Machine Learning Models for Return Forecasts
Recent studies have applied Machine Learning in order to forecast the risk/benefits of stock returns. These models show better performance than traditional finance methods in stock predictions and have helped in many stock portfolios. However, how precise are these predictions?
Rohit Allena, Ph.D., assistant professor at the C.T. Bauer College of Business presented his work during the HPE Data Science Institute’s seminar series on August 4, providing insight into the ex-ante precision surrounding the application of ML-based stock return forecasts. By adapting popular ML models used in finance, Allena is able to estimate standard errors (SEs) and covariances of risk premium forecasts from Lasso, Ridge, Elastic-Net, and Neural Networks. Allena further explains how to use these ML models -- and confidence intervals -- to form better investment approaches.
The standard practice of forming a trading strategy involves the use of conventional high-low portfolios (HLs), sorting stocks into deciles according to their return predictions. These portfolios help determine how to delegate stocks for higher returns for the next period. Whereas a written prediction proxies for the next period’s returns, Allena demonstrates that the confidence interval around the forecast proxies for squared forecast errors, thus generating more gains.
With this finding, Allena proposes the more effective method of “Confident-HL” trading strategies, focusing on dividing stocks according to both their written prediction and confidence interval measurements. Allena provides an in-depth statistical explanation of how the “Confident-ML” strategy works, reviewing the findings of his paper titled, “Confident Risk Premiums and Investments using Machine Learning Uncertainties.”
Allena finds his research interests in asset pricing and market microstructure, with an emphasis on the econometrics of Machine Learning and Bayesian inferences.