A Robust Portfolio Technique for Mitigating the Fragility of CVaR Minimization

Size: px
Start display at page:

Download "A Robust Portfolio Technique for Mitigating the Fragility of CVaR Minimization"

Transcription

1 A Robust Portfolio Technique for Mitigating the Fragility of CVaR Minimization Jun-ya Gotoh Department of Industrial and Systems Engineering Chuo University Kasuga, Bunkyo-ku, Tokyo 2-855, Japan Keita Shinozaki Department of Industrial and Systems Engineering Chuo University Kasuga, Bunkyo-ku, Tokyo 2-855, Japan k Abstract The conditional value-at-risk (CVaR) has gained growing popularity in financial risk management due to the coherence property and tractability in its optimization. However, optimal solutions to the CVaR minimization are highly susceptible to estimation error of the risk measure because the estimate depends on only a small portion of sampled scenarios. In this paper, by employing a robust optimization modeling for minimizing coherent risk measures, we present a simple and practical way for making the solution robust over a certain range of estimation errors. More specifically, we show that a worst-case coherent risk measure minimization leads to a penalized minimization of the empirical risk estimate. With an uncertainty set based on the factor model, which is often employed in practice for estimating the asset return distribution, the minimization becomes a tractable convex optimization problem, in which the penalty term can be represented as a standard deviation of portfolio residuals. Numerical results show that the robust CVaR minimization equipped with a parameter tuning technique achieves dominantly better performance than the existing CVaR minimizations. Keywords: robust portfolio; CVaR (conditional value-at-risk); coherent risk measure; factor model; regularization Introduction In the history of risk measurement for financial risk management, the conditional value-at-risk (CVaR) was originally introduced as an alternative risk measure to the value-at-risk (VaR), which had been used in practice for capturing a risk of large loss with small probability. While VaR has been criticized for its theoretical deficiency, CVaR is known to have nice properties such as the coherence (Artzner et al. 999) and the consistency with the second-order stochastic dominance (Ogryczak and Ruszczýnski 2002). Also, Rockafellar and Uryasev (2000) show that the minimization of CVaR results in a tractable optimization problem. For example, when the loss is defined as the minus return and a finite number of historical observations of returns are used in estimating the objective function, the CVaR minimization can be written as a linear program (LP) and solved efficiently. Due to these preferable properties, it has obtained growing popularity in practice as well, and we can say that CVaR is the most successful coherent risk measure so far. On the other hand, since the estimate of CVaR is computed by using only an upper tail part of the loss distribution, a large number of samples are required for assuring the statistical reliability of the estimate. Especially when CVaR is employed as the objective of a portfolio optimization, a much larger number of samples are required for ensuring the accuracy of the Corresponding author

2 optimal portfolio. For example, Takeda and Kanamori (2009) show the estimation error in the minimal CVaR is inversely proportional to β. This indicates that estimation error in CVaR is much severer than that of the mean loss since β 0.90 is usually applied while the mean loss is equal to CVaR with β = In addition, employing a large number of scenarios makes the resulting LP dense, i.e., to contain a prohibitively large number of nonzero entries. In order to obtain a sparser LP formulation of a large-scale problem, Konno, Waki and Yuuki (2002) propose to introduce a factor-model representation to the CVaR minimization as an analogy to the factor representation of the meanvariance model developed by Perold (984). Although they show through numerical results that the computation time can be significantly decreased by employing the factor representation, the out-of-sample performance is not examined. In practice, however, the number of samples which is available for the estimation is limited, and the estimated CVaR and the resulting optimal portfolio may contain considerable estimation error. To alleviate the estimation error of the CVaR minimization, Gotoh and Takeda (2008, 2009) employ a regularization technique so that the optimal portfolio vector will be less sensitive to the data, as in the mean-variance minimization by DeMiguel et al. (2009), Brodie et al. (2009). For the last decade, robust optimization techniques have been introduced in the field of financial portfolio selection, and shown to achieve a nice performance. See, e.g., Ben-Tal, El- Ghaoui and Nemirovski (2009) for the comprehensive survey of recent development in robust optimization. The robust portfolio takes into account the least favorable scenario or estimate. The set of possible scenarios (or estimates) is called uncertainty set. It should be noted that specifying the shape and size of uncertainty set is critical in the sense that robust formulation with improper selection of the uncertainty set can result in a useless solution. However, little has been examined how to specify the uncertainty sets in practical situation. Among such, Goldfarb and Iyengar (2003) presented a practical robust portfolio selection model, employing the confidence region of the factor model as an uncertainty set. Since their uncertainty set does not capture all the possible cases in the set, it may not be adequate to call it a robust optimization in the sense that a worse case may happen with a very small probability. Nevertheless, the uncertainty modeling based on the factor model is one of the most promising approaches since, in reality, we cannot actually recognize the worst case in a definite manner, but can do only in a probabilistic manner. The purpose of this article is to present a simple and practical approach for providing a CVaR minimizing portfolio with a robust property by employing the ordinary regression of the factor model similarly to Goldfarb and Iyengar (2003). In addition to the difference of the employed risk measures, we examine a more nonparametric way of exploiting the factor model, employing a statistical learning approach for determining the size of the uncertainty set. A robust portfolio model for the CVaR minimization has already been proposed by Zhu and Fukushima (2009). They focus on the uncertainty in the probability distribution used for defining CVaR. Such a modeling is called distributionally robust modeling. It is true that the probability estimation itself is under uncertainty and we cannot know the true one. However, it is not easy to imagine what form of uncertainty set is proper for the probability measure. In this sense, employing the uncertainty of probability distribution may not provide investors with a satisfactory solution. In contrast, we provide a more practical and simple way for making the CVaR minimizing portfolio robust against estimation errors. Another good point on our formulation is that our robust modeling is compatible with the distributionally robust modeling of, e.g., Zhu and Fukushima (2009). Therefore, if a reasonable uncertainty set for the probability distribution is given, then the two uncertainty sets can be taken into account simultaneously. The characteristics of the approach we take can be summarized as follows: Unlike Zhu and Fukushima (2009), we consider the uncertainty associated with the per- 2

3 turbation in observed return vectors in order to mitigate the fragility of the CVaR minimization. In particular, we apply a perturbation set associated with the distribution of the residual vectors arising from the ordinary least square method for multi-factor models. Since it is often assumed that the residual follows a normal distribution, it seems reasonable to employ the ellipsoidal uncertainty in the perturbation. Including CVaR as a special case, our robustifying technique is applicable to any coherent risk measure. In general, the minimization of the worst-case coherent risk measure turns out to be the minimization of the sum of the empirical risk measure and a penalty term. Especially when the perturbation is given by a norm term, the penalty term is represented by the dual norm of portfolio vector. This indicates that the norm-constrained CVaR minimization in Gotoh and Takeda (2009) is interpretable from the worst-case minimization. Another interesting fact is that the worst-case formulation presented in this paper can be combined with the distributionally robust minimization of the coherent risk, e.g., the CVaR minimization in Zhu and Fukushima (2009). Numerical results show that the developed robust CVaR model achieves significantly smaller CVaRs out-of-sample than the existing CVaR minimizations and two benchmark portfolios. Also, the regularization brings an effect of lowering the turnover of the portfolio vectors. The structure of this paper is as follows. The next section will be devoted to describing two existing implementation techniques of the CVaR minimization. Then a robust counterpart of the ordinary CVaR minimization will be provided in Section 3. Section 4 reports some numerical results, showing how the robust CVaR minimization improves the out-of-sample performance of non-robust CVaR minimizations. Finally, some remarks will be provided in the final section. 2 Formulations of CVaR Minimization Let there be n investable assets, and let R := (R,..., R n ) be their random return and π := (π,..., π n ) be a portfolio vector. Each component of π represents the investment ratio into an asset, and e n π = is fulfilled by definition, where e n := (,..., ) IR n. All the conditions on π are denoted by a set Π π IR n : e n π =. A portfolio optimization problem seeks to find a portfolio π Π so that the distribution of portfolio return R π would have a shape which is desirable to the investor. Typical criteria for tailoring the distribution are to minimize the risk and to maximize the mean return. It has been pointed out, however, that estimating the mean return in a satisfactory manner is a hard task. In fact, Jagannathan and Ma (2003) conclude that ignoring the mean return does not lose so much. Therefore, we in this paper limit the discussion to the minimization of a risk as in DeMiguel et al. (2009) where variance minimization is examined. Among many candidates for the risk measure, the conditional value-at-risk (CVaR) associated with the portfolio return is minimized since CVaR is important both in theory and practice but has fragility in estimation and optimization, as pointed out in Introduction. With a nonincreasing function L, let us define a portfolio loss by L(R π). The β-conditional value-at-risk (β-cvar) associated with the loss is defined by ϕ β (π) := min α α + β E[maxL(R π) α, 0] where β [0, ) is a significance level, and E[ ] is the operator of mathematical expectation. In the following part of this paper, a loss of the form R π is only considered. Also, we assume that the probability associated with R is independent of π as in Rockafellar and Uryasev (2002). 3,

