Robust portfolio optimization under multiperiod mean-standard deviation criterion

Transcription

1 Robust portfolio optimization under multiperiod mean-standard deviation criterion Spiridon Penev 1 Pavel Shevchenko 2 Wei Wu 1 1 The University of New South Wales, Sydney Australia 2 Macquarie University Sydney, Australia Wollongong, July 2017

2 Outline 1 Introduction 2 Motivation Technicalities 3 Main result 4 Numerical Examples. Quantification of Model Risk with Empirical Data. 5 Conclusion. 6 References

3 Introduction We look into model risk, defined as the loss due to uncertainty of the underlying distribution of the return of the assets in a portfolio. Uncertainty is measured by the Kullback-Leibler divergence (more generally by α-divergence). We show that in the worst case scenario, the optimal robust strategy can be obtained in a semi-analytical form. As a consequence, we quantify model risk. By combining with a Monte Carlo approach, the optimal robust strategy can be calculated numerically. Numerically compare performance of the robust strategy with the optimal non-robust strategy, the latter being calculated at a nominal distribution quantify model risk.

4 Introduction We look into model risk, defined as the loss due to uncertainty of the underlying distribution of the return of the assets in a portfolio. Uncertainty is measured by the Kullback-Leibler divergence (more generally by α-divergence). We show that in the worst case scenario, the optimal robust strategy can be obtained in a semi-analytical form. As a consequence, we quantify model risk. By combining with a Monte Carlo approach, the optimal robust strategy can be calculated numerically. Numerically compare performance of the robust strategy with the optimal non-robust strategy, the latter being calculated at a nominal distribution quantify model risk.

5 Introduction We look into model risk, defined as the loss due to uncertainty of the underlying distribution of the return of the assets in a portfolio. Uncertainty is measured by the Kullback-Leibler divergence (more generally by α-divergence). We show that in the worst case scenario, the optimal robust strategy can be obtained in a semi-analytical form. As a consequence, we quantify model risk. By combining with a Monte Carlo approach, the optimal robust strategy can be calculated numerically. Numerically compare performance of the robust strategy with the optimal non-robust strategy, the latter being calculated at a nominal distribution quantify model risk.

6 Introduction We look into model risk, defined as the loss due to uncertainty of the underlying distribution of the return of the assets in a portfolio. Uncertainty is measured by the Kullback-Leibler divergence (more generally by α-divergence). We show that in the worst case scenario, the optimal robust strategy can be obtained in a semi-analytical form. As a consequence, we quantify model risk. By combining with a Monte Carlo approach, the optimal robust strategy can be calculated numerically. Numerically compare performance of the robust strategy with the optimal non-robust strategy, the latter being calculated at a nominal distribution quantify model risk.

7 Introduction We look into model risk, defined as the loss due to uncertainty of the underlying distribution of the return of the assets in a portfolio. Uncertainty is measured by the Kullback-Leibler divergence (more generally by α-divergence). We show that in the worst case scenario, the optimal robust strategy can be obtained in a semi-analytical form. As a consequence, we quantify model risk. By combining with a Monte Carlo approach, the optimal robust strategy can be calculated numerically. Numerically compare performance of the robust strategy with the optimal non-robust strategy, the latter being calculated at a nominal distribution quantify model risk.

8 Motivation Which criterion should an investor choose to optimize portfolio wealth? Markowitz (1952): max u (E(W) κ Var(W)) Others: Safety-first: min u P(W d), targeting particular wealth level: max u E(W Ŵ 2 )), etc. One more: Mean-st. deviation (MSD) criterion: max u (E(W) κ Var(W)). Why the latter: for elliptically distributed returns, optimizing a risk measure form the whole class of Translation-invariant and positive-homogeneous risk measures (TIPH) is equivalent to optimizaing the MSD criterion (with a suitable κ).

