We show that out-of-sample tests used in the time-series predictability literature may suffer from test-size problems related to the common practice of exogenous specification of critical parameters, such as the choice of predictive variables, traded assets, and in-sample estimation periods. We perform specification searches across these parameters and find that rejections of the null hypothesis of no predictability are very sensitive to minor variations in parameter specification. We perform simulations using random factors to determine if the observed predictability in the data is real. The simulations suggest that much of the literatures' out-of-sample evidence of time-series based predictability is consistent with data-snooping.