This study develops a transfer-learning framework that extends high-quality expected-return information from firms where proxies work well to the broader U.S. equity universe. The method first learns how firm characteristics map into high-signal expected-return proxies in a reliable “teacher” domain. It then applies a disciplined correction, using realized returns only as a calibration device, so that the predictions remain useful outside that original domain. The resulting measures cover U.S. equities from 1957 to 2023 and deliver near-unbiased forecasts relative to existing alternatives.
in Externe publicatie door Julio Crego, Jens Soerlie Kvaerner, and Marc Stam