Dangers of Data-Driven Inference: The Case of Calendar Effects in Stock Returns

摘要: Economics is primarily a non-experimental science. Typically, we cannot generate new data sets on which to test hypotheses independently of the that may have led particular theory. The common practice using same set formulate and introduces data-snooping biases that, if not accounted for, invalidate assumptions underlying classical statistical inference. A striking example data-driven discovery presence calendar effects in stock returns. There appears be very substantial evidence systematic abnormal returns related day week, week month, month year, turn holidays, so forth. However, this has largely been considered without accounting for intensive search preceding it. In paper use 100 years daily bootstrap procedure allows us explicitly measure distortions inference induced by data-snooping. We find although nominal P-values individual rules are extremely significant, once evaluated context full universe from such were drawn, no longer remain significant.