4 CVaR can be roughly considered as the conditional expectation of the loss R π which exceeds the 00β-percentile of the loss, i.e., the β-value-at-risk (VaR). Following a custom of statistics, β is usually fixed at a value near.0, e.g., β = In order to minimize ϕ β (π), we have to estimate the risk measure since the true distribution of R is never known in practice. When only a collection of T historical observations R,..., R T of R are available, the empirical distribution is used to estimate the risk measure. More precisely, the empirical β-cvar, ϕ e β, is defined by ϕ e β (π) := min α α + ( β)t max R t π α, 0, () and the minimization is reduced to a linear programming (LP) problem: min α + ( β)t e T y s.t. y t R t π α, t =,..., T y 0, π Π. (2) as long as the set Π is given by a polyhedron. An optimal solution to (2) is an estimate of the portfolio that minimizes ϕ β (π). It is easy to see that ϕ e β (π) depends only on a portion of loss scenarios R [] π,..., R [τ] π with τ = ( β)t, where R [k] π indicates the k-th largest loss scenario and x is the smallest integer that is no less than x. See, e.g., Proposition 8 of Rockafellar and Uryasev (2002) for the details. This indicates that a perturbation in those large loss scenarios can make a big impact on the estimate of CVaR. The things become worse when we compute a portfolio that minimizes the empirical CVaR since the portfolio seeks to fit the perturbed sample. A practical way to alleviate the effect of such a perturbation is to employ a statistical model. Konno, Waki and Yuuki (2002) replace the observed returns R,..., R T in () with values estimated by a regression approach. Let us suppose that the structure of asset returns is described as a linear multi-factor model: K R j = β 0j + β kj F k + E j j =,..., n, (3) k= where F k are K factors used for explaining the asset returns R j, β kj are intercepts (k = 0) and factor loadings (k =,..., K) to be estimated, and E j are the residuals. The parameters β kj, k = 0,..., K, are usually estimated via the ordinary least squares (OLS) method for each asset j. Let F t,k be T observed values of the k-th factor, t =,..., T, and let β k := ( β k,..., β kn ) be the estimated intercept (k = 0) and factor loading (k =,..., K) vectors. Namely, T ( β 0j, β j,..., β K Kj ) arg min (R tj (β 0j + β kj F t,k )) 2, j =,..., n. β 0j,...,β Kj The use of the OLS method implies that any factor F k and residual E j are independent of each other, and satisfies E[E j ] = 0. The vectors R t := β 0 + K β k= k F t,k, t =,..., T, can be considered as T scenarios generated by the linear model (3). Let us define another CVaR by replacing the historical returns R t with the theoretical returns R t, and denote it by ϕ f β. ϕ f β (π) := min α α + ( β)t k= max R t π α, 0. 4

5 This estimate eliminates the residual return from the empirical CVaR, ϕ e β. Konno, Waki and Yuuki (2002) introduce this estimate, motivated by the fact that n j= E jπ j = 0 is expected to hold for a well-diversified portfolio π, and show that this simplification reduces the computation time for solving the LP (2). Let ε,..., ε T denote the realized residuals which are defined by ε t := R t R t = R t β 0 K β k F t,k. k= Then, the factor model-based CVaR can be rewritten by ϕ f β (π) = min α + max (R t ε t ) π α, 0. (4) α ( β)t When Π is represented by a system of linear inequalities or equalities, its minimization can be rewritten as a linear programming problem: min α + ( β)t e T y s.t. y t (R t ε t ) π α, t =,..., T y 0, π Π. (5) Since the variance of R t is larger than that of R t ε t while the mean values meet due to the nature of the OLS method, the factor-based CVaR ϕ f β (π) is expected to be (probabilistically) less than or equal to the empirical CVaR ϕ e β. In particular, if R and E follow normal distributions N(µ, V ) and N(0, Σ), respectively, one then has ϕ F β (π) := min α α + β E[ (R E) π α] + = µ π + C β π (V Σ)π µ π + C β π V π = ϕ β (π) where C β := exp( 2 F (β) 2 ) ( β). Since, for any ϵ > 0, lim 2π T P ϕ f β (π) ϕf β (π) > ϵ = 0 and lim T P ϕ e β (π) ϕ β(π) > ϵ = 0, one has lim T P ϕ f β (π) ϕe β (π)+c β π Σπ > ϵ = 0. In such a sense, the factor representation in Konno, Waki and Yuuki (2002) is expected to underestimate the empirical CVaR ϕ e β. 3 Robust CVaR Minimization Based on Multifactor Model 3. A Worst-case CVaR Recalling that both of the estimated CVaRs, ϕ e β in () and ϕf β in (4), are determined by only a small portion of T samples, R t and R t ε t, respectively, they can be highly affected by a small number of perturbed samples when the number of available samples, T, is small relatively to the number of assets, n. To overcome the fragility in ϕ e β, we below introduce a robust CVaR minimizing portfolio model in a different manner from Zhu and Fukushima (2009). Let us define an uncertainty set U by U := (r,..., r T ) : (r,..., r T ) = (R,..., R T ) (ε,..., ε T ) for some (ε,..., ε T ) Z. with a perturbation set of the form Z := (ε,..., ε T ) : ε t V, t =,..., T, 5

6 where V IR n is a compact set containing the origin in its interior. Among examples of V are () a finite set of possible vectors, denoted by V I := v,..., v k, so that the convex hull of this set contains the origin in its interior, and (2) a box-shaped polytope, denoted by V B := v : v L v v U with v Lj < 0 and v Uj > 0, j =,..., n. In what follows, we especially pay attention to a norm-based perturbation set of the form Z := (ε,..., ε T ) : ε t δ, t =,..., T, (6) where is a norm in IR n and δ is a parameter defining the size of the uncertainty set. Apparently, the corresponding uncertainty set is centered at the observed return vectors and the shape and size are determined by and δ, respectively. For example, if δ = 0, then U shrinks to the observed return vectors (R,..., R T ). Obviously, if ε follows a multivariate distribution, its confidence region may be employed as the perturbation set Z. For example, if ε follows N(0, Σ), the use of a norm of the form ε Σ := ε Σ ε is reasonable since a confidence region of the residual vector forms an ellipsoidal region. With the uncertainty set U, a worst-case CVaR, ϕ w β (π), of π is defined by ϕ w β (π) := max (r,...,r min α + T ) U α = max (ε,...,ε min T ) Z α α + ( β)t ( β)t max r t π α, 0 max (R t ε t ) π α, 0. (7) It is easy to see that ϕ e β (π) ϕw β (π). Combining the observation in the previous section, we can expect that ϕ f β (π) ϕw β (π) holds in a probabilistic sense. In particular, if we employ the norm-based perturbation (6) and set δ = max ε t : t =,..., T, one has ϕ f β (π) ϕw β (π). Also, it is contrastive that the worst-case CVaR (7) assumes the residual vectors, ε t, are variables, whereas the factor model-based CVaR (4) assumes they are fixed values, i.e., ε t = ε t. It should be noted that the worst-case CVaR, ϕ w β (π), is different from the worst-case CVaR proposed by Zhu and Fukushima (2009), where the (discrete distribution version of) worst-case CVaR is defined by max min p P α α + β p t max R t π α, 0 where P is a compact convex set in IR T, representing the uncertainty set of the probability distribution. Our worst-case CVaR takes care of the worst scenario in the return vectors (r,..., r T ) and keeps the empirical probability, p = (/T,..., /T ), intact, whereas that in Zhu and Fukushima (2009) do in the probability distributions (p,..., p T ) and keeps the observed return, (R,..., R T ), intact. Although Zhu and Fukushima (2009) describe how to specify the uncertainty set P of the probability distribution and present some numerical example showing how to apply their risk measure to practical data, it seems more difficult to agree about an uncertainty set of the probability distribution than that of the return vectors. In fact, a more commonly used idea in statistic analysis can be employed in specifying the uncertainty set of the return vector, as will be seen in Subsection 3.3. Despite the different assumptions on the origin of uncertainty, the resulting optimization can also be transformed into a convex optimization problem as the worst-case CVaR of Zhu and Fukushima (2009). 6