9 Motivation Which criterion should an investor choose to optimize portfolio wealth? Markowitz (1952): max u (E(W) κ Var(W)) Others: Safety-first: min u P(W d), targeting particular wealth level: max u E(W Ŵ 2 )), etc. One more: Mean-st. deviation (MSD) criterion: max u (E(W) κ Var(W)). Why the latter: for elliptically distributed returns, optimizing a risk measure form the whole class of Translation-invariant and positive-homogeneous risk measures (TIPH) is equivalent to optimizaing the MSD criterion (with a suitable κ).

10 Motivation Which criterion should an investor choose to optimize portfolio wealth? Markowitz (1952): max u (E(W) κ Var(W)) Others: Safety-first: min u P(W d), targeting particular wealth level: max u E(W Ŵ 2 )), etc. One more: Mean-st. deviation (MSD) criterion: max u (E(W) κ Var(W)). Why the latter: for elliptically distributed returns, optimizing a risk measure form the whole class of Translation-invariant and positive-homogeneous risk measures (TIPH) is equivalent to optimizaing the MSD criterion (with a suitable κ).

11 Motivation However: i) the joint distribution of the assets is unkonwn (e.g., slightly deviating from a nominal multivariate normal). ii) changes of the distribution over time (in multiple periods) (need dynamic approach). The deviation can be measured by: KL divergence, α-divergence.. Try to quantify the intuition: big divergence significant impact on an optimal investment decision that is calculated under the nominal distribution". Ultimate goal: If distributional assumptions are violated only slightly use the optimal investment strategy under the nominal model" (since robust approach may deliver too pessimistic strategy). Ideally: ball of radius η 0 around the nominal model: stay with the nominal inside", switch to robust outside". Need to quantify model risk from risk management perspective.

12 Motivation Technicalities Outline 1 Introduction 2 Motivation Technicalities 3 Main result 4 Numerical Examples. Quantification of Model Risk with Empirical Data. 5 Conclusion. 6 References

13 Motivation Technicalities Using KL divergence: reasonable for short term re-balancing (daily or weekly). Big advantage: closed form worst case distribution is available allows us to get the optimal (time-consistent) strategy under the worst case distribution in a semi-analytical form. Our previous work (BGPW, Automatica 2016) delivers the optimal strategy under the nominal model can compare performance under worst case scenario.

14 Motivation Technicalities Problem Formulation - d > 1 risky assets; fixed investment horizon [0,N]; return of each asset over the nth period [n,n+1], n=0,...,n 1: as r n+1 =(rn+1 1,...,rd n+1 ), withe rn+1 i, i = 1,...,d. Filtered probability space (Ω,F,(F n ),P) with the sample space Ω, the sigma-algebra F, filtration (F n ), probab. measure P, sigma-algebra F n = σ(r m,1 m n). Return vector r n+1 : E( r n+1 2 )<. At time n=0,...,n 1: re-balance using strategy u =(u 0,...,u N 1 ), all u n, taking values in U: { } U = u R d : 1 u = 1, for i = 1,...,d. For m>0, we use U m to denote the set of admissible sub-strategies u m =(u n ) n m

15 Motivation Technicalities Let W n denote the wealth at time n (n=0,...,n). Assume: W n and r n+1 are independent. During [n,n+1], wealth changes: W n+1 = W n (1+r n+1 ) u n = W n R n+1 u n, where R n+1 = ( 1+r n+1. At any time m, aim: optimize ) J m,x (u m )=E N 2 n=m J n,w n (W n+1 )+J N 1,WN 1 (W N ) W m = x, where J n,wn (W n+1 )=W n+1 κ n Var n,wn (W n+1 ) Var n,wn (W n+1 )=Var(W n+1 W n ), and Σ n = Var(r n+1 ), The above: multi-period selection criterion of MSD type; κ n characterizes investor s risk aversion. Details: BGPW.

