We show that returns to value strategies in individual equities, industries, commodities, currencies, global government bonds, and global stock indexes are predictable in the time series by their respective value spreads. In all these asset classes, expected value returns vary by at least as much as their unconditional level. A single common component of the value spreads captures about two-thirds of value return predictability and the remainder is asset-class-specifc. We argue that common variation in value premia is consistent with rationally time-varying expected returns, because (i) common value is closely associated with standard proxies for risk premia, such as the dividend yield, intermediary leverage and illiquidity, and (ii) value premia are globally high in bad times.

Previous research finds that machine learning methods predict short-term return variation in the cross-section of stocks, even when these methods do not impose strict economic restrictions. However, without such restrictions, the predictions from the models fail to generalize in a number of important ways, such as predicting time-series variation in market and long-short characteristic sorted portfolio returns across multiple horizons. I show this shortfall can be remedied by imposing economic restrictions in the architectural design of a neural network model and provide recommendations for using machine learning methods in asset pricing. Additionally, I shed light on the intermediate and long-run dynamics of the return forecasts generated by this model.

We study the returns to characteristic-sorted portfolios up to five years after portfolio formation. Among a set of 56 characteristics, we find large pricing errors between the contemporaneous returns of new and old sorts, where new sorts use only the most recent observations of firm characteristics. These relative pricing errors are not captured by existing asset pricing models and have been overlooked by standard tests using only returns to new sorts. Thus, pricing errors across horizons provide new and powerful information to test asset pricing models. Further, we show that these pricing errors are strongly related to a characteristic’s market beta and connected to the difference in return between new and old stocks in the characteristic-sorted portfolios. We argue that investors can improve the performance of characteristic-based strategies by considering past observations of firm characteristics.