7 3.2 Decomposition Property in Worst-case Coherent Risk Measures In order to discuss the tractability of the worst-case CVaR ϕ w β (π) in a general framework, let us first introduce the coherent risk measures (Artzner et al. 999). Let ρ(π) := ρ[ R π] be a coherent risk measure of portfolio π. Namely, ρ[ ] satisfies the following properties for random variables X, Y which are assumed to represent portfolio loss variables: (i) ρ[x ] ρ[y] if X Y; (ii) ρ[ax + b] = aρ[x ] + b for all a 0, b IR; (iii) ρ[x + Y] ρ[x ] + ρ[y]. It is known that any coherent risk measure can be written by a dual representation as follows: ρ[x ] = supep[x ] : p P where Ep[X ] is the mathematical expectation of X under probability distribution p, and P is a compact convex. Indeed, the empirical CVaR is a coherent measure with P = p IR T : e T p =, 0 p e T /(( β)t ). The empirical version of the coherent risk measure, ρ e (π), is then represented by ρ e (π) := max p t R t π : p P for some P. Also, the worst-case coherent risk, ρ w (π), can be similarly defined by ρ w (π) := max (ε max T p t (R t ε t ) π : p P. t ) Z With the generalized perturbation set, the worst-case coherent risk measure can be decomposed into the empirical version plus a penalty term. Theorem The worst-case coherent risk measure with the generalized perturbation can be represented by ρ w (π) = ρ e (π) + max s s π : s V. (8) See the appendix for the proof of the theorem. In the above sense, the worst-case CVaR does not lose the coherent property. Also, the worst-case coherent risk measure ρ w (π) is a convex function since both ρ e (π) and τ(π) := maxs π : s V are convex functions. Since CVaR is a coherent risk measure, the worst-case CVaR, ϕ w β (π), can also be written as the sum of the empirical version and a convex penalty term. Corollary The worst-case CVaR (7) can be represented by ϕ w β (π) = ϕe β (π) + max s s π : s V. In particular, if the norm-based perturbation (6) is employed, the worst-case CVaR can be decomposed into the empirical CVaR and a regularization term: ϕ w β (π) = ϕe β (π) + δ π (9) where π denotes the dual norm of π, i.e., π := maxss π : s. The second statement of Corollary states that when the norm-based perturbation is employed, the worst-case CVaR (7) can be decomposed into the empirical CVaR ϕ e β (π) and a norm of a portfolio. Similarly, the worst-case CVaRs with the finite perturbation set V I and the box perturbation set V B can be shown to be tractable convex functions. 7

8 Consequently, the robust counterpart of the CVaR minimization, which is written by a minmax-min optimization problem: min π Π ϕw β (π) = min max π Π (ε,...,ε min α + max (R t ε t ) π α, 0 T ) Z α ( β)t can be rewritten as a regularized minimization of empirical CVaR. Proposition Let Π be given as a system of linear inequalities and equalities. If the norm-based perturbation set is employed, the min-max-min robust counterpart of the CVaR minimization (2) can be reformulated as a convex optimization problem as follows: min α + ( β)t e T y + δ π s.t. y t R t π α, t =,..., T y 0, π Π. (0) Similarly, if the finite perturbation V I is employed, it results in an LP: min α + ( β)t e T y + ζ s.t. y t R t π α, t =,..., T y 0, π Π ζ v h π, h =,..., k. If the box perturbation V B is employed, it results in an LP: min α + ( β)t e T y + (v U v L ) w + v L π s.t. y t R t π α, t =,..., T y 0, π Π w π, w 0. The first and second statements are proved in a straightforward manner. See the appendix for the proof of the third one. It is noteworthy that the role of the norm-based regularization term in the CVaR minimization can also be justified by the generalization bounds for the ν-support vector machine, which is a collective term for statistical methods including classification and regression methods. As Gotoh and Takeda (2009) show, imposing a norm constraint can improve the performance of the CVaR type minimization. The discussion above explains that the regularized (or, equivalently, norm-constrained) CVaR minimization can be justified also in the context of robust optimization. Although Proposition 2 in Gotoh and Takeda (2009) also shows a constraint-robust formulation can be equivalent to the regularized formulation (0), we restate the statement above in order to clarify the relation between the usual min-max-min type formulation and the regularized formulation. On the other hand, Theorem also indicates that the robustification technique developed in this article can be applied to any coherent risk measure. In this sense, the worst-case CVaR ϕ w β is compatible with the worst-case CVaR proposed in Zhu and Fukushima (2009) since the latter is proved in their paper to be a coherent risk measure. For example, let us consider the box uncertainty set for the underlying probability: P = p IR T : p = e T T + η, for some η s.t. e T η = 0, η L η η U, 8

9 where η L and η U are given lower and upper bounds on probability distribution, respectively. Combining the argument in Zhu and Fukushima (2009) and Proposition with the normuncertainty (6) for the return vector, a worst-case CVaR which cares both the observation uncertainty and the distribution uncertainty can be written by the following convex program: min θ + δ π s.t. α + e T z ( β)t + β (η L ξ + η U ω) θ e T u + ξ + ω = z, ξ 0, ω 0 z t R t π α, z t 0, t =,..., T. Similarly, it is easy to see that the worst-case CVaR minimization (0) can cohabit with the worst-case CVaR minimization of Zhu and Fukushima (2009) with the ellipsoidal uncertainty P = p IR T : p = e T T + Aη 0 for some η s.t. e T Aη = 0, η η, where A IR T T. Needless to say, any combination of the above-mentioned uncertainty sets of perturbation and distribution can results in a convex optimization. Nevertheless, in what follows, we concentrate only on the single application of the norm-based perturbation set in order not to obscure the outline of our argument. 3.3 Formulation with Factor Model-based Uncertainty In order to implement the robust CVaR minimization (0), we have to set up how to determine the norm (or, equivalently, ) and the parameter δ. Goldfarb and Iyengar (2003) show how uncertainty sets of the mean return vector and the covariance matrix are constructed on the basis of the factor model. Although the same line can be traced for the CVaR minimization, we here give a simpler way for constructing the perturbation set Z. Suppose that the OLS method is applied to the estimation of the factor model (3). Then the statistical confidence regions of the residual vectors can be exploited as a perturbation set Z. In fact, the t-tests on the factor loadings of the factor model are often conducted under a normality assumption. Therefore, it is reasonable to employ a norm induced from the covariance matrix of the residual vectors, ε t, of the factor model. Let Σ := ε ε n T E E, E :=.., ε T ε T n and let us consider the norm of the form ε Σ := ε Σ ε for the perturbation set Z. As already mentioned, the use of this norm can be justified if the residual follows an elliptical distribution (not necessarily normal distribution). By Corollary, the worst-case CVaR becomes a convex function of the form ϕ w β (π) = min α α + ( β)t max R t π α, 0 + δ π Σπ. Noting that the term π Σπ indicates the variance of the residual of the portfolio return, reducing the term π Σπ is expected to improve the goodness-of-fit of the factor model. Accordingly, 9

10 it can be regarded that the worst-case CVaR minimization seeks to reduce the empirical CVaR and to improve the predictability simultaneously. The robust counterpart (0) can then be rewritten by a second order cone optimization problem, which can be efficiently solved by an interior point algorithm, of the form: min α + ( β)t e T y + δv s.t. y t R t π α, t =,..., T, y 0, π Π. v z z, z = /T Eπ, as long as Π is represented by a system of linear inequalities and equalities or second order cones. It is noteworthy that this optimization is well-defined even for the case where the number of samples is smaller than the size of the universe, i.e., T < n. Although such a case does not correspond to a norm in the sense that Σ is no longer a positive definite matrix, the quadratic term, π Σπ, certainly defines a regularization term of the portfolio vector. As for the parameter δ, we can apply a quantile of the distribution of the residual, corresponding to a confidence level, e.g., 0.95 and In the numerical experiment in the next section, however, the value of δ is determined by a data-driven approach so as to avoid arbitrariness in assuming a specific distribution or in determining the confidence level. 4 Numerical Experiments In this section, numerical results are presented for comparing the robust formulation () with two existing implementations (2) and (5) of the minimization of CVaR ϕ β (π) subject to π Π = π : e n π =, π 0. We use the 202 monthly return vectors of 85 stocks listed in the Nikkei225 index through February 993 to November The linear models (3) are estimated by the OLS method using the three factors advocated in Fama and French (993). The used factor data are computed along the line of the paper Kubota and Takehara (2007) and are available at the web site of Nikkei Media Marketing, Inc. (200). A rolling horizon scheme is employed for constructing portfolios and evaluating the performances. For the t-th time window of length T = 20, the factor model is estimated by using (R τ, F τ,k ) : τ = t, t +,..., t + 9, and an optimal portfolio, π t, is computed, t =,..., 82(= ). The out-of-sample portfolio returns are accordingly simulated by R t+20π t, t =,..., 82. The out-of-sample performance can be gauged by the following measures: Out-of-sample CVaR ˆϕβ := min α α + 82( β) 82 Out-of-sample turnover of portfolio vectors π := 8 where π t+,j := ( + r t+,j)πt,j n i= ( + r t+,i)πt,i. Out-of-sample mean ˆµ := R 20+tπ t Out-of-sample standard deviation ˆσ := max R 20+tπ t α, 0 8 j= (R 20+tπ t ˆµ)2 n π t+,j πt+,j ()

