Bayesian Dynamic Linear Models for Strategic Asset Allocation

Transcription

1 Bayesian Dynamic Linear Models for Strategic Asset Allocation Jared Fisher Carlos Carvalho, The University of Texas Davide Pettenuzzo, Brandeis University April 18, 2016 Fisher (UT) Bayesian Risk Prediction April 18, / 50

2 1 Introduction 2 Single Risky Asset 3 Multiple Risky Assets 4 Conclusion Fisher (UT) Bayesian Risk Prediction April 18, / 50

3 Excess Returns on an Index: is there Signal in the Noise? Percent Excess Return Stock Index 2 Year Bond Index 5 Year Bond Index Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

4 How should an investor optimally create a portfolio? Two step process: Establish predictions of the mean and variance of assets future excess returns Use these estimates to determine how much of portfolio to devote to each asset. Return on a portfolio is a weighted sum of the individual assets returns, where the weights are the proportions invested. Fisher (UT) Bayesian Risk Prediction April 18, / 50

5 Making Investments Given forecasted ˆµ t, ˆΣ t For an investor with power utility and risk aversion γ Portfolio weights vector is w t = 1 γ ( 1 ˆΣ t ˆµ t + 1 ) 2 diag(ˆσ t ) Fisher (UT) Bayesian Risk Prediction April 18, / 50

6 Understanding Excess Returns What is the distribution of Y i,t+1 = (R i,t+1 R f,t+1 ), given what we know at time t? E(Y i,t+1 D t ) ( risk premium ) = µ? (constant, no predictability) = µ t = f(x t ) = X tβ? = µt = X tβ t? ( time-varying parameters ) V ar(y i,t+1 D t ) = σ 2? (constant volatility) = σt 2? ( stochastic volatility ) Fisher (UT) Bayesian Risk Prediction April 18, / 50

7 Does Predictability Exist? Literature assumes linear relationship: Y i,t+1 = X tβ + ɛ t+1, V ar(ɛ) = σ 2 Tests are mostly in-sample, not out-of-sample (OOS). Welch and Goyal (2008) show that the good performance of popular variables in-sample don t hold OOS. More recently, authors show OOS predictability by deviating from the standard model. Time-varying parameters (e.g. Dangl and Halling, 2012) Stochastic volatility (e.g. Johannes, Korteweg and Polson, 2013) Parameter uncertainty (Bayesian models) Fisher (UT) Bayesian Risk Prediction April 18, / 50

8 Our Analysis Two research questions Predictability: is there useful information in X? Time-variation: are the parameter values (β and σ 2 ) constant with respect to time? Compare models with and without predictors and with and without variance discounting (of both regression coefficients and volatility) Benchmark: the constant model (i.e. X t = 1) Often called the expectation hypothesis model, it represents the efficient markets hypothesis/no predictability. Fisher (UT) Bayesian Risk Prediction April 18, / 50

9 Data description We will first look at portfolios of a risky asset (stock index or bond index) and a risk-free asset (3 month T-bill). We use the following data, spanning : Welch and Goyal s predictors of stock performance, updated to 2014, CRSP value weighted returns, Bonds data from Gargano, Pettenuzzo, and Timmermann (2015) Bond index for 2-5 year maturities, Cochrane and Piazzesi s (2005) linear combination of forward rates, Fama and Bliss (1987) forward spread, Ludvigson and Ng s (2009) macro factor. Fisher (UT) Bayesian Risk Prediction April 18, / 50

10 Our Model Y t = X t 1B t + v t B t = B t 1 + Ω t v t N(0, V t Σ t ) Ω t N(0, W t, Σ t ) (B 0, Σ 0 D 0 ) NW 1 n 0 (m 0, C 0, S 0 ) Σ t D t 1 W 1 δ vn t 1 (S t 1 ) W t = 1 δ C t 1 δ Fisher (UT) Bayesian Risk Prediction April 18, / 50

11 Y t = X t 1β t + ɛ t β DY OLS-Expanding Window Dyamic Linear Model DLM with Time-varying Parameters Full-term OLS Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

12 Model Advantages Bayesian model without need of MCMC Allows us to fits more models in the same amount of computation time Bridges gap between Recursive model vs. Rolling-window model Fisher (UT) Bayesian Risk Prediction April 18, / 50

