Can machine learning identify sector-level financial ratios that predict sector returns?

Abstract

Academics and practitioners use fundamental ratios to evaluate an industry's financial and operating performance. WRDS has large firm-level and sector/industry-level databases that provide over 70 financial ratios that are applied in finance classes to compare and assess relative financial performance. However, there has been a lack of sophisticated econometric methods assessing these ratios' importance in predicting sector-level stock return performance. Using Elastic Net methods, we identify financial ratios that significantly forecast out-of-sample sector stock returns and find that these predictive ratios vary across sectors. We form long and long-short portfolios that consistently outperform the market over-time. Long portfolios generate significant alpha and large utility gains, boost the Sharpe and Sortino ratios, and a cumulative investment portfolio exceeds the market benchmark by five times. Long-short portfolios generate Fama-French 4-factor and 6-factor alphas between 4–9% and cumulative investment gains from six to fourteen times. Our research establishes that machine learning can identify financial ratios that significantly predict sector returns and generate profitable portfolio allocation.