11 n : size of universe n : size of universe (a) β = 0.90 (b) β = 0.95 Figure : Out-of-sample CVaR Left figure shows the average of the achieved CVaRs with β = 0.90 while right one shows that with β = Each value indicates the average of ten CVaRs, each series of portfolios is constructed using randomly chosen universe. In computing an optimal portfolio π t, the parameter δ > 0 must already be given only on the basis of data available before time t We dynamically tune the value of δ by employing a six-fold cross validation (see, e.g., Hastie, Tibshirani and Friedman 200) in the following manner. For the t-th time window, the data set of size 20 is divided into six subsets of size 20, i.e., (R τ, F τ,k ) : τ = t, t +,..., t + 9, 2 (R τ, F τ,k ) : τ = t + 20,..., t + 39,..., 6 (R τ, F τ,k ) : τ = t + 00,..., t + 9. A portfolios, denoted by π k δ,t, is constructed for some δ by using the union of five subsets out of the six subsets, i.e.,,..., 6 \ k, and the performance of the optimal portfolio is simulated over the remaining subset k which is unused for computing the portfolio. Repeating the construction and simulation for all the six combinations of the subsets, 20 simulated portfolio returns are obtained, i.e., R t π δ,t,..., R t+9π 6 δ,t, and let µ δ,t := τ=0 k= R 20(k )+τ π k δ,t and σ δ,t := τ=0 k= (R 20(k )+τ π k δ,t µ δ,t) 2. The value of δ at time window t is chosen from k : k = 0,, 2, 3, 4 so that the value µ δ,t + C β σ δ,t is the smallest, and the optimal portfolio π t is computed using that δ and all the data,..., 6. This is repeated by sliding the time window. We computed portfolios for various asset universes of size n = 20, 40,..., 60. For each n, ten sets of assets are randomly chosen from the 85 assets contained in the Nikkei225 index. Figure shows the average of the out-of-sample CVaRs, ˆϕ β, with β = 0.90 and We see that the robust model achieves the smallest CVaRs out-of-sample for any size of universe. On the other hand, the minimization of the factor model-based CVaR ϕ f β (π) constantly results in the largest CVaR values. In addition, the difference between ϕ f β (π) and ϕw β (π) becomes larger as the size of universe grows. We guess that this is because minimizing ϕ f β (π) excessively fit the past data, while the other models successfully reduced the out-of-sample CVaR by diversification. We see from Figure 2 that the turnover of the robust CVaR model is smaller than the two CVaR alternatives, while worse than the equally weighted portfolio. The regularization term works for decreasing the turnover as well as improving the out-of-sample CVaR. Figures 3 and 4 show the out-of-sample mean and the standard deviation, respectively, of the out-of-sample returns. Although the mean return of the robust model is smaller than the index and the equally weighted portfolio on average, the robust model achieves higher mean return than the other two alternatives except the case of (β, n) = (0.95, 20). On the other hand, the robust model achieves the smallest standard deviation among all the models. In addition, the value decreases as the

12 n : size of universe n : size of universe (a) β = 0.90 (b) β = 0.95 Figure 2: Out-of-sample Turnover n : size of universe n : size of universe (a) β = 0.90 (b) β = 0.95 Figure 3: Mean of Out-of-sample Returns number of assets decreases. Such a low standard deviation is partly led by the small CVaR values in Figure, and besides, the tuning of δ seems to contribute to a certain degree. Table reports the p-values of the Wilcoxon signed rank test on the out-of-sample performance of the robust model in comparison with the nominal CVaR minimization model. From this table, we see that the robust CVaR model combined with the tuning of δ outperforms the ordinary CVaR models in various aspects. In particular, when the number of assets is large, the difference is significantly supported. 5 Concluding Remarks In this paper, we present a new formulation for the CVaR minimization by introducing a robust optimization technique. When a norm-based perturbation is employed, it results in a convex optimization whose objective function consists of the usual nominal CVaR and a regularization term. The distribution of the residual vectors of the factor model is exploited so as to define the regularization term. More generally, the CVaR can be replaced with any coherent measure of risk. What is interesting here is that the robust modeling employed in the present paper is compatible with the distributionally robust model for coherent risk measure minimization. 2

13 n : size of universe n : size of universe (a) β = 0.90 (b) β = 0.95 Figure 4: Standard Deviation of Out-of-sample Returns Table : p-values of the Wilcoxon signed rank test of the out-of-sample performance of the robust model relative to the nominal model CVaR ˆϕ β Turnover π Mean ˆµ Standard deviation ˆσ β = 0.90 β = 0.95 β = 0.90 β = 0.95 β = 0.90 β = 0.95 β = 0.90 β = n Each value indicates the p-value of the Wilcoxon signed rank test for the out-of-sample performance of the robust model versus that of the nominal model over the ten trials, each using a randomly chosen asset universe of size n. The null hypothesis H 0 : The robust model achieved as good performance as the nominal model; The alternate hypothesis H : The robust model is superior to the nominal model. 3

14 Through some numerical experiments, the robust model constantly achieves smaller CVaRs, turnover, standard deviation and higher mean than the other CVaR minimization models. In particular, the simple factor representation developed in Konno, Waki and Yuuki (2002) results in the worst performance in all aspects among the several models, so that it should not be used. In contrast, employing the regularization term in minimizing CVaR mitigates the estimation error or the effect of outlying data. It should be recalled that the use of the regularization term can also be justified from the viewpoint of the generalization theory for machine learning (Schölkopf et al. 2000, Gotoh and Takeda 2008, 2009). In this paper, we omit the mean maximizing term according to DeMiguel et al. (2009), which focuses on the variance minimization instead of the mean-variance model. It is easy to see that the robust modeling developed in this article can also be applied to the mean maximizing term. The investigation on the effect of the regularization term in the mean-coherent risk minimization remains as a future research. Acknowledgment The research of the first author is partly supported by a MEXT Grant-in-Aid for Young Scientists (B) A Proof of Theorem Proof. The worst-case coherent risk measure ρ w can be rewritten as ρ w (π) = max q ε t R t,q t π + q t ε t π : q P, ε t δ, t =,..., T. Note that for q P, we have max (ε t) Z q t ε t π = so the proof is complete. q t max (ε ε t π = max t) Z ε ε π : ε V q t = max ε ε π : ε V, B Proof of Proposition Proof. When the box perturbation V B is employed, the penalty term is represented by an LP of the maximization form: maxπ s : v L s v U for some π Π. By the duality theorem, it can be replaced with the dual LP of the minimization form: min(v U v L ) w + v L π : w π, w 0, and the formulation of the third statement is obtained. References Artzner P, Delbaen F, Eber JM, Heath D (999) Coherent Measures of Risk. Mathematical Finance 9: Ben-Tal A, ElGhaoui L, Nemirovski A (2009) Robust Optimization. Princeton Univ. Press. 4

15 Brodie J, Daubechiesa I, De Mol C, Giannone D, Lorisc I (2009) Sparse and Stable Markowitz Portfolios, PNAS 06: DeMiguel V, Garlappi L, Nogales FJ, Uppal R (2009) A Generalized Approach to Portfolio Optimization: Improving Performance by Constraining Portfolio Norms. Management Science 55: Fama E, French K (993) Common Risk Factors in the Returns on Stocks and Bonds. Journal of Financial Economics 33: Goldfarb D, Iyengar G (2003) Robust portfolio selection problems. Mathematics of Operations Research 28: 38. Gotoh J, Takeda A (2009) On the Role of Norm Constraints in Portfolio Selection, Department of Industrial and Systems Engineering Discussion Paper Series ISE09-03, Chuo University. Gotoh J, Takeda A (2008) Minimizing Loss Probability Bounds for Portfolio Selection, Department of Industrial and Systems Engineering Discussion Paper Series ISE08-04, Chuo University. Hastie T, Tibshirani R, Friedman J (200) The Elements of Statistical Learning Data Mining, Inference, and Prediction. Springer, NY. Jagannathan R, Ma T (2003) Risk Reduction in Large Portfolios: Why Imposing the Wrong Constraints Helps. The Journal of Finance 4: Konno H, Yuuki A, Waki H (2002) Portfolio Optimization under Lower Partial Risk Measures. Asia- Pacific Financial Markets 9: Kubota K, Takehara H (2007) Fama-French Factor Model no Yukousei no Saikenshou, (in Japanese) Nikkei Media Marketing, Inc. (200) (in Japanese) Ogryczak W, Ruszczýnski A (2002) Dual stochastic dominance and related mean risk models. SIAM Journal on Optimization 3: Perold AF (984) Large Scale Portfolio Optimization. Management Science 30: Rockafellar TR, Uryasev S (2000) Optimization of Conditional Value-At-Risk. The Journal of Risk 2:2 4. Rockafellar TR, Uryasev S (2002) Conditional Value-at-Risk for General Loss Distributions. Journal of Banking and Finance 26: Schölkopf B, Smola AJ, Williamson RC, Bartlett PL (2000) New support vector algorithms. Neural Computation 2: Takeda A, Kanamori T (2009) A robust approach based on conditional value-at-risk measure to statistical learning problems. European Journal of Operational Research 98: Zhu S, Fukushima M (2009) Worst-Case Conditional Value-at-Risk with Application to Robust Portfolio Management. Operations Research 57:

