Asset Allocation and Risk Assessment with Gross Exposure Constraints

Transcription

1 Asset Allocation and Risk Assessment with Gross Exposure Constraints Forrest Zhang Bendheim Center for Finance Princeton University A joint work with Jianqing Fan and Ke Yu, Princeton Princeton University Asset Allocation with Gross Exposure Constraints 1/25

2 Introduction Princeton University Asset Allocation with Gross Exposure Constraints 2/25

3 Markowitz s Mean-variance analysis Problem: minw w T Σw, s.t. w T 1 = 1, and w T µ = r 0. Solution: w = c 1 Σ 1 µ + c 2 Σ 1 1 Cornerstone of modern finance where CAPM and many portfolio theory is built upon. Too sensitive on input vectors and their estimation errors. Can result in extreme short positions (Green and Holdfield, 1992). More severe for large portfolio. Princeton University Asset Allocation with Gross Exposure Constraints 3/25

4 Markowitz s Mean-variance analysis Problem: minw w T Σw, s.t. w T 1 = 1, and w T µ = r 0. Solution: w = c 1 Σ 1 µ + c 2 Σ 1 1 Cornerstone of modern finance where CAPM and many portfolio theory is built upon. Too sensitive on input vectors and their estimation errors. Can result in extreme short positions (Green and Holdfield, 1992). More severe for large portfolio. Princeton University Asset Allocation with Gross Exposure Constraints 3/25

5 Challenge of High Dimensionality Estimating high-dim cov-matrices is intrinsically challenging. Suppose we have 500 (2000) stocks to be managed. There are 125K (2 m) free parameters! Yet, 2-year daily returns yield only about sample size n = 500. Accurately estimating it poses significant challenges. Impact of dimensionality is large and poorly understood: Risk: w T ˆΣw. Allocation: ĉ1ˆσ ĉ 2ˆΣ 1ˆµ. Accumulating of millions of estimation errors can have a devastating effect. Princeton University Asset Allocation with Gross Exposure Constraints 4/25

6 Challenge of High Dimensionality Estimating high-dim cov-matrices is intrinsically challenging. Suppose we have 500 (2000) stocks to be managed. There are 125K (2 m) free parameters! Yet, 2-year daily returns yield only about sample size n = 500. Accurately estimating it poses significant challenges. Impact of dimensionality is large and poorly understood: Risk: w T ˆΣw. Allocation: ĉ1ˆσ ĉ 2ˆΣ 1ˆµ. Accumulating of millions of estimation errors can have a devastating effect. Princeton University Asset Allocation with Gross Exposure Constraints 4/25

7 Efforts in Remedy Reduce sensitivity of estimation. Shrinkage and Bayesian: Expected return (Klein and Bawa, 76; Chopra and Ziemba, 93; ) Cov. matrix (Ledoit & Wolf, 03, 04) Factor-model based estimation (Fan, Fan and Lv, 2008; Pesaran and Zaffaroni, 2008) Robust portfolio allocation (Goldfarb and Iyengar, 2003) No-short-sale portfolio (De Roon et al., 2001; Jagannathan and Ma, 2003; DeMiguel et al., 2008; Bordie et al., 2008) None of them are far enough; no theory. Princeton University Asset Allocation with Gross Exposure Constraints 5/25

8 Efforts in Remedy Reduce sensitivity of estimation. Shrinkage and Bayesian: Expected return (Klein and Bawa, 76; Chopra and Ziemba, 93; ) Cov. matrix (Ledoit & Wolf, 03, 04) Factor-model based estimation (Fan, Fan and Lv, 2008; Pesaran and Zaffaroni, 2008) Robust portfolio allocation (Goldfarb and Iyengar, 2003) No-short-sale portfolio (De Roon et al., 2001; Jagannathan and Ma, 2003; DeMiguel et al., 2008; Bordie et al., 2008) None of them are far enough; no theory. Princeton University Asset Allocation with Gross Exposure Constraints 5/25

