Discussion of Portfolio Choice with Model Misspeci…cation by Pesaran, Uppal, and Za¤aroni Xiaoji Lin Ohio State University UBC Summer Finance Conference July 31, 2015 1 / 17 Context: Portfolio choice, APT, and misspeci…cation The literature on APT Ross ’76, ’77, Huberman ’82, Chamberlain and Rothschild ’83, Ingersoll ’84 Portfolio selection in the presence of "mispricing" Dybvig and Ross ’85, Green ’86, Grinblatt and Titman ’89, Kosowski et al ’06 The implications of the no-arbitrage constraint imposed by APT for the mean-variance portfolio estimation Jobson and Korkie ’80, Black and Litterman ’90, Green and Holli…eld ’92, MacKinlay and Pastor ’00, Jagannathan and Ma ’03, Contribution: Extends the original APT to allow for large pricing errors due to model misspeci…cation and decomposes the mean-variance portfolio into an "alpha" portfolio and a beta portfolio. 2 / 17 Context: Portfolio choice, APT, and misspeci…cation The literature on APT Ross ’76, ’77, Huberman ’82, Chamberlain and Rothschild ’83, Ingersoll ’84 Portfolio selection in the presence of "mispricing" Dybvig and Ross ’85, Green ’86, Grinblatt and Titman ’89, Kosowski et al ’06 The implications of the no-arbitrage constraint imposed by APT for the mean-variance portfolio estimation Jobson and Korkie ’80, Black and Litterman ’90, Green and Holli…eld ’92, MacKinlay and Pastor ’00, Jagannathan and Ma ’03, Contribution: Extends the original APT to allow for large pricing errors due to model misspeci…cation and decomposes the mean-variance portfolio into an "alpha" portfolio and a beta portfolio. 2 / 17 Outline 1 Main results in the paper 2 Tests of the theory implications in the data 3 / 17 The APT In matrix form: R = A + BF + " E["jF ] = 0 E[""0 jF ] = R is N 1 vector of asset returns A is N 1 vector of intercepts B = [b1 ; b2 ; :::bN ] is a N K matrix of factor sensitivities " is N 1 vector of disturbance term. 4 / 17 Notes Given the structure, Ross (1976) shows that the absence of arbitrage in large economies implies that =( ) +B K where is Nx1 vector of expected returns and is the model zero-beta parameter, and equals the risk-free rate if such a rate exists This relationship holds exactly only as the number of assets !in…nity. 5 / 17 Key result 1 Extends the interpretation of the APT and shows that it can capture large pricing errors that arise from mis-measured or missing factors |{z} pricing errors = A |{z} loadings m |{z} latent/missing factors + a |{z} idio. pricing errors The original APT interpretation pricing errors are small pricing errors are unrelated to factors 6 / 17 Key result 1 Extends the interpretation of the APT and shows that it can capture large pricing errors that arise from mis-measured or missing factors |{z} pricing errors = A |{z} loadings m |{z} latent/missing factors + a |{z} idio. pricing errors The original APT interpretation pricing errors are small pricing errors are unrelated to factors 6 / 17 Key result 2 Under the extended APT the optimal mean-variance portfolio = an alpha portfolio + {z } | depends on pricing errors a beta portfolio {z } | depends on factor risk premia/loadings The original APT implies the optimal mean-variance portfolio = a beta portfolio {z } | depends on factor risk premia/loadings Obtains similar decompositions for the global minimum-variance portfolio and the e¢ cient-frontier portfolios. 7 / 17 Key result 2 Under the extended APT the optimal mean-variance portfolio = an alpha portfolio + {z } | depends on pricing errors a beta portfolio {z } | depends on factor risk premia/loadings The original APT implies the optimal mean-variance portfolio = a beta portfolio {z } | depends on factor risk premia/loadings Obtains similar decompositions for the global minimum-variance portfolio and the e¢ cient-frontier portfolios. 