Robust portfolio selection under norm uncertainty

Robust portfolio selection under norm uncertainty Wang and Cheng Journal of Inequalities and Applications (2016) 2016:164 DOI 10.1186/s13660-016-1102-4 R E S E A R C H Open Access Robust portfolio selection under norm uncertainty Lei Wang 1 and Xi Cheng

More information

Worst-Case Conditional Value-at-Risk with Application to Robust Portfolio Management

Worst-Case Conditional Value-at-Risk with Application to Robust Portfolio Management Worst-Case Conditional Value-at-Risk with Application to Robust Portfolio Management Shu-Shang Zhu Department of Management Science, School of Management, Fudan University, Shanghai 200433, China, sszhu@fudan.edu.cn

More information

A SECOND ORDER STOCHASTIC DOMINANCE PORTFOLIO EFFICIENCY MEASURE

A SECOND ORDER STOCHASTIC DOMINANCE PORTFOLIO EFFICIENCY MEASURE K Y B E R N E I K A V O L U M E 4 4 ( 2 0 0 8 ), N U M B E R 2, P A G E S 2 4 3 2 5 8 A SECOND ORDER SOCHASIC DOMINANCE PORFOLIO EFFICIENCY MEASURE Miloš Kopa and Petr Chovanec In this paper, we introduce

More information

Robust Portfolio Selection Based on a Joint Ellipsoidal Uncertainty Set

Robust Portfolio Selection Based on a Joint Ellipsoidal Uncertainty Set Robust Portfolio Selection Based on a Joint Ellipsoidal Uncertainty Set Zhaosong Lu March 15, 2009 Revised: September 29, 2009 Abstract The separable uncertainty sets have been widely used in robust portfolio

More information

Short Course Robust Optimization and Machine Learning. Lecture 6: Robust Optimization in Machine Learning

Short Course Robust Optimization and Machine Learning. Lecture 6: Robust Optimization in Machine Learning Short Course Robust Optimization and Machine Machine Lecture 6: Robust Optimization in Machine Laurent El Ghaoui EECS and IEOR Departments UC Berkeley Spring seminar TRANSP-OR, Zinal, Jan. 16-19, 2012

More information

FE670 Algorithmic Trading Strategies. Stevens Institute of Technology

FE670 Algorithmic Trading Strategies. Stevens Institute of Technology FE670 Algorithmic Trading Strategies Lecture 8. Robust Portfolio Optimization Steve Yang Stevens Institute of Technology 10/17/2013 Outline 1 Robust Mean-Variance Formulations 2 Uncertain in Expected Return

More information

Risk Management of Portfolios by CVaR Optimization

Risk Management of Portfolios by CVaR Optimization Risk Management of Portfolios by CVaR Optimization Thomas F. Coleman a Dept of Combinatorics & Optimization University of Waterloo a Ophelia Lazaridis University Research Chair 1 Joint work with Yuying

More information

Semidefinite and Second Order Cone Programming Seminar Fall 2012 Project: Robust Optimization and its Application of Robust Portfolio Optimization

Semidefinite and Second Order Cone Programming Seminar Fall 2012 Project: Robust Optimization and its Application of Robust Portfolio Optimization Semidefinite and Second Order Cone Programming Seminar Fall 2012 Project: Robust Optimization and its Application of Robust Portfolio Optimization Instructor: Farid Alizadeh Author: Ai Kagawa 12/12/2012

More information

Department of Social Systems and Management. Discussion Paper Series

Department of Social Systems and Management. Discussion Paper Series Department of Social Systems and Management Discussion Paper Series No. 1114 The Downside Risk-Averse News-Vendor Minimizing Conditional Value-at-Risk Jun-ya Gotoh and Yuichi Takano April 25 UNIVERSITY

More information

Expected Shortfall is not elicitable so what?

Expected Shortfall is not elicitable so what? Expected Shortfall is not elicitable so what? Dirk Tasche Bank of England Prudential Regulation Authority 1 dirk.tasche@gmx.net Finance & Stochastics seminar Imperial College, November 20, 2013 1 The opinions

More information

Włodzimierz Ogryczak. Warsaw University of Technology, ICCE ON ROBUST SOLUTIONS TO MULTI-OBJECTIVE LINEAR PROGRAMS. Introduction. Abstract.

Włodzimierz Ogryczak. Warsaw University of Technology, ICCE ON ROBUST SOLUTIONS TO MULTI-OBJECTIVE LINEAR PROGRAMS. Introduction. Abstract. Włodzimierz Ogryczak Warsaw University of Technology, ICCE ON ROBUST SOLUTIONS TO MULTI-OBJECTIVE LINEAR PROGRAMS Abstract In multiple criteria linear programming (MOLP) any efficient solution can be found

More information

Robustness and bootstrap techniques in portfolio efficiency tests

Robustness and bootstrap techniques in portfolio efficiency tests Robustness and bootstrap techniques in portfolio efficiency tests Dept. of Probability and Mathematical Statistics, Charles University, Prague, Czech Republic July 8, 2013 Motivation Portfolio selection

More information

Motivation General concept of CVaR Optimization Comparison. VaR and CVaR. Přemysl Bejda.

Motivation General concept of CVaR Optimization Comparison. VaR and CVaR. Přemysl Bejda. VaR and CVaR Přemysl Bejda premyslbejda@gmail.com 2014 Contents 1 Motivation 2 General concept of CVaR 3 Optimization 4 Comparison Contents 1 Motivation 2 General concept of CVaR 3 Optimization 4 Comparison

More information

Coherent risk measures

Coherent risk measures Coherent risk measures Foivos Xanthos Ryerson University, Department of Mathematics Toµɛας Mαθηµατ ικὼν, E.M.Π, 11 Noɛµβρὶoυ 2015 Research interests Financial Mathematics, Mathematical Economics, Functional

More information

Expected Shortfall is not elicitable so what?

Expected Shortfall is not elicitable so what? Expected Shortfall is not elicitable so what? Dirk Tasche Bank of England Prudential Regulation Authority 1 dirk.tasche@gmx.net Modern Risk Management of Insurance Firms Hannover, January 23, 2014 1 The

More information

Handout 8: Dealing with Data Uncertainty

Handout 8: Dealing with Data Uncertainty MFE 5100: Optimization 2015 16 First Term Handout 8: Dealing with Data Uncertainty Instructor: Anthony Man Cho So December 1, 2015 1 Introduction Conic linear programming CLP, and in particular, semidefinite

More information

Reformulation of chance constrained problems using penalty functions

Reformulation of chance constrained problems using penalty functions Reformulation of chance constrained problems using penalty functions Martin Branda Charles University in Prague Faculty of Mathematics and Physics EURO XXIV July 11-14, 2010, Lisbon Martin Branda (MFF

More information

Choosing the best set of variables in regression analysis using integer programming

Choosing the best set of variables in regression analysis using integer programming DOI 10.1007/s10898-008-9323-9 Choosing the best set of variables in regression analysis using integer programming Hiroshi Konno Rei Yamamoto Received: 1 March 2007 / Accepted: 15 June 2008 Springer Science+Business

More information

Coherent Risk Measures. Acceptance Sets. L = {X G : X(ω) < 0, ω Ω}.

Coherent Risk Measures. Acceptance Sets. L = {X G : X(ω) < 0, ω Ω}. So far in this course we have used several different mathematical expressions to quantify risk, without a deeper discussion of their properties. Coherent Risk Measures Lecture 11, Optimisation in Finance

More information

Robust portfolio selection based on a joint ellipsoidal uncertainty set

Robust portfolio selection based on a joint ellipsoidal uncertainty set Optimization Methods & Software Vol. 00, No. 0, Month 2009, 1 16 Robust portfolio selection based on a joint ellipsoidal uncertainty set Zhaosong Lu* Department of Mathematics, Simon Fraser University,

More information

arxiv: v3 [math.oc] 25 Apr 2018

arxiv: v3 [math.oc] 25 Apr 2018 Problem-driven scenario generation: an analytical approach for stochastic programs with tail risk measure Jamie Fairbrother *, Amanda Turner *, and Stein W. Wallace ** * STOR-i Centre for Doctoral Training,

More information

Stochastic Optimization with Risk Measures

Stochastic Optimization with Risk Measures Stochastic Optimization with Risk Measures IMA New Directions Short Course on Mathematical Optimization Jim Luedtke Department of Industrial and Systems Engineering University of Wisconsin-Madison August

More information

Solving Classification Problems By Knowledge Sets

Solving Classification Problems By Knowledge Sets Solving Classification Problems By Knowledge Sets Marcin Orchel a, a Department of Computer Science, AGH University of Science and Technology, Al. A. Mickiewicza 30, 30-059 Kraków, Poland Abstract We propose

More information

Optimizing over coherent risk measures and nonconvexities: a robust mixed integer optimization approach

Optimizing over coherent risk measures and nonconvexities: a robust mixed integer optimization approach Optimizing over coherent risk measures and nonconvexities: a robust mixed integer optimization approach The MIT Faculty has made this article openly available. Please share how this access benefits you.

