Machines Picking Emerging Market Stocks

Are models based on advanced machine learning adept at predicting returns for individual emerging market stocks? In the November 2022 version of their paper entitled "Machine Learning and the Cross-section of Emerging Market Stock Returns", Matthias Hanauer and Tobias Kalsbach compare abilities of machine learning models to predict emerging market stock returns. They consider nine alternatives: two traditional linear models (ordinary least squares and elastic net); two tree-based models (gradient boosted regression trees and random forest); and, five neural networks (one to five layers). Tree-based methods and neural networks identify non-linearities and variable interactions. They further consider a combination of the five neural networks and a combination of all tree-based plus neural network methods. For each model at the end of each month, they rank stocks into country-neutral fifths, or quintiles, based on next-month expected returns and reform a portfolio that is long (short) the quintile with the highest (lowest) expected returns. For tests of long-only net performance, they assume 1-way trading frictions are half the estimated bid-ask spread and apply trading cost mitigation rules. Using returns and 36 accounting/trading variables for 15,152 unique stocks from 32 emerging market countries as included in the MSCI Emerging Markets Index during July 1995 through December 2021 (with out-of-sample testing starting January 2002), they find that: