How Global is Predictability? The Power of Financial Transfer Learning

Abstract

We demonstrate that a common global model predicts stock returns more effectively than local models estimated individually for each country, even when the global model excludes local data. We introduce a "generalized elastic net" (GENet) to estimate a combined global-and-local model, showing theoretically and empirically that it efficiently (i) transfers information from global data to local countries and (ii) detects unique local components. The resulting model is 94% global-nearly the same function predicts stock returns across countries and over the past century. These findings have broad implications for asset pricing, highlighting the stability of the stochastic discount factor as a function of characteristics.

Keywords: Predictability, asset pricing, transfer learning, machine learning JEL Codes: C53, C58, F3, G1, G4

JEL Classification: C53, C58, F3, G1, G4