Data snooping bias in the performance evaluation of asset allocation strategies

There are considerable asset allocation models which have been devoted to the issue of handling estimation error. However, the effective performance of these models ignores the effects of data snooping bias. For the first time, this paper examines and tests the effect of data snooping bias in allocation strategies, using the reality check and the superior predictive ability, as well as their stepwise extensions. First, a universe of 4653 portfolio strategies based on the choice of parameters is constructed for test. Meanwhile, the common equal weight strategy is employed as the benchmark strategy. The paper evaluates across four datasets for weekly and daily data of Chinese stock market in terms of mean return, Sharpe ratio, and certainty-equivalent return. Even though individual strategy significantly outperform a 1/N strategy at both daily and weekly frequencies, when considered in isolation, their outperformance in whole period and the first half period, generally, does not remain significant after correcting for data snooping. The simulation further confirms that these tests can eliminate the data snooping bias in the evaluation of asset allocation models. © 2018, Editorial Board of Journal of Systems Engineering Society of China. All right reserved.