9 About this talk Propose utility maximization with gross-sale constraint. It bridges no-short-sale constraint to no-constraint on allocation. Oracle (Theoretical), actual and empirical risks are very close. No error accumulation effect. Elements in covariance can be estimated separately; facilitates the use of non-synchronized high-frequency data. Provide theoretical understanding why wrong constraint can even beat Markowitz s portfolio (Jagannathan and Ma, 2003). Portfolio selection and tracking. Select or track a portfolio with limited number of stocks. Improve any given portfolio with modifications of weights on limited number of stocks. Princeton University Asset Allocation with Gross Exposure Constraints 6/25

10 About this talk Propose utility maximization with gross-sale constraint. It bridges no-short-sale constraint to no-constraint on allocation. Oracle (Theoretical), actual and empirical risks are very close. No error accumulation effect. Elements in covariance can be estimated separately; facilitates the use of non-synchronized high-frequency data. Provide theoretical understanding why wrong constraint can even beat Markowitz s portfolio (Jagannathan and Ma, 2003). Portfolio selection and tracking. Select or track a portfolio with limited number of stocks. Improve any given portfolio with modifications of weights on limited number of stocks. Princeton University Asset Allocation with Gross Exposure Constraints 6/25

11 Outline 1 Portfolio optimization with gross-exposure constraint. 2 Portfolio selection and tracking. 3 Simulation studies 4 Empirical studies: Princeton University Asset Allocation with Gross Exposure Constraints 7/25

12 Short-constrained portfolio selection maxw E[U(w T R)] s.t. w T 1 = 1, w 1 c, Aw = a. Equality Constraint: A = µ = expected portfolio return. A can be chosen so that we put constraint on sectors. Short-sale constraint: When c = 1, no short-sale allowed. When c =, problem becomes Markowitz s. Portfolio selection: solution is usually sparse. Princeton University Asset Allocation with Gross Exposure Constraints 8/25

13 Short-constrained portfolio selection maxw E[U(w T R)] s.t. w T 1 = 1, w 1 c, Aw = a. Equality Constraint: A = µ = expected portfolio return. A can be chosen so that we put constraint on sectors. Short-sale constraint: When c = 1, no short-sale allowed. When c =, problem becomes Markowitz s. Portfolio selection: solution is usually sparse. Princeton University Asset Allocation with Gross Exposure Constraints 8/25

14 Risk optimization Theory Actual and Empirical risks: R(w) = w T Σw, R n (w) = w T ˆΣw. w opt = argmin R(w), w 1 c ŵ opt = argmin R n (w) w 1 c Risks: R(w opt ) oracle, R n (ŵ opt ) empirical; R(ŵopt ) actual risk of a selected portfolio. Theorem 1: Let a n = ˆΣ Σ. Then, we have R(ŵ opt ) R(w opt ) 2a n c 2 R(ŵ opt ) R n (ŵ opt ) a n c 2 R(w opt ) R n (ŵ opt ) a n c 2. Princeton University Asset Allocation with Gross Exposure Constraints 9/25

15 Risk optimization Theory Actual and Empirical risks: R(w) = w T Σw, R n (w) = w T ˆΣw. w opt = argmin R(w), w 1 c ŵ opt = argmin R n (w) w 1 c Risks: R(w opt ) oracle, R n (ŵ opt ) empirical; R(ŵopt ) actual risk of a selected portfolio. Theorem 1: Let a n = ˆΣ Σ. Then, we have R(ŵ opt ) R(w opt ) 2a n c 2 R(ŵ opt ) R n (ŵ opt ) a n c 2 R(w opt ) R n (ŵ opt ) a n c 2. Princeton University Asset Allocation with Gross Exposure Constraints 9/25

16 Accuracy of Covariance: I Theorem 2: If for a sufficiently large x, max P { n σ ij ˆσ ij > x} < exp( Cx 1/a ), i,j for some two positive constants a and C, then Σ ˆΣ = O P ( (log p) a n ). Impact of dimensionality is limited. Princeton University Asset Allocation with Gross Exposure Constraints 10/25

17 Accuracy of Covariance: I Theorem 2: If for a sufficiently large x, max P { n σ ij ˆσ ij > x} < exp( Cx 1/a ), i,j for some two positive constants a and C, then Σ ˆΣ = O P ( (log p) a n ). Impact of dimensionality is limited. Princeton University Asset Allocation with Gross Exposure Constraints 10/25

