Backwards Search for the Most Important Firm/Stock Characteristics

Instead of searching among hundreds of firm/stock characteristics to identify those that best predict stock returns, what about first finding the stocks with the highest and lowest past returns and then examining the characteristics of those stocks? In his June 2023 paper entitled “Essence of the Cross Section”, Sina Seyfi identifies the strongest determinants of expected stock returns by:

1. Sorting stocks into fifths (quintiles) at the end of each month during the last 10 years based on monthly returns (120 sets of quintile portfolios).
2. Computing the average monthly value of each of 206 firm/stock characteristics among stocks in each quintile across the last 10 years.
3. Forming each month out-of-sample quintiles that are as similar as possible regarding these 206 average characteristics to the in-sample returns-sorted quintiles.
4. Studying variations of the 206 characteristics across these out-of-sample quintiles to identify the most important drivers of future stock returns.

This method allows for non-linearities and interactions among characteristics, which a conventional linear regression method does not. Using returns and characteristics data for publicly listed U.S. common stocks and the U.S. risk-free rate as available during 1926 through 2021, he finds that:

(more…)