More information

Distributionally Robust Discrete Optimization with Entropic Value-at-Risk

Distributionally Robust Discrete Optimization with Entropic Value-at-Risk Distributionally Robust Discrete Optimization with Entropic Value-at-Risk Daniel Zhuoyu Long Department of SEEM, The Chinese University of Hong Kong, zylong@se.cuhk.edu.hk Jin Qi NUS Business School, National

More information

Miloš Kopa. Decision problems with stochastic dominance constraints

Miloš Kopa. Decision problems with stochastic dominance constraints Decision problems with stochastic dominance constraints Motivation Portfolio selection model Mean risk models max λ Λ m(λ r) νr(λ r) or min λ Λ r(λ r) s.t. m(λ r) µ r is a random vector of assets returns

More information

Asymptotic distribution of the sample average value-at-risk

Asymptotic distribution of the sample average value-at-risk Asymptotic distribution of the sample average value-at-risk Stoyan V. Stoyanov Svetlozar T. Rachev September 3, 7 Abstract In this paper, we prove a result for the asymptotic distribution of the sample

More information

Robust Optimization for Risk Control in Enterprise-wide Optimization

Robust Optimization for Risk Control in Enterprise-wide Optimization Robust Optimization for Risk Control in Enterprise-wide Optimization Juan Pablo Vielma Department of Industrial Engineering University of Pittsburgh EWO Seminar, 011 Pittsburgh, PA Uncertainty in Optimization

More information

Robust Growth-Optimal Portfolios

Robust Growth-Optimal Portfolios Robust Growth-Optimal Portfolios! Daniel Kuhn! Chair of Risk Analytics and Optimization École Polytechnique Fédérale de Lausanne rao.epfl.ch 4 Technology Stocks I 4 technology companies: Intel, Cisco,

More information

Multivariate Stress Testing for Solvency

Multivariate Stress Testing for Solvency Multivariate Stress Testing for Solvency Alexander J. McNeil 1 1 Heriot-Watt University Edinburgh Vienna April 2012 a.j.mcneil@hw.ac.uk AJM Stress Testing 1 / 50 Regulation General Definition of Stress

More information

Finanzrisiken. Fachbereich Angewandte Mathematik - Stochastik Introduction to Financial Risk Measurement

Finanzrisiken. Fachbereich Angewandte Mathematik - Stochastik Introduction to Financial Risk Measurement ment Convex ment 1 Bergische Universität Wuppertal, Fachbereich Angewandte Mathematik - Stochastik @math.uni-wuppertal.de Inhaltsverzeichnis ment Convex Convex Introduction ment Convex We begin with the

More information

Worst-Case Violation of Sampled Convex Programs for Optimization with Uncertainty

Worst-Case Violation of Sampled Convex Programs for Optimization with Uncertainty Worst-Case Violation of Sampled Convex Programs for Optimization with Uncertainty Takafumi Kanamori and Akiko Takeda Abstract. Uncertain programs have been developed to deal with optimization problems

More information

Size Matters: Optimal Calibration of Shrinkage Estimators for Portfolio Selection

Size Matters: Optimal Calibration of Shrinkage Estimators for Portfolio Selection Size Matters: Optimal Calibration of Shrinkage Estimators for Portfolio Selection Victor DeMiguel Department of Management Science and Operations, London Business School,London NW1 4SA, UK, avmiguel@london.edu

More information

Cross-Validation with Confidence

Cross-Validation with Confidence Cross-Validation with Confidence Jing Lei Department of Statistics, Carnegie Mellon University UMN Statistics Seminar, Mar 30, 2017 Overview Parameter est. Model selection Point est. MLE, M-est.,... Cross-validation

More information

Research Article Optimal Portfolio Estimation for Dependent Financial Returns with Generalized Empirical Likelihood

Research Article Optimal Portfolio Estimation for Dependent Financial Returns with Generalized Empirical Likelihood Advances in Decision Sciences Volume 2012, Article ID 973173, 8 pages doi:10.1155/2012/973173 Research Article Optimal Portfolio Estimation for Dependent Financial Returns with Generalized Empirical Likelihood

More information

Machine Learning for OR & FE

Machine Learning for OR & FE Machine Learning for OR & FE Regression II: Regularization and Shrinkage Methods Martin Haugh Department of Industrial Engineering and Operations Research Columbia University Email: martin.b.haugh@gmail.com

More information

On Backtesting Risk Measurement Models

On Backtesting Risk Measurement Models On Backtesting Risk Measurement Models Hideatsu Tsukahara Department of Economics, Seijo University e-mail address: tsukahar@seijo.ac.jp 1 Introduction In general, the purpose of backtesting is twofold:

More information

Time Series Models for Measuring Market Risk

Time Series Models for Measuring Market Risk Time Series Models for Measuring Market Risk José Miguel Hernández Lobato Universidad Autónoma de Madrid, Computer Science Department June 28, 2007 1/ 32 Outline 1 Introduction 2 Competitive and collaborative

More information

Sparse PCA with applications in finance

Sparse PCA with applications in finance Sparse PCA with applications in finance A. d Aspremont, L. El Ghaoui, M. Jordan, G. Lanckriet ORFE, Princeton University & EECS, U.C. Berkeley Available online at www.princeton.edu/~aspremon 1 Introduction

More information

Cross-Validation with Confidence

Cross-Validation with Confidence Cross-Validation with Confidence Jing Lei Department of Statistics, Carnegie Mellon University WHOA-PSI Workshop, St Louis, 2017 Quotes from Day 1 and Day 2 Good model or pure model? Occam s razor We really

More information

The Canonical Model Space for Law-invariant Convex Risk Measures is L 1

The Canonical Model Space for Law-invariant Convex Risk Measures is L 1 The Canonical Model Space for Law-invariant Convex Risk Measures is L 1 Damir Filipović Gregor Svindland 3 November 2008 Abstract In this paper we establish a one-to-one correspondence between lawinvariant

More information

Optimization Tools in an Uncertain Environment

Optimization Tools in an Uncertain Environment Optimization Tools in an Uncertain Environment Michael C. Ferris University of Wisconsin, Madison Uncertainty Workshop, Chicago: July 21, 2008 Michael Ferris (University of Wisconsin) Stochastic optimization

More information

Distributionally Robust Stochastic Optimization with Wasserstein Distance

Distributionally Robust Stochastic Optimization with Wasserstein Distance Distributionally Robust Stochastic Optimization with Wasserstein Distance Rui Gao DOS Seminar, Oct 2016 Joint work with Anton Kleywegt School of Industrial and Systems Engineering Georgia Tech What is

More information

MFE Financial Econometrics 2018 Final Exam Model Solutions

MFE Financial Econometrics 2018 Final Exam Model Solutions MFE Financial Econometrics 2018 Final Exam Model Solutions Tuesday 12 th March, 2019 1. If (X, ε) N (0, I 2 ) what is the distribution of Y = µ + β X + ε? Y N ( µ, β 2 + 1 ) 2. What is the Cramer-Rao lower

More information

The newsvendor problem with convex risk

The newsvendor problem with convex risk UNIVERSIDAD CARLOS III DE MADRID WORKING PAPERS Working Paper Business Economic Series WP. 16-06. December, 12 nd, 2016. ISSN 1989-8843 Instituto para el Desarrollo Empresarial Universidad Carlos III de

More information

Robust conic quadratic programming with ellipsoidal uncertainties

Robust conic quadratic programming with ellipsoidal uncertainties Robust conic quadratic programming with ellipsoidal uncertainties Roland Hildebrand (LJK Grenoble 1 / CNRS, Grenoble) KTH, Stockholm; November 13, 2008 1 Uncertain conic programs min x c, x : Ax + b K

More information

Generalized Hypothesis Testing and Maximizing the Success Probability in Financial Markets

Generalized Hypothesis Testing and Maximizing the Success Probability in Financial Markets Generalized Hypothesis Testing and Maximizing the Success Probability in Financial Markets Tim Leung 1, Qingshuo Song 2, and Jie Yang 3 1 Columbia University, New York, USA; leung@ieor.columbia.edu 2 City

More information

Testing Downside-Risk Efficiency Under Distress

Testing Downside-Risk Efficiency Under Distress Testing Downside-Risk Efficiency Under Distress Jesus Gonzalo Universidad Carlos III de Madrid Jose Olmo City University of London XXXIII Simposio Analisis Economico 1 Some key lines Risk vs Uncertainty.

