Investors' searches for successful forecasting models cause the data generating process for financial returns to change over time, which means that individual return forecasting models can, at best, hope to uncover evidence of 'local' predictability. We illustrate this point on a suite of forecasting models used to predict US stock returns, and propose an adaptive forecast combination approach. Most of the time the forecasting models perform rather poorly, but there is evidence of relatively short-lived periods with modest return predictability. The short duration of the episodes where return predictability appears to be present and the relatively weak degree of predictability even during such periods makes predicting returns an extraordinarily challenging task.
