(\xi \). The aim of the problem is construct of flow of products y which minimize transportation costs for satisfying customers demand, plus penalties for any unsatisfied demand z.

For a problem with a monotonic loss function, the following result defines a conservative risk region.

In this section, we characterize exactly the risk region of the portfolio selection problem when the distribution of asset returns belongs to a certain class of distributions.

In the portfolio selection problem one aims to choose a portfolio of financial assets with uncertain returns. For \(i = 1,\ldots , d\), let \(x_{i}\) denote the amount to invest in asset i, and \(\varvec{\xi }_{i}\) the random return of asset i. The loss function in this problem is the negative total return, that is \(f(x, \varvec{\xi }) = \sum _{i=1}^d -x_i \varvec{\xi }_i = -x^T \varvec{\xi }\), and \(\varXi = {\mathbb {R}}^d\). The set \({\mathcal {X}}\subset {\mathbb {R}}^d\) of feasible portfolios may encompass constraints like no short-selling (\(x \ge 0\)), total investment (\(\sum _{i=1}^{d} x_{i} = 1\)) and quotas on certain stocks (\(x \le c\)).

The following corollary gives sufficient conditions for the risk region to satisfy the aggregation condition, and for aggregation sampling to be consistent.

Elliptical distributions are a general class of distributions which include among others the multivariate Normal and multivariate t-distributions. See [16] for a full overview of the subject.

By a slight abuse of notation, for a set \({\mathcal {A}}\subset {\mathbb {R}}^d\) and a matrix \(T\in {\mathbb {R}}^{d\times d}\), we write \({T\left( {\mathcal {A}}\right) := \{ T\xi : \xi \in {\mathcal {A}}\}}\). Finally, recall that the conic hull of a set \({\mathcal {A}}\subset {\mathbb {R}}^d\), which we denote \({\text {conic}}\left( {\mathcal {A}} \right) \), is the smallest convex cone containing \({\mathcal {A}}\).

In this section, we test the performance of the methodology developed in this paper. For the portfolio selection problem, when \({\mathcal {X}}\subseteq {\mathbb {R}}_{+}^{d}{\setminus }\{0\}\) the loss function \(f(x,\xi ) = -x^{T}\xi \) is monotonic decreasing. We therefore use this problem throughout this section to test both the conservative risk region presented in Sect. 6, and the exact risk region presented in Sect. 7.

In order to test whether a point belongs to the exact non-risk region in (41) requires the projection of a point onto a convex cone. This can be done by solving a small linear complementarity problem. See [36] or our follow-up paper [15] for more details. We solve linear complementarity problems using code from the Siconos numerics library [1]. To test whether a point \(\xi \in \varXi \) belongs to the conservative risk region in (38), involves the evaluation of the probability \({\mathbb {P}}\left( \varvec{\xi }< \xi \right) \). Since calculating this probability exactly involves evaluating a multidimensional integral we approximate the probability by taking a large sample from \(\varvec{\xi }\), and using the empirical distribution function of this sample. Repeatedly testing membership of both types of risk region is therefore computationally intensive. Ways of mitigating this issue are discussed in our follow-up paper [15]. These membership tests, and the aggregation sampling algorithm have been implemented and made available as a package for the Julia programming language [14]. All experiments were conducted on a laptop with an Intel Core i7-720QM CPU at 1.6 GHz.

8.1 Probability of risk regions

As discussed in Sect. 4.1, the performance of the aggregation sampling algorithm with respect to standard Monte Carlo sampling improves as the probability of the aggregation region increases. In this first experiment we observe the behavior of this probability over a range of dimensions.

For this experiment, we suppose that \(K = {\text {conic}}\left( {\mathcal {X}} \right) = {\mathbb {R}}^d_{+}\), and that the random vector follows a multivariate Normal distribution \({\mathcal {N}}(0, \varLambda (\rho ))\), where the covariance matrix \(\varLambda (\rho )\), for \(0\le \rho <1\), is defined as follows:

$$\begin{aligned} \varLambda _{ij}(\rho ) = {\left\{ \begin{array}{ll} \rho &{} \text { if } i\ne j,\\ 1 &{} \text { otherwise.} \end{array}\right. } \end{aligned}$$

In particular, we calculate the probability for the case \(\rho = 0\), that is where the asset returns are independently distributed, and the case \(\rho = 0.3\), that is where the asset returns are positively correlated. The probabilities of the non-risk regions are estimated by sampling and testing membership for 20,000 points.

