Testing for the best alternative with an application to performance measurement

Suppose that we are searching for the maximum of many unknown and analytically untractable quantities or, say, the 'best alternative' among several candidates. If our decision is based on historical or simulated data there is some sort of selection bias and it is not evident if our choice is significantly better than any other. In the present work a large sample test for the best alternative is derived in a rather general setting. The test is demonstrated by an application to financial data and compared with the Jobson-Korkie test for the Sharpe ratios of two asset portfolios. We find that ignoring conditional heteroscedasticity and non-normality of asset returns can lead to misleading decisions. In contrast, the presented test for the best alternative accounts for these kinds of phenomena.