Abstract
Robust portfolio selection explicitly incorporates a model of parameter uncertainty in the problem formulation, and optimizes for the worst-case scenario. We consider robust mean–variance portfolio selection involving a trade-off between the worst-case utility and the worst-case regret, or the largest difference between the best utility achievable under the model and that achieved by a given portfolio. While optimizing for the worst-case utility may yield an overly pessimistic portfolio, optimizing for the worst-case regret may result in a complete loss of robustness; we theoretically demonstrate this point. Robust trade-off portfolio compromises these two extremes, enabling more informative selections. We show that, under a widely used ellipsoidal uncertainty model, the entire optimal trade-off curve can be found via solving a series of semidefinite programs (SDPs). We then extend the model to handle the union of finitely many ellipsoids, and show that trade-off analysis under this quite general uncertainty model also reduces to a series of SDPs. For more general uncertainties, we propose an iterative algorithm based on the cutting-set method. Under the finite-union-of-ellipsoids model, this algorithm offers an alternative to the SDP in exploring the optimal trade-off curve. We illustrate the promises of the trade-off portfolios by using historical stock returns data.
Similar content being viewed by others
References
Alizadeh F, Goldfarb D (2003) Second-order cone programming. Math Program 95(1):3–51
Ben-Tal A, El Ghaoui L, Nemirovski A (2009) Robust optimization. Princeton University Press, Princeton
Ben-Tal A, Bertsimas D, Brown D (2010) A soft robust model for optimization under ambiguity. Oper Res 58(4–2):1220–1234
Bertsimas D, Brown D, Caramanis C (2011) Theory and applications of robust optimization. SIAM Rev 53(3):464–501
Black F, Litterman R (1992) Global portfolio optimization. Financ Anal J 48(5):28–43
Boyd S, Vandenberghe L (2004) Convex optimization. Cambridge University Press, Cambridge
Boyd S, El Ghaoui L, Feron E, Balakrishnan V (1994) Linear matrix inequalities in system and control theory. Society for Industrial and Applied Mathematics, Philadelphia
Ceria S, Stubbs R (2006) Incorporating estimation errors into portfolio selection: robust portfolio construction. J Asset Manag 7(2):109–127
Cornuejols G, Tütüncü R (2006) Optimization methods in finance. Cambridge University Press, Cambridge
El Ghaoui L, Oks M, Oustry F (2003) Worst-case value-at-risk and robust portfolio optimization: a conic programming approach. Oper Res 51(4):543–556
Fabozzi F, Kolm P, Pachamanova D, Focardi S (2007) Robust portfolio optimization. J Portf Manag 34(1):40–48
Garlappi L, Uppal R, Wang T (2007) Portfolio selection with parameter and model uncertainty: a multi-prior approach. Rev Financ Stud 20(1):41–81
Glasserman P, Xu X (2013) Robust portfolio control with stochastic factor dynamics. Oper Res 61(4):874–893
Glasserman P, Xu X (2014) Robust risk measurement and model risk. Quant Finance 14(1):29–58
Goldfarb D, Iyengar G (2003) Robust portfolio selection problems. Math Oper Res 28(1):1–38
Grant M, Boyd S (2008) Graph implementations for nonsmooth convex programs. In: Blondel V, Boyd S, Kimura H (eds) Recent advances in learning and control (a tribute to M. Vidyasagar). Springer, Berlin, pp 95–110
Grant M, Boyd S (2014) CVX: matlab software for disciplined convex programming, version 2.1. http://cvxr.com/cvx. Accessed 14 Jan 2020