The results of this experiment are plotted in Fig. 1. In Fig. 1a, b are plotted the probabilities of the conservative and exact aggregation regions. To aid the readers’ intuition we have also plotted a reduced scenario set in two dimensions using conservative and exact risk regions in Fig. 1c, d for \(\rho =0.3\) and \(\beta =0.95\).

The figures show that not only is the probability of the conservative aggregation region smaller than that of the exact aggregation region but also it decays much more quickly. This emphasizes the importance of using an exact risk region for aggregation sampling if possible. Interestingly the probability of the aggregation regions for the correlated asset returns is greater and decays more slows than that of the independent asset returns. This tells us that, in addition to the loss function, the performance of our methodology depends strongly on the distribution of the random vector. Although the probability of the conservative aggregation region decays fairly rapidly, it remains non-negligible for random vectors of a moderate dimension, around 15, for the correlated asset returns. For exact aggregation regions, the probability remains high for the correlated asset returns for up to a dimension of 40.

Open image in new window

8.2 Performance of aggregation sampling

We now test the performance of the aggregation sampling algorithm using conservative and exact risk regions against standard Monte Carlo sampling in terms of the quality of the solutions each method yields.

Experimental Set-up We use the following problem:

$$\begin{aligned} \underset{x\ge 0}{{\text {minimize}}}\ \,&\beta {\text {-CVaR}}(-x^T\varvec{\xi })\nonumber \\ \text {subject to}\, x^T\mu&\ge t\nonumber \\ \sum _{i=1}^{d}x_{i}&= 1\\ x&\ge 0. \end{aligned}$$

(P)

where the asset returns follow a multivariate Normal distribution \({\mathcal {N}}(\mu , \varSigma )\). We use two distributions: one of dimension 5 and another of dimension 10. These distributions have been fitted from monthly return data for randomly selected companies in the FTSE 100 index. The problem is thus to select a portfolio which minimizes the conditional value-at-risk of the one-month return, subject to a minimimum expected return of t, and no short-selling. These distributions have been made available online [13] in an HDF5 file, and can be accessed using the keys “normal/dim = 5/dist 1” and “normal/dim = 10/dist 1”. We use the target expected one-month return \(t=0.005\) which is feasible for the constructed problems.

This problem has been chosen so that we can solve the problem exactly for Normally distributed returns, and so calculate the optimality gap for solutions found from solving scenario-based approximations. The following formula is easily verified by recalling that for continuous probability distributions, the \(\beta {\text {-CVaR}}\) is just the conditional expectation of the random variable above the \(\beta \)-quantile (see [29] for instance):

$$\begin{aligned} \beta {\text {-CVaR}}(-x^T \varvec{\xi }) = (1 - \beta )\mu ^T x + \sqrt{x^T\varSigma x} \int _{\varPhi ^{-1}(\beta )}^\infty z\ d\varPhi (z) \end{aligned}$$

(42)

where \(\varPhi \) denotes the distribution function of the standard Normal distribution. The problem (P) can therefore be solved exactly using an interior point algorithm and in our experiments we use the software package IPOPT [37] to do this.

Denote by \(\{(\xi _{s}, p_{s})\}_{s = 1}^{n}\) a scenario set of size n, where \(\xi _{s}\) denotes the vector of asset returns in scenario s, and \(p_{s}\) the corresponding probability. Then, the scenario-based approximation to (P) using this scenario set, is the following linear program:

$$\begin{aligned} \underset{x, y, \alpha }{{\text {minimize}}}\ \,&\alpha + \frac{1}{1 - \beta } \sum _{s=1}^{n} p_{s}y_{s}\\ \text {subject to } y_{s}&\ge -x^{T}\xi _{s} - \alpha \\ x^{T}\mu&\ge t\\ \sum _{i=1}^{d} x_{i}&= 1\\ x, y&\ge 0. \end{aligned}$$

See [29] for more details on how \(\beta {\text {-CVaR}}\) is linearized for discrete random vectors in this way. These scenario-based problems are modelled using JuMP [10] and solved using Gurobi 7.5 [18].

