We measure regression model performance (as perceived by a conservative investor betting on a complete market) via the out-of-sample expected utility for the allocation that maximizes expected utility under a most adverse model-consistent measure. This robust allocation is optimal under the minimum generalized relative entropy (MGRE) measure. We analyze our performance measure in the (practical) case of an investor whose utility function is a member of a three-parameter logarithmic family with a wide range of possible risk aversions. Here, our performance measure is independent of the market prices, and the MGRE measure minimizes the Kullback–Leibler relative entropy.

We consider the 12-month moving average aggregate default rate of S&P-rated US-bonds. We estimate the conditional probability distribution of this default rate as a function of a weighted average bond rating, a lagged default rate and a preliminary predictor that is based on lagged new issuance. Our modeling approach is asymptotically optimal for an expected utility maximizing investor. The resulting conditional probability density is consistent with our intuition. We measure the model’s performance by the out-of-sample expected utility. According to this measure, our model clearly outperforms a simple regression model, a regression model with ARMA error terms and a Poisson model.