Hauser R, Krishnamurthy V, Tütüncü R (2013) Relative robust portfolio optimization. arXiv preprint arXiv:1305.0144
Hettich R, Kortanek K (1993) Semi-infinite programming: theory, methods, and applications. SIAM Rev 35(5):380–429
Huang D, Zhu SS, Fabozzi F, Fukushima M (2010) Portfolio selection under distributional uncertainty: a relative robust CVaR approach. Eur J Oper Res 203(3):185–194
Iancu DA, Trichakis N (2013) Pareto efficiency in robust optimization. Manag Sci 60(1):130–147
Iyengar G, Ma AKC (2010) A robust optimization approach to pension fund management. J Asset Manag 11(2–3):163–177
Jorion P (1979) Bayes-Stein estimation for portfolio analysis. J Financ Quant Anal 21(3):279–292
Kim JH, Kim WC, Fabozzi FJ (2014) Recent developments in robust portfolios with a worst-case approach. J Optim Theory Appl 161(1):103–121
Kim WC, Kim JH, Mulvey JM, Fabozzi FJ (2015) Focusing on the worst state for robust investing. Int Rev Financ Anal 39:19–31
Künsch H (1989) The jackknife and the bootstrap for general stationary observations. Ann Stat 17:1217–1241
Kwon RH, Wu D (2017) Factor-based robust index tracking. Optim Eng 18(2):443–466
Lim A, Shanthikumar J, Watewai T (2011) Robust asset allocation with benchmarked objectives. Math Finance 21(4):643–679
Lim A, Shanthikumar J, Vahn GY (2012) Robust portfolio choice with learning in the framework of regret: single-period case. Manag Sci 58(9):1732–1746
Liu R, Singh K (1992) Moving blocks jackknife and bootstrap capture weak dependence. In: LePage R, Billard L (eds) Exploring the limits of bootstrap. Wiley, Hoboken, pp 225–248
Lobo M, Boyd S (2000) The worst-case risk of a portfolio. Unpublished manuscript. https://web.stanford.edu/~boyd/papers/pdf/risk_bnd.pdf. Accessed 14 Jan 2020
Lobo M, Vandenberghe L, Boyd S, Lebret H (1998) Applications of second-order cone programming. Linear Algebra Appl 284:193–228
Lotfi S, Zenios SA (2018) Robust VaR and CVaR optimization under joint ambiguity in distributions, means, and covariances. Eur J Oper Res 269(2):556–576
Lotfi S, Salahi M, Mehrdoust F (2017) Adjusted robust mean-value-at-risk model: less conservative robust portfolios. Optim Eng 18(2):467–497
Lutgens F, Schotman PC (2010) Robust portfolio optimisation with multiple experts. Rev Finance 14:343–383
Markowitz H, Usmen N (2003) Resampled frontiers versus diffuse Bayes: an experiment. J Invest Manag 1(4):9–25
Michaud R (1989) The Markowitz optimization enigma: Is ‘optimized’ optimal? Financ Anal J 45(1):31–42
Michaud R, Michaud R (2008) Estimation error and portfolio optimization: a resampling approach. J Invest Manag 6(1):8–28
Mutapcic A, Boyd S (2009) Cutting-set methods for robust convex optimization with worst-case oracles. Optim Methods Softw 24(3):381–406
Natarajan K, Sim M, Uichanco J (2010) Tractable robust expected utility and risk models for portfolio optimization. Math Finance 20(4):695–731
Peng Y, Freund R (2004) Computation of minimum volume covering ellipsoids. Oper Res 52(5):690–706
Politis D, Romano R (1992) A circular block-resampling procedure for stationary data. In: LePage R, Billard L (eds) Exploring the limits of bootstrap. Wiley, Hoboken, pp 263–270
Politis D, Romano R (1994) The stationary bootstrap. J Am Stat Assoc 89:1303–1313
Rustem B, Becker R, Marty W (2000) Robust min–max portfolio strategies for rival forecast and risk scenarios. J Econ Dyn Control 24(11–12):1591–1621
Toh K, Todd M, Tütüncü R (2009) SDPT3 Version 4.0. A Matlab software for semidefinite-quadratic-linear programming. www.math.nus.edu.sg/~mattohkc/sdpt3.html. Accessed 14 Jan 2020