We are interested in the quality and stability of the solutions that are yielded by our scenario generation method as compared to standard Monte Carlo sampling for a given scenario set size. To this end, in each experiment, for a range of scenario set sizes, we construct 100 scenario sets using sampling and aggregation sampling with conservative and exact risk regions, solve the resulting problems, and calculate the optimality gaps for the solutions that these yield.

Denote by \(z^{*}\) the optimal solution value for problem (P), and by \({\tilde{x}}\) a solution found by solving a scenario-based approximation. Then the optimality gap of \({\tilde{x}}\) is given by

$$\begin{aligned} \beta {\text {-CVaR}}(-{\tilde{x}}^{T}\varvec{\xi }) - z^{*} \end{aligned}$$

where \(\beta {\text {-CVaR}}(-{\tilde{x}}^{T}\varvec{\xi })\) calculated using (42).

Results In Fig. 2 are presented the results of these stability tests for two different problems. In the first problem we have \(d = 5\) and \(\beta =0.95\). In the second problem we have \(d = 10\) and \(\beta = 0.99\). For each scenario set size and scenario generation method we have drawn a box plot of the optimality gap of the 100 constructed scenario sets. In the legend of each plot we have given the estimated probability of the aggregation regions, a, and the true optimal value \(z^{*}\) is included in the title. Note that Cons. Agg. sampling and Exact Agg. Sampling are abbrieviations for, respectively, aggregation sampling using the conservative risk region, and aggregation sampling using the exact risk region.

Open image in new window

In both cases, both aggregation sampling methods outperform standard Monte Carlo sampling for all scenario set sizes in terms of the size and variability of the calculated optimality gaps. This is because for aggregation sampling we are effectively sampling more scenarios compared with standard Monte Carlo sampling. Aggregation sampling with exact risk regions also significantly outperforms aggregation sampling with conservative risk regions. The improved performance can be expected given that its probability is greater than that of the conservative risk region which gives a greater effective sample size.

In this paper we have demonstrated that for stochastic programs which use a tail risk measure, a significant portion of the support of the random vector in the problem may not participate in the calculation of that tail risk measure, whatever feasible decision is used. As a consequence, for scenario-based problems, if we concentrate our scenarios in the region of the distribution which is important to the problem, the risk region, we can represent the uncertainty in our problem in a more parsimonious way, thus reducing the computational burden of solving it.

We have proposed and analyzed two specific methods of scenario generation using risk regions: aggregation sampling and aggregation reduction. Both of these methods were shown to be more effective, in comparison to standard Monte Carlo sampling, as the probability of the non-risk region increases: in essence the higher this probability the more redundancy there is in the original distribution. The application of our methodology relies on having a convenient characterization of a risk region. For portfolio selection problems we derived the exact risk region when returns have an elliptical distribution. However, a characterization of the exact risk region will generally not be possible. Nevertheless, it is sufficient to have a conservative risk region. For stochastic programs with monotonic loss functions, a wide problem class which includes some network design problems, we were able to derive such a region.

The effectiveness of our methodology depends on the probability of the aggregation region, that is the exact or conservative non-risk region used in our scenario generation algorithms. We observed that for both the stochastic programs with monotonic loss function and portfolio selection problems that this probability tends to zero as the dimension of the random vector in the problem increases. However, in some circumstances this effect is mitigated. We observed that small positive correlations slowed down this convergence for the portfolio selection problem.

We tested the performance of our aggregation sampling algorithm for portfolio selection problems using both the exact non-risk region and the conservative risk region for monotonic loss functions. This demonstrated a significant improvement on the performance of standard Monte Carlo sampling, particularly when an exact non-risk region was used.

The methodology has much potential. For some small to moderately-sized network design problems this methodology could yield much better solutions. In particular the methodology is agnostic to the presence of integer variables, and so could be used to solve difficult mixed integer programs.

In our follow-up paper [15] we demonstrate that our methodology may be applied to more difficult and realistic portfolio selection problems such as those involving integer variables, and for which the asset returns are no longer elliptically distributed. In the same paper we also discuss some of the technical issues involved in applying the method, such as finding the conic hull of the feasible region, and methods of projecting points onto this. We also investigate the use of artificial constraints as a way of making our methodology more effective.