Decision-Improved Predictions

Data-driven optimization faces uncertain costs that can be predicted from historical data. Recent work in decision-focused learning (DFL) improves the decision performance directly, beyond standard ''predict-then-optimize'', but often leads to nonconvex losses and introduces greater computational burden. Most critically, DFL no longer recovers the Bayes-optimal predictor, which harms interpretability and safety in consequential settings. We introduce decision-improved predictions that incorporate decision-loss into the prediction error and therefore improve the decision-regret of PTO. We propose reweighing the mean-squared error by context-predicted decision regret, RWERM+, which we show preserves consistency for Bayes-optimal predictions when well-specified. We extend to orthogonal estimation and lift our method to projected-regret-reweighted basis expansion, which further improves under misspecification. Our methods can all be implemented via reweighted least squares, which enables flexible extension. In experiments, our method achieves similar decision-regret improvements as decision-focused learning does, at a substantial improvement in prediction error.