18 Algorithms min w T Σw. w T 1=1, w 1 c 1 Quadratic programming for each given c (Exact). 2 Coordinatewise minimization. 3 LARS approximation. Princeton University Asset Allocation with Gross Exposure Constraints 11/25

19 Connections with penalized regression Regression problem: Letting Y = R p and X j = R p R j, var(w T R) = min b E(w T R b) 2 = min b E(Y w 1 X 1 w p 1 X p 1 b) 2, Gross exposure: w 1 = w T w c, not equivalent to w 1 d. d = 0 picks X p, but c = 1 picks multiple stocks. Princeton University Asset Allocation with Gross Exposure Constraints 12/25

20 Connections with penalized regression Regression problem: Letting Y = R p and X j = R p R j, var(w T R) = min b E(w T R b) 2 = min b E(Y w 1 X 1 w p 1 X p 1 b) 2, Gross exposure: w 1 = w T w c, not equivalent to w 1 d. d = 0 picks X p, but c = 1 picks multiple stocks. Princeton University Asset Allocation with Gross Exposure Constraints 12/25

21 Approximate solution LARS: to find solution path w (d) for PLS min b, w E(Y 1 d w T X b) 2, Approximate solution: PLS provides a suboptimal solution to risk optimization problem with c = d T w opt(d). Take Y = optimal no-short-sale constraint (c = 1). Multiple Y helps. e.g. Also take Y = solution to c = 2 Princeton University Asset Allocation with Gross Exposure Constraints 13/25

22 Portfolio tracking and improvement PLS regarded as finding a portfolio to minimize the expected tracking error portfolio tracking. PLS interpreted as modifying weights to improve the performance of Y Portfolio improvements. with limited number of stocks limited exposure. empirical risk path R n (d) helps decision making. Remark: PLS min b, w 1 d n t=1 (Y i w T X t b) 2 is equivalent to PLS using sample covariance matrix. Princeton University Asset Allocation with Gross Exposure Constraints 14/25

23 An illustration Data: Y = CRSP; X = 10 industrial portfolios. Today = 1/8/05. Sample Cov: one-year daily return. Actual: hold one year. (a) Portfolio Weights (b) Portfolio risk Ex ante Ex post weight Y annualized volatility(%) gross exposure constraint(d) number of stocks Princeton University Asset Allocation with Gross Exposure Constraints 15/25

24 Fama-French three-factor model Model: R i = b i1 f 1 + b i2 f 2 + b i3 f 3 + ε i or R = Bf + ε. f 1 = CRSP index; f 2 = size effect; f 3 = book-to-market effect Covariance: Σ = Bcov(f)B T + diag(σ 2 1,,σ2 p). Parameters for factor loadings Parameters for factor returns µ b cov b µ f cov f Parameters: Calibrated to market data (5/1/02 8/29/05, from Fan, Fan and Lv, 2008) Parameters: Factor loadings: b i i.i.d. N(µ b, cov b ) Noise: σ i i.i.d. Gamma(3.34,.19) conditioned on σ i >.20. Simulation: Factor returns f t i.i.d. N(µ f, cov f ), ε it i.i.d. σ i t 6 Princeton University Asset Allocation with Gross Exposure Constraints 16/25

25 Fama-French three-factor model Model: R i = b i1 f 1 + b i2 f 2 + b i3 f 3 + ε i or R = Bf + ε. f 1 = CRSP index; f 2 = size effect; f 3 = book-to-market effect Covariance: Σ = Bcov(f)B T + diag(σ 2 1,,σ2 p). Parameters for factor loadings Parameters for factor returns µ b cov b µ f cov f Parameters: Calibrated to market data (5/1/02 8/29/05, from Fan, Fan and Lv, 2008) Parameters: Factor loadings: b i i.i.d. N(µ b, cov b ) Noise: σ i i.i.d. Gamma(3.34,.19) conditioned on σ i >.20. Simulation: Factor returns f t i.i.d. N(µ f, cov f ), ε it i.i.d. σ i t 6 Princeton University Asset Allocation with Gross Exposure Constraints 16/25