13 Recursion. β t 1, Σ t 1 D t 1 NW 1 n t 1 (m t 1, C t 1, S t 1) β t, Σ t 1 D t 1 NW 1 n t 1 (m t 1, R t, S t 1) Y t D t 1 T nt 1 (X t 1m t 1, Q ts t 1) R t = C t 1 + W t = 1 δ Ct 1 Q t = V t + X t 1R tx t 1 β t, Σ t D t NW 1 n t (m t, C t, S t) e t = Y t X t 1m t 1. A t = R tx t 1/Q t n t = δ vn t m t = m t 1 + A te t C t = R t A ta tq t S t = n 1 t ( δ vn t 1S t 1 + ete t Q t ) Fisher (UT) Bayesian Risk Prediction April 18, / 50

14 Modeling Details Prior created on data Models evaluated on Evaluated on both economic and statistical criteria. Economic Measure: Certainty Equivalent Returns, using power utility (CRRA) U(wealth) = 1 1 γ (wealth)1 γ γ = 5 Statistical Prediction Measure: Mean Squared Prediction Error Ratio Statistical Fit Measure: Average Log Score Restrict: Portfolio weights wt [ 2, 3] Coefficient variance discount factor δ [0.98, 1.0] Volatility discount factor δ v [0.9, 1.0] Fisher (UT) Bayesian Risk Prediction April 18, / 50

15 Results: No Discounting - Comparison to Literature Stock Index Bond Index Predictor CER ALS MSE Mat. Pred. CER ALS MSE (none) (none) Log D/P CP Log D/Y FB Log E/P LN Smooth E/P (none) Log D/Payout CP B/M FB T Bill Rate LN LngTerm Yld (none) LngTerm Ret CP Term Spread FB Def.Yld.Sprd LN Def.Ret.Sprd (none) Stock Var CP Net Eqty Exp FB Inflation LN Fisher (UT) Bayesian Risk Prediction April 18, / 50

16 Discount Factor Heatmap - Grid of δ, δ v 1.00 LN 2 years CER δ v δ Fisher (UT) Bayesian Risk Prediction April 18, / 50

17 Average Over Models Many models beat the benchmark, given the correct discount factors. But, we don t know a priori how much to discount or which predictors will perform well. Solution: average and share strength across models. For each time t, weight each of the models prediction based on its performance up through time t 1. Create different averaged models by weighting on utility and score, as well as an equal-weighted model. w U i,τ+1 = w S i,τ+1 = ( 1 1 γ τ ( τ t=1 τ t=1 U i,t ) 1 1 γ ln(score i,t ) ) min j ( τ ) ln(score j,t ) t=1 Fisher (UT) Bayesian Risk Prediction April 18, / 50

18 Modeling Details A model is fit for every combination of predictor, δ, and δ v. 10 values of δ and δ v are considered, equally spaced in the range δ [0.98, 1.0], δ v [0.9, 1.0], for a grid of 100 possibilities. Fisher (UT) Bayesian Risk Prediction April 18, / 50

19 Model Averaging Results: Stocks Pred TVP SV Models Weights CER ALS MSE (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

20 Model Averaging Results: Stocks Stocks CER Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted ALS Fisher (UT) Bayesian Risk Prediction April 18, / 50

21 Model Averaging Results: Bonds, 2 Year Maturity Pred TVP SV Models Weights CER ALS MSE (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

22 Model Averaging Results: Bonds, 2 Year Maturity Bonds_2 CER ALS Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted Fisher (UT) Bayesian Risk Prediction April 18, / 50

23 Model Averaging Results: Bonds, 3 Year Maturity Pred TVP SV Models Weights CER ALS MSE (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

24 Model Averaging Results: Bonds, 3 Year Maturity Bonds_3 CER ALS Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted Fisher (UT) Bayesian Risk Prediction April 18, / 50

25 Model Averaging Results: Bonds, 4 Year Maturity Pred TVP SV Models Weights CER ALS MSE (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

26 Model Averaging Results: Bonds, 4 Year Maturity Bonds_4 CER ALS Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted Fisher (UT) Bayesian Risk Prediction April 18, / 50

27 Model Averaging Results: Bonds, 5 Year Maturity Pred TVP SV Models Weights CER ALS MSE (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

28 Model Averaging Results: Bonds, 5 Year Maturity Bonds_5 CER ALS Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted Fisher (UT) Bayesian Risk Prediction April 18, / 50

29 Conclusions on Single Asset Models The best single risky asset models include predictors and stochastic volatility, perhaps with time-varying parameters for bonds. Does predictability exist? Yes, the best averaged model in most cases include predictors. Is time variation important? Yes, especially stochastic volatility. Fisher (UT) Bayesian Risk Prediction April 18, / 50