More information

MFM Practitioner Module: Risk & Asset Allocation. John Dodson. February 18, 2015

MFM Practitioner Module: Risk & Asset Allocation. John Dodson. February 18, 2015 MFM Practitioner Module: Risk & Asset Allocation February 18, 2015 No introduction to portfolio optimization would be complete without acknowledging the significant contribution of the Markowitz mean-variance

More information

The Subdifferential of Convex Deviation Measures and Risk Functions

The Subdifferential of Convex Deviation Measures and Risk Functions The Subdifferential of Convex Deviation Measures and Risk Functions Nicole Lorenz Gert Wanka In this paper we give subdifferential formulas of some convex deviation measures using their conjugate functions

More information

Second Order Cone Programming, Missing or Uncertain Data, and Sparse SVMs

Second Order Cone Programming, Missing or Uncertain Data, and Sparse SVMs Second Order Cone Programming, Missing or Uncertain Data, and Sparse SVMs Ammon Washburn University of Arizona September 25, 2015 1 / 28 Introduction We will begin with basic Support Vector Machines (SVMs)

More information

Robust Fisher Discriminant Analysis

Robust Fisher Discriminant Analysis Robust Fisher Discriminant Analysis Seung-Jean Kim Alessandro Magnani Stephen P. Boyd Information Systems Laboratory Electrical Engineering Department, Stanford University Stanford, CA 94305-9510 sjkim@stanford.edu

More information

Valid Inequalities and Restrictions for Stochastic Programming Problems with First Order Stochastic Dominance Constraints

Valid Inequalities and Restrictions for Stochastic Programming Problems with First Order Stochastic Dominance Constraints Valid Inequalities and Restrictions for Stochastic Programming Problems with First Order Stochastic Dominance Constraints Nilay Noyan Andrzej Ruszczyński March 21, 2006 Abstract Stochastic dominance relations

More information

Quantifying Stochastic Model Errors via Robust Optimization

Quantifying Stochastic Model Errors via Robust Optimization Quantifying Stochastic Model Errors via Robust Optimization IPAM Workshop on Uncertainty Quantification for Multiscale Stochastic Systems and Applications Jan 19, 2016 Henry Lam Industrial & Operations

More information

Robustness in Stochastic Programs with Risk Constraints

Robustness in Stochastic Programs with Risk Constraints Robustness in Stochastic Programs with Risk Constraints Dept. of Probability and Mathematical Statistics, Faculty of Mathematics and Physics Charles University, Prague, Czech Republic www.karlin.mff.cuni.cz/~kopa

More information

AN ADAPTIVE PROCEDURE FOR ESTIMATING COHERENT RISK MEASURES BASED ON GENERALIZED SCENARIOS. Vadim Lesnevski Barry L. Nelson Jeremy Staum

AN ADAPTIVE PROCEDURE FOR ESTIMATING COHERENT RISK MEASURES BASED ON GENERALIZED SCENARIOS. Vadim Lesnevski Barry L. Nelson Jeremy Staum Proceedings of the 2006 Winter Simulation Conference L. F. Perrone, F. P. Wieland, J. Liu, B. G. Lawson, D. M. Nicol, and R. M. Fujimoto, eds. AN ADAPTIVE PROCEDURE FOR ESTIMATING COHERENT RISK MEASURES

More information

The deterministic Lasso

The deterministic Lasso The deterministic Lasso Sara van de Geer Seminar für Statistik, ETH Zürich Abstract We study high-dimensional generalized linear models and empirical risk minimization using the Lasso An oracle inequality

More information

RISK AND RELIABILITY IN OPTIMIZATION UNDER UNCERTAINTY

RISK AND RELIABILITY IN OPTIMIZATION UNDER UNCERTAINTY RISK AND RELIABILITY IN OPTIMIZATION UNDER UNCERTAINTY Terry Rockafellar University of Washington, Seattle AMSI Optimise Melbourne, Australia 18 Jun 2018 Decisions in the Face of Uncertain Outcomes = especially

More information

A Test of Cointegration Rank Based Title Component Analysis.

A Test of Cointegration Rank Based Title Component Analysis. A Test of Cointegration Rank Based Title Component Analysis Author(s) Chigira, Hiroaki Citation Issue 2006-01 Date Type Technical Report Text Version publisher URL http://hdl.handle.net/10086/13683 Right

More information

Inverse Stochastic Dominance Constraints Duality and Methods

Inverse Stochastic Dominance Constraints Duality and Methods Duality and Methods Darinka Dentcheva 1 Andrzej Ruszczyński 2 1 Stevens Institute of Technology Hoboken, New Jersey, USA 2 Rutgers University Piscataway, New Jersey, USA Research supported by NSF awards

More information

CS295: Convex Optimization. Xiaohui Xie Department of Computer Science University of California, Irvine

CS295: Convex Optimization. Xiaohui Xie Department of Computer Science University of California, Irvine CS295: Convex Optimization Xiaohui Xie Department of Computer Science University of California, Irvine Course information Prerequisites: multivariate calculus and linear algebra Textbook: Convex Optimization

More information

Mathematical Optimization Models and Applications

Mathematical Optimization Models and Applications Mathematical Optimization Models and Applications Yinyu Ye Department of Management Science and Engineering Stanford University Stanford, CA 94305, U.S.A. http://www.stanford.edu/ yyye Chapters 1, 2.1-2,

More information

Robust Performance Hypothesis Testing with the Variance. Institute for Empirical Research in Economics University of Zurich

Robust Performance Hypothesis Testing with the Variance. Institute for Empirical Research in Economics University of Zurich Institute for Empirical Research in Economics University of Zurich Working Paper Series ISSN 1424-0459 Working Paper No. 516 Robust Performance Hypothesis Testing with the Variance Olivier Ledoit and Michael

More information

MS-C1620 Statistical inference

MS-C1620 Statistical inference MS-C1620 Statistical inference 10 Linear regression III Joni Virta Department of Mathematics and Systems Analysis School of Science Aalto University Academic year 2018 2019 Period III - IV 1 / 32 Contents

More information

Robust Optimization: Applications in Portfolio Selection Problems

Robust Optimization: Applications in Portfolio Selection Problems Robust Optimization: Applications in Portfolio Selection Problems Vris Cheung and Henry Wolkowicz WatRISQ University of Waterloo Vris Cheung (University of Waterloo) Robust optimization 2009 1 / 19 Outline

More information

Worst-Case Expected Shortfall with Univariate and Bivariate. Marginals

Worst-Case Expected Shortfall with Univariate and Bivariate. Marginals Worst-Case Expected Shortfall with Univariate and Bivariate Marginals Anulekha Dhara Bikramjit Das Karthik Natarajan Abstract Worst-case bounds on the expected shortfall risk given only limited information

More information

Research Reports on Mathematical and Computing Sciences

Research Reports on Mathematical and Computing Sciences ISSN 1342-2804 Research Reports on Mathematical and Computing Sciences A Modified Algorithm for Nonconvex Support Vector Classification Akiko Takeda August 2007, B 443 Department of Mathematical and Computing

More information

Introduction to Convex Optimization

Introduction to Convex Optimization Introduction to Convex Optimization Daniel P. Palomar Hong Kong University of Science and Technology (HKUST) ELEC5470 - Convex Optimization Fall 2018-19, HKUST, Hong Kong Outline of Lecture Optimization

More information

Robust Portfolio Risk Minimization Using the Graphical Lasso

Robust Portfolio Risk Minimization Using the Graphical Lasso Robust Portfolio Risk Minimization Using the Graphical Lasso Tristan Millington & Mahesan Niranjan Department of Electronics and Computer Science University of Southampton Highfield SO17 1BJ, Southampton,

More information

CVaR and Examples of Deviation Risk Measures

CVaR and Examples of Deviation Risk Measures CVaR and Examples of Deviation Risk Measures Jakub Černý Department of Probability and Mathematical Statistics Stochastic Modelling in Economics and Finance November 10, 2014 1 / 25 Contents CVaR - Dual

More information

ORIGINS OF STOCHASTIC PROGRAMMING

ORIGINS OF STOCHASTIC PROGRAMMING ORIGINS OF STOCHASTIC PROGRAMMING Early 1950 s: in applications of Linear Programming unknown values of coefficients: demands, technological coefficients, yields, etc. QUOTATION Dantzig, Interfaces 20,1990

More information

Robustness of data-driven CVaR optimization using smoothing technique

Robustness of data-driven CVaR optimization using smoothing technique Robustness of data-driven CVaR optimization using smoothing technique by Johnny Chow Aresearchpaper presented to the University of Waterloo in partial fulfillment of the requirement for the degree of Master

More information

Machine Learning And Applications: Supervised Learning-SVM

Machine Learning And Applications: Supervised Learning-SVM Machine Learning And Applications: Supervised Learning-SVM Raphaël Bournhonesque École Normale Supérieure de Lyon, Lyon, France raphael.bournhonesque@ens-lyon.fr 1 Supervised vs unsupervised learning Machine

More information

Ordinary Least Squares Linear Regression

Ordinary Least Squares Linear Regression Ordinary Least Squares Linear Regression Ryan P. Adams COS 324 Elements of Machine Learning Princeton University Linear regression is one of the simplest and most fundamental modeling ideas in statistics

More information

MFM Practitioner Module: Risk & Asset Allocation. John Dodson. February 4, 2015

MFM Practitioner Module: Risk & Asset Allocation. John Dodson. February 4, 2015 & & MFM Practitioner Module: Risk & Asset Allocation February 4, 2015 & Meucci s Program for Asset Allocation detect market invariance select the invariants estimate the market specify the distribution

More information

ECON4515 Finance theory 1 Diderik Lund, 5 May Perold: The CAPM

ECON4515 Finance theory 1 Diderik Lund, 5 May Perold: The CAPM Perold: The CAPM Perold starts with a historical background, the development of portfolio theory and the CAPM. Points out that until 1950 there was no theory to describe the equilibrium determination of

More information

R = µ + Bf Arbitrage Pricing Model, APM

R = µ + Bf Arbitrage Pricing Model, APM 4.2 Arbitrage Pricing Model, APM Empirical evidence indicates that the CAPM beta does not completely explain the cross section of expected asset returns. This suggests that additional factors may be required.

