A0993
Title: Simple out-of-sample tests for asset pricing
Authors: Svetlana Bryzgalova - London Business School (United Kingdom) [presenting]
Alberto Quaini - Erasmus University of Rotterdam (Netherlands)
Ashish Sahay - Man AHL (United Kingdom)
Abstract: The aim is to show that traditional measures of out-of-sample model performance in asset pricing ignore model estimation risk and significantly underscore true standard errors. As a result, typical tests overestimate t-stats associated with out-of-sample Sharpe ratios, alphas, etc. A simple split-sample estimation design is proposed, that allows to effectively measure out-of-sample model performance and provides valid statistical inference for both in-sample and out-of-sample parameters. Empirically, the performance of popular linear factor models is revisited, and it is found that model estimation risk has a nontrivial impact on the out-of-sample tests. The results have important implications for the evaluation of asset pricing models (linear, nonlinear, and those estimated via machine learning techniques), the use of spanning tests in model comparison, and measuring risk-adjusted returns out of sample.