Tütüncü R, Koenig M (2004) Robust asset allocation. Ann Oper Res 132(1–4):157–187
Vandenberghe L, Boyd S, Wu SP (1998) Determinant maximization with linear matrix inequality constraints. SIAM J Matrix Anal Appl 19(2):499–533
Zhu S, Fukushima M (2009) Worst-case conditional value-at-risk with application to robust portfolio management. Oper Res 57(5):1155–1168
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix 1: Proofs
1.1 Appendix 1.1: Proof of Theorem 1
1.1.1 Part (a)
With the introduction of an auxiliary variable t, maximization of \(U_{\text {wc}}(w)\) (5) can be written
Then the Lagrangian of this concave program is
where \(\lambda\) is a measure defined on the Borel \(\sigma\)-field of \({\mathcal {U}}\). Then the Lagrangian dual function of problem (5) is
i.e., \(g(\lambda )\) is finite if and only if \(\lambda\) is a probability measure on \({\mathcal {U}}\). Note
where \(\bar{U}(\mu ,\varSigma )=\sup _{z:{\mathbf {1}}^Tz=1}U(z,\mu ,\varSigma )\); \(\mathbf {E}_{\lambda }(\cdot )\) denotes the expectation with respect to probability measure \(\lambda\), \(\mu _{\lambda } = \mathbf {E}_{\lambda }[\mu ]\), and \(\varSigma _{\lambda } = \mathbf {E}_{\lambda }[\varSigma ]\).
Thus the dual optimization problem of (5) is to find \(\lambda\) for which \((\mu _{\lambda }, \varSigma _{\lambda })\) solves
where the last equation is due to the strong duality. The latter holds because Slater’s condition for the primal (5) is satisfied: \(\inf _{(\mu ,\varSigma )\in {\mathcal {U}}} U(w,\mu ,\varSigma ) > -\infty\) since \({\mathcal {U}}={\mathcal {M}}\times \mathcal {S}\) is compact by the assumption; note that the last optimization problem is precisely of the same form as problem (5).
Under the given separable uncertainty model, it is easy to verify that the optimal tuple \((w^{\star },\mu _{\lambda ^{\star }}, \varSigma _{\lambda ^{\star }})\) satisfies
which implies that under the optimal probability measure \(\lambda ^{\star }\), \(\mu _{\lambda ^{\star }}=\mathbf {E}_{\lambda ^{\star }}[\mu ]\) is a point on the boundary of \({\mathcal {M}}\). This can only occur when \(\lambda ^{\star }\) is a product measure of a Dirac measure that puts all the mass at a boundary point of \({\mathcal {M}}\) and any probability measure on \(\mathcal {S}\).
1.1.2 Part (b)
Similarly to part (a), write the problem of minimizing \(R_{\text {wc}}(w)\) (6) as
whose Lagrangian is
for \(w\in {\mathbb {R}}^n\), \(t\in {\mathbb {R}}\), and \(\lambda\) a measure on the Borel \(\sigma\)-field of \({\mathcal {U}}\). The associated dual optimization problem is to maximize
for a probability measure \(\lambda\) on \({\mathcal {U}}\). The last inequality is due to the convexity of \(\bar{U}\) in \((\mu , \varSigma )\) and Jensen’s inequality.
Slater’s condition is satisfied for (P) since \(\sup _{(\mu ,\varSigma )\in {\mathcal {U}}}\{\bar{U}(\mu ,\varSigma ) - U(w,\mu ,\varSigma )\} < \infty\), which in turn is because of the assumed compactness of \({\mathcal {U}}={\mathcal {M}}\times {\mathcal {S}}\). Thus the strong duality holds, which translates
for the optimal tuple \((w^{\star }, t^{\star }, \lambda ^{\star })\). The fourth inequality shows that the opitimal \(\lambda ^{\star }\) is a degenerate probability measure that puts all the masses on the maximizers of \(\bar{U}(\mu , \varSigma ) - U(w^{\star },\mu ,\varSigma )\), where \(w^{\star }=\text {arg max}_{w:{\mathbf {1}}^Tw=1} U(w, \mu _{\lambda ^{\star }}, \varSigma _{\lambda ^{\star }})\).
