Most published return predictors failed out of sample, in an update to 2021
What it found: Re-tested with US data to 2021, over a third of 29 newer market-forecasting signals failed even when fitted to the whole period, and half the rest did poorly on new data. A few still worked.
What it means for you: Treat published market-forecasting signals sceptically: most haven't held up on later data.
The figures and the limit
Goyal, Welch and Zafirov re-tested 29 variables from 26 papers published after Goyal and Welch (2008), plus the original 17, with data to the end of 2021. More than one-third of the new variables were no longer significant even in sample. Of those that were, half performed poorly out of sample. A small number still performed reasonably well both in and out of sample.
The limit: US equity premium data. Valuation ratios are among the variables tested. Updates the 2008 study cited at Levels 2 and 4 rather than repeating it. Statistical performance alone doesn't show that a timing rule would have paid after costs.