7 / 17 Key result 2 Under the extended APT the optimal mean-variance portfolio = an alpha portfolio + {z } | depends on pricing errors a beta portfolio {z } | depends on factor risk premia/loadings The original APT implies the optimal mean-variance portfolio = a beta portfolio {z } | depends on factor risk premia/loadings Obtains similar decompositions for the global minimum-variance portfolio and the e¢ cient-frontier portfolios. 7 / 17 Key result 3 Practical implications: The decomposition results and the restriction arising from the extended APT can be used to improve the estimation of portfolio weights when the model misspeci…cation is present. 8 / 17 My discussion: Test the theory results in the data 1 A single factor model: Market 2 A two factor model: Market + Equity issuance shock (Belo, Lin, and Yang ’15) 3 A two factor model: Market + a latent factor without the APT restriction on pricing errors ( 0 < 0) 4 A two factor model: Market + a latent factor with the APT restrictions on pricing errors ( 0 < 0) Test assets: 5 BM portfolios, 5 investment portfolio, 5 size portfolios 9 / 17 Standard asset pricing tests Single factor model: CAPM Two factor model: Market factor + Equity issuance shock (ICS) A CAPM B 16 16 14 14 Market+ ICS SZ1 SZ1 IK5 10 SZ2 SZ4 SZ3 BM1 8 Predicted Predicted SZ2 12 SZ5 BM2 BM3BM4 IK1 IK4 IK3 IK2 6 6 8 C 10 12 BM5 BM3 16 IK1 IK5 SZ5 BM1 6 Latent factor (δ=∞) 8 D BM2 IK2 IK4 IK3 10 12 14 16 Latent factor (δ= 1) 12 SZ1 IK5 BM1 SZ2 SZ4 SZ3 latent w/ restriction 16 14 Predicted SZ4 BM4 10 6 14 BM5 SZ3 8 16 10 12 14 12 SZ1 10 IK5 SZ2 SZ3 SZ4 BM4 BM5 10 / 17 Model misspeci…cation with no restriction Model: Market + latent factor The APT restriction: A =1 CAPM B 16 16 14 14 Market+ ICS SZ1 Predicted Predicted SZ2 12 SZ1 IK5 10 SZ2 SZ4 SZ3 BM1 8 SZ5 BM2 BM3BM4 IK1 IK4 IK3 IK2 6 6 8 C 10 12 12 SZ4 BM4 10 BM3 SZ5 BM1 6 14 16 IK1 IK5 8 BM5 BM5 SZ3 6 8 BM2 IK2 IK4 IK3 10 12 14 16 Latent factor (δ=∞) 16 Predicted 14 12 10 SZ1 IK5 BM1 SZ5 8 6 6 8 SZ2 SZ4 SZ3 BM2BM3BM4 IK1 IK4 IK3 IK2 10 12 BM5 14 16 11 / 17 Model misspeci…cation with restriction Model: Market + latent factor The APT restriction: =1 A CAPM B 16 16 14 14 12 12 Market+ ICS SZ1 SZ1 IK5 10 SZ2 SZ4 SZ3 BM1 8 Predicted Predicted SZ2 SZ5 BM2 BM3BM4 IK1 IK4 IK3 IK2 6 6 8 C 10 12 16 Market+latent w/ restriction Predicted SZ1 IK5 BM1 SZ5 6 8 SZ2 SZ4 SZ3 BM2BM3BM4 IK1 IK4 IK3 IK2 10 12 BM5 14 8 D 12 6 SZ5 BM1 6 Latent factor (δ=∞) 14 8 BM3 16 IK1 IK5 6 14 16 10 SZ4 BM4 10 8 BM5 BM5 SZ3 BM2 IK2 IK4 IK3 10 12 14 16 Latent factor (δ= 1) 16 14 12 SZ1 10 IK5 8 SZ2 SZ3 SZ4 BM4 BM2BM3 IK1 BM5 BM1 SZ5 IK4 IK2IK3 6 6 8 10 12 14 16 12 / 17 Model misspeci…cation with tighter restriction Model: Market + latent factor The APT restriction: = 1 and = 0:01 B L a te n t fa c to r (δ=0 .0 1 ) 16 15 15 14 14 13 13 Ma r ke t+la ten t w/ r e str ictio n Ma r ke t+la ten t w/ r e str ictio n A L ate n t fa c to r (δ=1 ) 16 12 11 SZ1 10 SZ2 IK5 SZ3 SZ4 9 BM5 BM4 12 11 SZ1 10 SZ2 IK5 SZ3 SZ4 9 BM3 IK1 BM1 BM2 SZ5 8 BM3 IK1 BM1 BM2 SZ5 7 IK4 IK3 IK2 7 IK4 IK3 IK2 6 5 BM5 BM4 8 6 6 8 10 12 14 16 5 6 8 10 12 14 16 13 / 17 Model misspeci…cation with restriction The ICS factor and the latent factor are positively correlated 10 50 ICS-MI Latent Correlation = 0.27 0 - 10 1975 0 1980 1985 1990 1995 Year 2000 2005 2010 - 50 2015 14 / 17 Model misspeci…cation with restriction The HML and SMB factors and the latent factor are correlated SMB Latent 40 20 0 -20 -40 1975 Corr. = -0.13 1980 1985 1990 1995 Year 2000 2005 2010 40 2015 HML Latent 20 0 -20 -40 1975 Corr. = 0.26 1980 1985 1990 1995 Year 2000 2005 2010 2015 15 / 17 The mean-variance, alpha, and beta portfolios The alpha portfolio drives most of the variations of the mean-variance portfolio 4 Alpha Portfolio, SR. = 1.02 Beta Portfolio, SR. = 0.67 MV Portfolio, SR. = 1.13 3 Excess Return (%) 2 1 0 -1 -2 -3 1975 1980 1985 1990 1995 Year 2000 2005 2010 2015 16 / 17 Conclusion Very nice paper! Understanding model misspeci…cation is important for asset pricing test and portfolio choice 17 / 17