26 Risk Improvements and decision making (a) Empirical and actual risks sample cov actual risk empirical risk 250 (b) Number of stocks sample cov Sample Annual volatility (%) 8 6 number of stocks Exposure parameter d Exposure parameter d (c) Empirical and actual risks factor cov actual risk empirical risk 250 (d) Number of stocks factor cov 200 Factor Annual volatility (%) number of stocks Exposure parameter d Exposure parameter d Factor-model based estimation is more accurate. Princeton University Asset Allocation with Gross Exposure Constraints 17/25

27 Empirical studies (I) Princeton University Asset Allocation with Gross Exposure Constraints 18/25

28 Some details Data: 100 portfolios from the website of Kenneth French from (10 years) Portfolios: two-way sort according to the size and book-to-equity ratio, 10 categories each. Evaluation: Rebalance monthly, and record daily returns. Covariance matrix: Estimate by sample covariance matrix, factor model used last twelve months daily data, and RiskMetrics. Princeton University Asset Allocation with Gross Exposure Constraints 19/25

29 Risk, Sharpe-Ratio, Maximum Weight, Annualized return Risk of Portfolios factor sample risk metrics Sharpe Ratio factor sample risk metrics Annualized Risk (%) Sharpe ratio Exposure Constraint (C) Exposure Constraint (C) Max Max of portfolios factor sample risk metrics Annualized Return (%) Annualized Return factor sample risk metrics Exposure Constraint (C) Exposure Constraint (C) Princeton University Asset Allocation with Gross Exposure Constraints 20/25

30 Short-constrained MV portfolio (Results I) Methods Mean Std Sharpe-R Max-W Min-W Long Short Sample Covariance Matrix Estimator No short(c = 1) Exact(c = 1.5) Exact(c = 2) Exact(c = 3) Approx. (c = 2) Approx. (c = 3) GMV Unmanaged Index Equal-W CRSP Princeton University Asset Allocation with Gross Exposure Constraints 21/25

31 Empirical studies (II) Princeton University Asset Allocation with Gross Exposure Constraints 22/25

32 Some details Data: 1000 stocks with missing data selected from Russell 3000 from (5 years). Allocation: Each month, pick 400 stocks at random and allocate them (mitigating survivor biases). Evaluation: Rebalance monthly, and record daily returns. Covariance matrix: Estimate by sample covariance matrix, factor model used last twenty-four months daily data, and RiskMetrics. Princeton University Asset Allocation with Gross Exposure Constraints 23/25

33 15 14 (a)risk of portfolios(ns) factor sample risk metrics (b)risk of portfolios (mkt) factor sample risk metrics Annuzlied Volatility(%) Annualized Volatility(%) number of stocks number of stocks Princeton University Asset Allocation with Gross Exposure Constraints 24/25

34 Conclusion Utility maximization with gross-sale constraint bridges no-short-sale constraint to no-constraint on allocation. It makes oracle (theoretical), actual and empirical risks close: No error accumulation effect for a range of c; Elements in covariance can be estimated separately; facilitates use of non-synchronize high-frequency data. Provide theoretical understanding why wrong constraint help. Portfolio selection, tracking, and improvement. Select or track a portfolio with limited number of stocks. Improve any given portfolio with modifications of weights on limited number of stocks. Provide tools for checking efficiency of a portfolio. Princeton University Asset Allocation with Gross Exposure Constraints 25/25

35 Conclusion Utility maximization with gross-sale constraint bridges no-short-sale constraint to no-constraint on allocation. It makes oracle (theoretical), actual and empirical risks close: No error accumulation effect for a range of c; Elements in covariance can be estimated separately; facilitates use of non-synchronize high-frequency data. Provide theoretical understanding why wrong constraint help. Portfolio selection, tracking, and improvement. Select or track a portfolio with limited number of stocks. Improve any given portfolio with modifications of weights on limited number of stocks. Provide tools for checking efficiency of a portfolio. Princeton University Asset Allocation with Gross Exposure Constraints 25/25