We perform a large scale stress test of modern Multiple Testing Procedures (MTPs) that are used to evaluate the superior predictive ability of many forecasting models, documenting how their size depends on methodological choices and data sample properties in environments plagued by data snooping. We specifically focus on the evaluation of technical trading rules, whose number has been exponentially increasing in recent years. We find that false discoveries (Type I errors) increase when the sample average return and volatility decrease. Notably, MTPs fail to provide adequate finite-sample control of aggregate error rates when tests are performed on bearish data samples or when market frictions (trading fees, liquidity costs, short selling restrictions) are ignored. To control for asymmetric data snooping bias, researchers should increase the statistical significance threshold used to detect superior forecasting performance in downward trending markets.