16 Motivation Technicalities The value function of this control problem: V (m,x)= sup u m U m J m,x (u m ). (1) ( Use KL divergence R(E)=E E loge ), where E is the ratio of the density of an alternative distribution to model distribution. For a given η > 0, a KL divergence ball is: B η ={E : R(E) η}. (2)

17 Motivation Technicalities Next: robust version of the problem. Let balls B ηn ={E : R(E) η n }, where n=0,...,n 1. Given any starting time m=0,...,n 1, we denote the set of E m =(E m,...,e N 1 ) such that each E n B ηn, where n=m,...,n 1, by B m ( E n : ratio of density of alternative to density of the nominal distribution over [n, n + 1]). Then: V(m,x)= sup κ m u m Σ m u m )+ u m U m J m,x (E m,u m )=E N 1 n=m+1 inf J m,x (E m,u m ), E m B m ( E m W m (R m+1 u m e η nγ n E n J n,wn (W n+1 ) W m = x ),

18 Motivation Technicalities Remark. γ n is a weighting to reflect the fact that a large uncertainty in the future should not impact too much the decision at the current stage. In addition, this also guarantees the existence an optimal solution. As η n 0 return to the non-robust case.

19 Main result Semi-Analytical Optimal Solution under KL Divergence. We require strongly time consistent optimal robust strategy. It represents a robustified version of a strong time consistent optimal strategy inspired by Kang and Filar 2006 (see also BGPW). Definition A strategy u m, =(um,...,u N 1 ) is strongly time consistent optimal robust w.r.t. J m,x (E m,u m ) if: 1: Let A m U m be a set of strategies u m =(v,u m+1,...,u N 1 ). Then E m, B m s.t. sup u m A m inf E m B m J m,x(e m,u m )=J m,x (E m,,u m, ). 2: For n=m+1,...,n 1, E n, B n s.t. sup u n U n inf E n B n J n,x(e n,u n )=J n,x (E n,,u n, ). If only 1 is satisfied: call it weakly time consistent optimal robust strategy w.r.t. J n,x ( ).

20 Main result Semi-Analytical Optimal Solution under KL Divergence. We require strongly time consistent optimal robust strategy. It represents a robustified version of a strong time consistent optimal strategy inspired by Kang and Filar 2006 (see also BGPW). Definition A strategy u m, =(um,...,u N 1 ) is strongly time consistent optimal robust w.r.t. J m,x (E m,u m ) if: 1: Let A m U m be a set of strategies u m =(v,u m+1,...,u N 1 ). Then E m, B m s.t. sup u m A m inf E m B m J m,x(e m,u m )=J m,x (E m,,u m, ). 2: For n=m+1,...,n 1, E n, B n s.t. sup u n U n inf E n B n J n,x(e n,u n )=J n,x (E n,,u n, ). If only 1 is satisfied: call it weakly time consistent optimal robust strategy w.r.t. J n,x ( ).

21 Main result Semi-Analytical Optimal Solution under KL Divergence. We require strongly time consistent optimal robust strategy. It represents a robustified version of a strong time consistent optimal strategy inspired by Kang and Filar 2006 (see also BGPW). Definition A strategy u m, =(um,...,u N 1 ) is strongly time consistent optimal robust w.r.t. J m,x (E m,u m ) if: 1: Let A m U m be a set of strategies u m =(v,u m+1,...,u N 1 ). Then E m, B m s.t. sup u m A m inf E m B m J m,x(e m,u m )=J m,x (E m,,u m, ). 2: For n=m+1,...,n 1, E n, B n s.t. sup u n U n inf E n B n J n,x(e n,u n )=J n,x (E n,,u n, ). If only 1 is satisfied: call it weakly time consistent optimal robust strategy w.r.t. J n,x ( ).

22 Main result Since the value function of the robust control problem is separable (in the sense that it can be written as a sum of expectations), we know: weakly time consistent optimal strategy, which can be found by period-wise optimization, is also a strongly time consistent optimal strategy.