Hence we now consider the problem
Recall \({\mathcal {M}}=\{{\bar{\mu }} + Pu: \Vert u\Vert _2 \le 1\}\) and that \(\bar{U}(\mu ,\varSigma )\) attains a closed form:
where
Also since \(w^{\star }\) maximizes \(U(w, \mu _{\lambda ^{\star }}, \varSigma _{\lambda ^{\star }})\), it follows that
(see Sect. 3.2).
If \(\mathcal {S}\) is singleton, i.e., \({\mathcal {S}}=\{{\bar{\varSigma }}\}\), then
since \(\mathbf {E}_{\lambda }[f(\varSigma )]=f({\bar{\varSigma }})\) for any function f of \(\varSigma\); \(\pi\) is the probability measure on \(\{u: \Vert u\Vert \le 1\}\) induced from \(\lambda\) by the transformation \(\mu ={\bar{\mu }} + Pu\), and \(u_{\max }\) solves the following optimization problem
Note that \(-u_{\max }\) also solves (34). The inequality in the last line holds with equality if and only if \(\pi\) puts equal masses on \(u_{\max }\) and \(-u_{\max }\). Thus \(\lambda ^{\star }\) is a degenerate probability measure that puts equal masses on \(({\bar{\mu }}\pm Pu_{\max }, {\bar{\varSigma }})\), yielding \(\mu _{\lambda ^{\star }}=\mathbf {E}_{\lambda ^{\star }}[\mu ]={\bar{\mu }}\). Trivially \(\varSigma _{\lambda ^{\star }} = {\bar{\varSigma }}\). Therefore, we have
as claimed.
For general \(\mathcal {S}\), if the worst-case regret criterion
has a unique maximizer \(\varSigma ^{\star }(w)\) for each weight vector w, then problem (R) has a unique maximizer \(\varSigma ^{\star }=\varSigma ^{\star }(w^{\star })\) and the optimal probability measure \(\lambda ^{\star }\) puts all the masses on \(\varSigma ^{\star }\) in \(\mathcal {S}\). Thus it suffices to consider probability measures \(\lambda\) on \({\mathcal {M}}\times \{\varSigma ^{\star }\}\) in maximizing \(g(\lambda )\) in (D). In particular, for these \(\lambda\) there holds \(\mathbf {E}_{\lambda }[f(\varSigma )]=f(\varSigma ^{\star })\) for any function f. Thus by the same argument as the singleton case above, \(\lambda ^{\star }\) is a degenerate probability measure that puts equal masses on \(({\bar{\mu }}\pm Pu_{\max }^{\star }, \varSigma ^{\star })\), where \(u_{\max }^{\star }\) is a solution to problem (34) with \({\bar{\varSigma }}\) replaced by \(\varSigma ^{\star }\). This yields \(\mu _{\lambda ^{\star }}=\mathbf {E}_{\lambda ^{\star }}[\mu ]={\bar{\mu }}\) and \(w^{\star } = \beta (\varSigma ^{\star }) + \varGamma (\varSigma ^{\star }){\bar{\mu }}\), as desired.
1.2 Appendix 1.2: Proof of Theorem 2
Assume that \(\lambda _1 = \cdots = \lambda _r > \lambda _{r+1}\) for some \(r \ge 1\), and let \(V_1\) denote the first r columns of the V. The Lagrangian of (18) is
The dual function is thus
interpreting \(g(\lambda _1)=\infty\) if \(V_1^T\nu \ne 0\). Note \({\bar{\nu }}_1 = \cdots = {\bar{\nu }}_r = 0\) if \(V_1^T\nu = 0\). The dual optimization problem is to minimize \(g(\lambda )\), which is equivalent to the SDP (19) in that \(t^{\star } = g(\lambda ^{\star })\).
For \(\lambda > \lambda _1\), \(g(\lambda )\) is strictly convex and also strictly increasing for sufficiently large \(\lambda\), thus \(\lambda ^{\star } > \lambda _1\) if the equation
has a root on \(\lambda > \lambda _1\), otherwise \(g'(\lambda ) \ge 0\) and \(\lambda ^{\star }=\lambda _1\).
