Events

Prof. Alan Timmermann
Distinguished Professor of Finance and Economics
University of California San Diego

Hyperparameters determine the architecture of machine learning (ML) models and can greatly affect their forecasting performance, yet there is little consensus on how to choose the range and grid of hyperparameters to search over. We provide an extensive examination of which hyperparameters are most important for popular ML models’ out-of-sample forecasting performance using a large U.S. dataset on individual stock returns and firm characteristics. We find that some choices of hyperparameters virtually guarantee good out-of-sample return forecasts while others lock in poor forecasts. This poses a challenge because many empirical studies fail to provide details on how they set their hyperparameters. We also find that time-series validation methods do not offer a definitive solution to the dependence of out-of-sample return forecasting  performance on the underlying range of hyperparameters.