23 Main result Theorem. Suppose (um ), m=0,1,...,n ( 1 is a strategy where there exists a sequence (θm) s.t. E exp ( R m+1 u m 2 θm) ) <, and u m S m = = S m X m = E ( E m log(e m) ) = η m, ( κ m ( E m X m b mσm 1 1 ) a m Σ 1 1 a m 1 h m + (b m) 2 κm 2 = 1 a m κma 2 m m 1 + Σ κ 2 m g m exp ( R m+1 u m 1 ) ) θm Rm+1 ( E exp ( R m+1 u m 1 ) ) θm +e η m+1γ m+1 G m+1 (u m+1,θ m+1 )E(R m+1), a m, (3), (4)

24 Main result where gm = hm (b m )2, hm =(X a m) T Σm 1 Xm, a m = 1 Σ 1 m 1, m bm = 1 Σ 1 m X m, E m = exp( R ) m+1 u 1 m θm ( E exp ( R m+1 u m 1 ) ) P-a.s., θm G m (u m,θ m )= θ m loge (exp ( R m+1 u m 1 θ m) ) + e η m+1γ m+1 G m+1 (u m+1,θ m+1 )E( R m+1 u m) κm S m η m θ m, G N (u N,θ N )=0. Then, (u m) is optimal, and the value function is given by where x (0, ). V(m,x)=xG m (u m,θ m),

25 Numerical Examples. d = 3 stocks: Navitas, Domino and Tabcorp, N = 5. Historical daily prices collected: 1 Jan Dec The 261 daily returns calculated. Set κ n = 3, W 0 = 1, returns are assumed to be i.i.d. over the investment horizon. Comparison of Optimal Robust and Non-Robust Portfolio. For a nominal 3-dim. MVN, mean µ, cov. mat. Σ, and alternative model: 3-dimensional MVN, mean µ, cov. mat. Σ, KL: ( R(E)= 1 trace(σ 1 Σ)+(µ µ) ( det(σ) ) ) Σ 1 (µ µ) d+log. 2 det( Σ) For illustration consider µ = c n µ,for some c n R,and Σ= Σ. Assume expected returns and covariance matrices stay constant over the investment horizon (i.e. E(r n )= µ, and Σ n = Σ).

26 Numerical Examples. In the worst case scenario for model disturbance, the alternative model is on the boundary of the KL div. ball (see Theorem) the divergence between the two models is equal to η n we can find c n. Assume η n to be constant over the entire investment horizon. For simplicity, simply write η and c. Choose (γ 2,γ 3,γ 4,γ 5 ) such that (ηγ 2,ηγ 3,ηγ 4,ηγ 5 )=(7.5,8.0,8.5,9.0) (reflecting investor s risk tolerance for uncertainty of distribution) (in contrast to κ which is the risk aversion of the investor s preference for a fixed distribution). So: the investor will have its own freedom to choose the amount of penalization.

27 Numerical Examples. Suppose: the alternative model is the true model. By generating data from the alternative model compare the performance under the optimal robust and non-robust strategies for different η. The optimal robust strategy: calculated by using 500, 000 Monte Carlo simulations. Then, we simulate 500,000 daily return paths (over 5 days), and compare the performance over three reasonable metrics:

28 Numerical Examples Performance Comparison 1 5 Number of Outperform Times η Figure: the number of times robust outperforms non-robust

29 Numerical Examples. Performance Comparison E(W robust N ) E(W non robust N ) Dollars $ η Figure: robust vs non-robust: expected terminal wealth

30 Numerical Examples. Performance Comparison E(W robust N ) W0 V ar(w robust ) N E(W non robust N ) W0 V ar(w non robust ) N -0.5 Dollars $ η Figure: robust vs non-robust: ratio of the difference between the expected terminal wealth and the initial wealth to the standard deviation of the terminal wealth

31 Numerical Examples. Table: Performance of Robust and Non-Robust Optimal Solution (Shift of Mean) c η Times robust % E(W N ) outperforms model robust non-robust difference % % % % % % %