More information

IV. Matrix Approximation using Least-Squares

IV. Matrix Approximation using Least-Squares IV. Matrix Approximation using Least-Squares The SVD and Matrix Approximation We begin with the following fundamental question. Let A be an M N matrix with rank R. What is the closest matrix to A that

More information

A Second Full-Newton Step O(n) Infeasible Interior-Point Algorithm for Linear Optimization

A Second Full-Newton Step O(n) Infeasible Interior-Point Algorithm for Linear Optimization A Second Full-Newton Step On Infeasible Interior-Point Algorithm for Linear Optimization H. Mansouri C. Roos August 1, 005 July 1, 005 Department of Electrical Engineering, Mathematics and Computer Science,

More information

COHERENT APPROACHES TO RISK IN OPTIMIZATION UNDER UNCERTAINTY

COHERENT APPROACHES TO RISK IN OPTIMIZATION UNDER UNCERTAINTY COHERENT APPROACHES TO RISK IN OPTIMIZATION UNDER UNCERTAINTY Terry Rockafellar University of Washington, Seattle University of Florida, Gainesville Goal: a coordinated view of recent ideas in risk modeling

More information

Generalized quantiles as risk measures

Generalized quantiles as risk measures Generalized quantiles as risk measures Bellini, Klar, Muller, Rosazza Gianin December 1, 2014 Vorisek Jan Introduction Quantiles q α of a random variable X can be defined as the minimizers of a piecewise

More information

Multivariate Stress Scenarios and Solvency

Multivariate Stress Scenarios and Solvency Multivariate Stress Scenarios and Solvency Alexander J. McNeil 1 1 Heriot-Watt University Edinburgh Croatian Quants Day Zagreb 11th May 2012 a.j.mcneil@hw.ac.uk AJM Stress Testing 1 / 51 Regulation General

More information

The Value of Adaptability

The Value of Adaptability The Value of Adaptability Dimitris Bertsimas Constantine Caramanis September 30, 2005 Abstract We consider linear optimization problems with deterministic parameter uncertainty. We consider a departure

More information

Nonparametric estimation of tail risk measures from heavy-tailed distributions

Nonparametric estimation of tail risk measures from heavy-tailed distributions Nonparametric estimation of tail risk measures from heavy-tailed distributions Jonthan El Methni, Laurent Gardes & Stéphane Girard 1 Tail risk measures Let Y R be a real random loss variable. The Value-at-Risk

More information

Information Choice in Macroeconomics and Finance.

Information Choice in Macroeconomics and Finance. Information Choice in Macroeconomics and Finance. Laura Veldkamp New York University, Stern School of Business, CEPR and NBER Spring 2009 1 Veldkamp What information consumes is rather obvious: It consumes

More information

X

X Correlation: Pitfalls and Alternatives Paul Embrechts, Alexander McNeil & Daniel Straumann Departement Mathematik, ETH Zentrum, CH-8092 Zürich Tel: +41 1 632 61 62, Fax: +41 1 632 15 23 embrechts/mcneil/strauman@math.ethz.ch

More information

Analysis of Fast Input Selection: Application in Time Series Prediction

Analysis of Fast Input Selection: Application in Time Series Prediction Analysis of Fast Input Selection: Application in Time Series Prediction Jarkko Tikka, Amaury Lendasse, and Jaakko Hollmén Helsinki University of Technology, Laboratory of Computer and Information Science,

More information

Birgit Rudloff Operations Research and Financial Engineering, Princeton University

Birgit Rudloff Operations Research and Financial Engineering, Princeton University TIME CONSISTENT RISK AVERSE DYNAMIC DECISION MODELS: AN ECONOMIC INTERPRETATION Birgit Rudloff Operations Research and Financial Engineering, Princeton University brudloff@princeton.edu Alexandre Street

More information

Least squares under convex constraint

Least squares under convex constraint Stanford University Questions Let Z be an n-dimensional standard Gaussian random vector. Let µ be a point in R n and let Y = Z + µ. We are interested in estimating µ from the data vector Y, under the assumption

More information

A full-newton step infeasible interior-point algorithm for linear programming based on a kernel function

A full-newton step infeasible interior-point algorithm for linear programming based on a kernel function A full-newton step infeasible interior-point algorithm for linear programming based on a kernel function Zhongyi Liu, Wenyu Sun Abstract This paper proposes an infeasible interior-point algorithm with

More information

Probabilistic Regression Using Basis Function Models

Probabilistic Regression Using Basis Function Models Probabilistic Regression Using Basis Function Models Gregory Z. Grudic Department of Computer Science University of Colorado, Boulder grudic@cs.colorado.edu Abstract Our goal is to accurately estimate

More information

Chap 1. Overview of Statistical Learning (HTF, , 2.9) Yongdai Kim Seoul National University

Chap 1. Overview of Statistical Learning (HTF, , 2.9) Yongdai Kim Seoul National University Chap 1. Overview of Statistical Learning (HTF, 2.1-2.6, 2.9) Yongdai Kim Seoul National University 0. Learning vs Statistical learning Learning procedure Construct a claim by observing data or using logics

More information

Efficient optimization of the reward-risk ratio with polyhedral risk measures

Efficient optimization of the reward-risk ratio with polyhedral risk measures Math Meth Oper Res (2017) 86:625 653 https://doi.org/10.1007/s00186-017-0613-1 ORIGINAL ARTICLE Efficient optimization of the reward-risk ratio with polyhedral risk measures Wlodzimierz Ogryczak 1 Michał

More information

RESEARCH REPORT SPARSE ROBUST PORTFOLIO OPTIMIZATION VIA NLP REGULARIZATIONS. No December 2016

RESEARCH REPORT SPARSE ROBUST PORTFOLIO OPTIMIZATION VIA NLP REGULARIZATIONS. No December 2016 Akademie věd České republiky Ústav teorie informace a automatizace Czech Academy of Sciences Institute of Information Theory and Automation RESEARCH REPORT Martin BRANDA, Michal ČERVINKA, Alexandra SCHWARTZ

More information

Combinatorial Data Mining Method for Multi-Portfolio Stochastic Asset Allocation

Combinatorial Data Mining Method for Multi-Portfolio Stochastic Asset Allocation Combinatorial for Stochastic Asset Allocation Ran Ji, M.A. Lejeune Department of Decision Sciences July 8, 2013 Content Class of Models with Downside Risk Measure Class of Models with of multiple portfolios

More information

Empirical properties of large covariance matrices in finance

Empirical properties of large covariance matrices in finance Empirical properties of large covariance matrices in finance Ex: RiskMetrics Group, Geneva Since 2010: Swissquote, Gland December 2009 Covariance and large random matrices Many problems in finance require

More information

EE 227A: Convex Optimization and Applications April 24, 2008

EE 227A: Convex Optimization and Applications April 24, 2008 EE 227A: Convex Optimization and Applications April 24, 2008 Lecture 24: Robust Optimization: Chance Constraints Lecturer: Laurent El Ghaoui Reading assignment: Chapter 2 of the book on Robust Optimization

More information

Selected Examples of CONIC DUALITY AT WORK Robust Linear Optimization Synthesis of Linear Controllers Matrix Cube Theorem A.

Selected Examples of CONIC DUALITY AT WORK Robust Linear Optimization Synthesis of Linear Controllers Matrix Cube Theorem A. . Selected Examples of CONIC DUALITY AT WORK Robust Linear Optimization Synthesis of Linear Controllers Matrix Cube Theorem A. Nemirovski Arkadi.Nemirovski@isye.gatech.edu Linear Optimization Problem,

More information

A Note on Robust Representations of Law-Invariant Quasiconvex Functions

A Note on Robust Representations of Law-Invariant Quasiconvex Functions A Note on Robust Representations of Law-Invariant Quasiconvex Functions Samuel Drapeau Michael Kupper Ranja Reda October 6, 21 We give robust representations of law-invariant monotone quasiconvex functions.

More information

CORC Technical Report TR Robust portfolio selection problems

CORC Technical Report TR Robust portfolio selection problems CORC Technical Report TR-2002-03 Robust portfolio selection problems D. Goldfarb G. Iyengar Submitted: Dec. 26, 200. Revised: May 25th, 2002, July 24th, 2002 Abstract In this paper we show how to formulate

More information