Financial Transfer Learning

Abstract

Stock return predictability is demonstrated to be largely global by showing that a common global model predicts stock returns better than local models estimated individually in each country, even when the global model has never seen stock data from the local country. Small but significant local components are shown to exist but cannot be estimated by purely local models. Instead, new transfer learning methods are introduced that take advantage of the existence of a global component to better estimate local components. Lastly, it is estimated that predictive coefficients are more than 90% global.