View article

The investment policy that worked best in the past may not work as well in the future; it may have been best in the past in part because it was lucky rather than because it is best in fact. If the best-performing portfolio is selected from historical data, the person selecting is data mining. The problem is to adjust past performance in estimating future performance to reflect such data mining. Three models of different degrees of generality are proposed, and data mining corrections for these models are presented. If it is assumed that the policies tested are drawn at random from a number of policies, plausible data mining corrections emerge. The exact correction depends on the assumed model of how expected returns, and then realized returns, are drawn.