From (35), \(u^{\star } = (1/2)(\lambda ^{\star } I -P^TBP)^{\dagger }\nu + v\), where \(v \in {\mathcal {N}}(\lambda ^{\star } I - P^TBP)\). If \(\lambda ^{\star } > \lambda _1\), \(v=0\) and \(u^{\star } = (1/2)(\lambda ^{\star } I -P^TBP)^{-1}\nu = (1/2)V{{\,\mathrm{diag}\,}}{(1/(\lambda ^{\star }-\lambda _1),\ldots ,1/(\lambda ^{\star }-\lambda _n))}V^T\nu\). Thus \(\Vert u^{\star }\Vert _2^2 = (1/4)\sum _{i=1}^n {\bar{\nu }}_i^2/(\lambda ^{\star }-\lambda _i)^2 = 1\) due to (36), which is also confirmed by the Kuhn–Karush–Tucker (KKT) condition \(\lambda ^{\star }(\Vert u^{\star }\Vert _2^2-1) = 0\) for (18). If \(\lambda ^{\star } = \lambda _1\), then \(\lambda _1v=P^TBPv\) hence \(v=kv_{\max }\) for some scalar k. This also implies that \({\bar{\nu }}_1 = \cdots = {\bar{\nu }}_r = 0\). From the KKT condition, \(\Vert u^{\star }\Vert _2^2 = (1/4)\sum _{i=r+1}^n{\bar{\nu }}_i^2/(\lambda _1-\lambda _i)^2 + k^2=1\), yielding \(k=\pm \sqrt{1-(1/4)\sum _{i=r+1}^n{\bar{\nu }}_i^2/(\lambda _1-\lambda _i)^2}\). The latter is well-defined since \(g'(\lambda _1) \ge 0\).
Appendix 2: Performance metrics
We first show how the wealth varies under the periodic rebalancing strategy described above and describe some descriptive statistics used to assess its performance. A dollar invested to asset i at the beginning of the \(j-1\)st period grows to \(d^{(j-1)}_i = \prod ^{N_{\mathrm {estim}}+jL}_{t=N_{\mathrm {estim}}+1+(j-1)L}(1+r_{it})\). The initial portfolio \(w^{(j-1)}\) held at the beginning of the \(j-1\)st period is therefore changed to the portfolio \(\tilde{w}^{(j-1)}\) with weights
For the first period, we take \({{\tilde{w}}}^{(0)}=0\), i.e., the initial holdings of the assets are zero.
At the start date of the jth period, asset weights are changed from \({{\tilde{w}}}^{(j-1)}\) to \(w^{(j)}\). The turnover ratio, i.e., the relative amount of the purchase and sale of securities in a fund’s portfolio, associated with the change is computed as
The percentage of the portfolio that is bought and sold to exchange for other stocks is \(100{\mathrm {TO}}_j\%\).
Portfolio rebalancing incurs transaction costs. Let the cost to buy or sell one share of stock i be \(\eta _i\). The transaction cost \({\mathrm {TC}}_j\) due to the rebalancing at the beginning of the jth period is given by
Then the (normalized) wealth grows according to the recursion
for \(t = N_{\mathrm {estim}}, \ldots , N_{\mathrm {estim}}+KL\), with \(W_{N_{\mathrm {estim}}} = 1\).
We are interested in several performance metrics which include the annualized return, risk, and Sharpe ratio (SR) of the normalized wealth time series. (The annualized Sharpe ratio is the ratio of the annualized excess expected return of relative to the risk-free return \(\mu _{\mathrm {rf}}\)) Another performance metric of interest is the maximum drawdown, defined as
i.e., the largest drawdown of the wealth experienced by the trading strategy over the entire trading periods.
Appendix 3: Portfolio rebalancing data basics
Rights and permissions
About this article
Cite this article
Won, JH., Kim, SJ. Robust trade-off portfolio selection. Optim Eng 21, 867–904 (2020). https://doi.org/10.1007/s11081-020-09485-z
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11081-020-09485-z