30 Our Multivariate Model Ideal portfolio probably contains more than one risky asset. Use this same model, but fit for multiple risky assets. Portfolio of the stock index and a bond index, for a given maturity. Each model can include one stock predictor and one bond predictor Fisher (UT) Bayesian Risk Prediction April 18, / 50

31 Multivariate Model Averaging Results, 2 year maturity Pred TVP SV Models Weights CER ALS MSE S. MSE B (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

32 Multivariate Model Averaging Results, 2 year maturity StocksBonds_2 CER Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted ALS Fisher (UT) Bayesian Risk Prediction April 18, / 50

33 Multivariate Model Averaging Results, 3 year maturity Pred TVP SV Models Weights CER ALS MSE S. MSE B (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

34 Multivariate Model Averaging Results, 3 year maturity StocksBonds_3 CER ALS Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted Fisher (UT) Bayesian Risk Prediction April 18, / 50

35 Multivariate Model Averaging Results, 4 year maturity Pred TVP SV Models Weights CER ALS MSE S. MSE B (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

36 Multivariate Model Averaging Results, 4 year maturity StocksBonds_4 CER ALS Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted Fisher (UT) Bayesian Risk Prediction April 18, / 50

37 Multivariate Model Averaging Results, 5 year maturity Pred TVP SV Models Weights CER ALS MSE S. MSE B (none) Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Equal Utility Score Fisher (UT) Bayesian Risk Prediction April 18, / 50

38 Multivariate Model Averaging Results, 5 year maturity StocksBonds_5 CER Benchmark SV TVP TVP SV w/predictors Equal Weighted Utility Weighted Score Weighted ALS Fisher (UT) Bayesian Risk Prediction April 18, / 50

39 Summary The best single risky asset models include predictors and stochastic volatility, perhaps with time-varying parameters for bonds. If optimizing statistical fit (ALS), the best models of multiple risky assets include stochastic volatility, usually with predictors. If optimizing economic significance (CER), the best models of multiple risky assets include Predictors alone for shorter maturities. Time-varying parameters and stochastic volatility with no predictors for larger maturities, equal or utility weighted (also the balanced choice). Fisher (UT) Bayesian Risk Prediction April 18, / 50

40 Limitations The literature has shown that the time period used affects results. However, showing that there is predictability from runs against Welch and Goyal s finding that predictability disappears in the more recent data. Fisher (UT) Bayesian Risk Prediction April 18, / 50

41 Conclusions We demonstrate a Bayesian methodology that can quickly estimate a time-series model without requiring MCMC or another computation-intensive sampling algorithm. Time-varying parameters, stochastic volatility, and predictors generally show improvements over the benchmark model. Does predictability exist? Yes, the best averaged model in most cases include predictors. Is time variation important? Yes, especially stochastic volatility. Fisher (UT) Bayesian Risk Prediction April 18, / 50

42 Questions, Comments? Thank you! Fisher (UT) Bayesian Risk Prediction April 18, / 50

43 Different Risk Aversion What if γ = 10? Fisher (UT) Bayesian Risk Prediction April 18, / 50

44 Multivariate Portfolio Weights, 2 Year Maturity, γ = 5 Historic Mean Model - Portfolio Weights Weight - Percent Invested Stocks Bonds Start Eval Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

45 Multivariate Portfolio Weights, 2 Year Maturity, γ = 10 Historic Mean Model - Portfolio Weights Weight - Percent Invested Stocks Bonds Start Eval Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

46 Multivariate Portfolio Weights, 2 Year Maturity, γ = 5 Score-weighted Model, no Discounting - Portfolio Weights Weight - Percent Invested Stocks Bonds Start Eval Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

47 Multivariate Portfolio Weights, 2 Year Maturity, γ = 10 Score-weighted Model, no Discounting - Portfolio Weights Weight - Percent Invested Stocks Bonds Start Eval Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

48 Multivariate Portfolio Weights, 2 Year Maturity, γ = 5 Score-weighted Model, with Discounting - Portfolio Weights Weight - Percent Invested Stocks Bonds Start Eval Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

49 Multivariate Portfolio Weights, 2 Year Maturity, γ = 10 Score-weighted Model, with Discounting - Portfolio Weights Weight - Percent Invested Stocks Bonds Start Eval Year Fisher (UT) Bayesian Risk Prediction April 18, / 50

50 Intervention Intervention: Expected risk premium should be non-negative if not positive. Fisher (UT) Bayesian Risk Prediction April 18, / 50