We develop a simulation-based procedure to test for stock return predictability with multiple regressors. The process governing the regressors is left completely free and the test procedure remains valid in small samples even in the presence of non-normalities and GARCH-type effects in the stock returns.