Divide and Conquer: Financial Ratios and Industry Returns Predictability

Abstract

We study whether a large set of financial ratios provides valuable information about future excess stock returns. Confronted with a data-rich environment, we propose a novel ``divide and conquer" methodology that allows to efficiently retain all of the information available to investors. In particular, our method does not assume, a priori, that some of the financial ratios may be irrelevant or easily reducible. We compare our methodology against standard, recursive sparse and dense predictive regression methodologies, as well as benchmark forecast combination strategies and non-linear machine learning methods. Forecasts based on our method, not only outperform in out-of-sample predictive comparisons, but translate into out-of-sample economic gains that are greater than the historical averages and all of the competing forecasting strategies. Our results lend strong support for using accounting-based information in forecasting stock returns, both at the industry and at the market level.

Keywords: Financial ratios, forecast combination, machine learning, returns predictability, data-rich models, industry portfolios.

JEL Classification: C11, G11, G17, C32, C53, C55.