32 Numerical Examples. Another case with a closed form formula for the KL div.: when both the nominal and alternative models are multivariate skew-normal. Given a nominal model Y SN d (µ,σ, ξ) and an alternative model Ȳ SN d( µ, Σ, ξ), then models is given by: R skew (E)= R(E)+2 ( +E 2 π (µ µ) Σ 1 Σ 1 2 ξ E ( log 2Φ ( Ξ 1 1 ξ T ξ ))), ) Ξ 1 SN 1 (0, ξ ξ, ξ ξ, Ξ 2 SN 1 (ξ T Σ 1 2 (µ µ), ξ T Σ 1 2 ΣΣ 1 2 ξ, ( ( log 2Φ ( Ξ 2 1 ξ T ξ ))) ξ T Σ 2 1 Σ 1 2 ξ ξ T Σ 1 2 ΣΣ 2 1 ξ ).

33 Numerical Examples Performance Comparison 1 5 Number of Outperform Times η Figure: the number of times robust outperform non-robust

34 Numerical Examples. Performance Comparison Dollars $ E(W robust N ) E(W non robust N ) η Figure: robust vs non-robust: expected terminal wealth

35 Numerical Examples Performance Comparison Dollars $ E(W robust N ) W0 V ar(w robust ) N E(W non robust N ) W0 V ar(w non robust ) N η Figure: robust vs non-robust: ratio of the difference between the expected terminal wealth and the initial wealth to the standard deviation of the terminal wealth

36 Numerical Examples. Table: Performance of Robust and Non-Robust Optimal Solution (skew-normal) c η Times robust % E(W N ) outperforms model robust non-robust difference % % % % % % % W.r. to the first performance comparison, we notice that now the number of times that the optimal non-robust strategy outperforms the robust strategy decreases as the radius of the divergence increases. However, the percentages of outperform times are all above 50%

37 Numerical Examples. Quantification of Model Risk with Empirical Data. Outline 1 Introduction 2 Motivation Technicalities 3 Main result 4 Numerical Examples. Quantification of Model Risk with Empirical Data. 5 Conclusion. 6 References

38 Numerical Examples. Quantification of Model Risk with Empirical Data. Although in some cases it may be worth choosing the non-robust optimal strategy, we still need to quantify the amount of model risk involved in doing so. Let Q denote the probability measure of an alternative model (i.e. the empirical measure for the true distribution). The optimal portfolio is said to have a model risk of θ with a confidence level q if ( Q W non robust N W robust N ) θ = 1 q. (In other words, model risk is the (1 q)th-quantile of the distribution of the difference between the terminal wealth under the non-robust strategy and the robust strategy.)

39 Numerical Examples. Quantification of Model Risk with Empirical Data Figure: The distribution of (W non robust N W robust N )

40 Numerical Examples. Quantification of Model Risk with Empirical Data. Divide the dataset: the first (dataset 1), is used to estimate the expected value and the covariance matrix of the nominal, which yields: µ = , Σ= (5) For illustration assume that the nominal is a 3-dimensional MVN with above mean vector and cov. matrix. Dataset 2 (60 data points) is used to evaluate a forecasted distribution of the incoming daily returns (our estimated alternative model to be used in the following 5 days). Use estimation procedure based on the kth-nearest-neighbor approach. Each time, a sample of 60 is generated from the MVN and the divergence between MNV and alternative model is estimated by using this sample, and dataset 2. Repeat 100,000 times.

41 Numerical Examples. Quantification of Model Risk with Empirical Data. Estimated divergence ( using ) kth-nearest-neighbor approach: ˆR(E)= d S S i=1 log Syk (i) (S 1)ỹ k (i). where ỹ k (i) is the Euclidean distance of the kth-nearest-neighbor of Ỹi in (Ỹj) j i, y k (i) is the Euclidean distance of the kth-nearest-neighbor of Ỹi in (Y i ), d = 3. Get: ˆR(E) By knowing the KL divergence, we use a bootstrapping to sample 100, 000 data points from dataset 2 to construct the distribution of W non robust N WN robust. The estimated model risk at q = 95% confidence level is , i.e., if the optimal non-robust strategy is applied but the optimal robust strategy turns to be more appropriate, then 95% of the time we would lose no more than 0.66 cents for every one dollar.

42 Conclusion. Derived a semi-analytical form of optimal robust strategy for an investment portfolio when uncertainty of distribution of the returns is involved. We have applied our approach to several numerical examples and have suggested whether or not we should take the optimal robust or non-robust strategy. We also define model risk from risk management perspective and present an algorithm for quantifying the model risk by using empirical data. This produces a convenient way of quantifying model risk in practice.

43 References H. Bannister, B. Goldys, S. Penev and W. Wu (2016), "Multiperiod mean-standard-deviation time consistent portfolio selection",automatica,73, P. Glasserman and X. B. Xu (2013), "Robust portfolio control with stochastic factor dynamics," Oper. Res., 61, P. Glasserman and X. B. Xu (2014),"Robust Risk Measurement and Model Risk," Quant. Finance, 14, J. C. Schneider and N. Schweizer (2015), "Robust measurement of (heavy-tailed) risks: Theory and implimentation,", J. Econom. Dynam. Control, 61, T. Breuer and I. Csiszár (2016),"Measuring distribution model risk," Math. Finance, 26,

44 References H. Bannister, B. Goldys, S. Penev and W. Wu (2016), "Multiperiod mean-standard-deviation time consistent portfolio selection",automatica,73, P. Glasserman and X. B. Xu (2013), "Robust portfolio control with stochastic factor dynamics," Oper. Res., 61, P. Glasserman and X. B. Xu (2014),"Robust Risk Measurement and Model Risk," Quant. Finance, 14, J. C. Schneider and N. Schweizer (2015), "Robust measurement of (heavy-tailed) risks: Theory and implimentation,", J. Econom. Dynam. Control, 61, T. Breuer and I. Csiszár (2016),"Measuring distribution model risk," Math. Finance, 26,

45 References H. Bannister, B. Goldys, S. Penev and W. Wu (2016), "Multiperiod mean-standard-deviation time consistent portfolio selection",automatica,73, P. Glasserman and X. B. Xu (2013), "Robust portfolio control with stochastic factor dynamics," Oper. Res., 61, P. Glasserman and X. B. Xu (2014),"Robust Risk Measurement and Model Risk," Quant. Finance, 14, J. C. Schneider and N. Schweizer (2015), "Robust measurement of (heavy-tailed) risks: Theory and implimentation,", J. Econom. Dynam. Control, 61, T. Breuer and I. Csiszár (2016),"Measuring distribution model risk," Math. Finance, 26,

46 References H. Bannister, B. Goldys, S. Penev and W. Wu (2016), "Multiperiod mean-standard-deviation time consistent portfolio selection",automatica,73, P. Glasserman and X. B. Xu (2013), "Robust portfolio control with stochastic factor dynamics," Oper. Res., 61, P. Glasserman and X. B. Xu (2014),"Robust Risk Measurement and Model Risk," Quant. Finance, 14, J. C. Schneider and N. Schweizer (2015), "Robust measurement of (heavy-tailed) risks: Theory and implimentation,", J. Econom. Dynam. Control, 61, T. Breuer and I. Csiszár (2016),"Measuring distribution model risk," Math. Finance, 26,

47 References H. Bannister, B. Goldys, S. Penev and W. Wu (2016), "Multiperiod mean-standard-deviation time consistent portfolio selection",automatica,73, P. Glasserman and X. B. Xu (2013), "Robust portfolio control with stochastic factor dynamics," Oper. Res., 61, P. Glasserman and X. B. Xu (2014),"Robust Risk Measurement and Model Risk," Quant. Finance, 14, J. C. Schneider and N. Schweizer (2015), "Robust measurement of (heavy-tailed) risks: Theory and implimentation,", J. Econom. Dynam. Control, 61, T. Breuer and I. Csiszár (2016),"Measuring distribution model risk," Math. Finance, 26,