Distributionally Robust Decision Focused Learning with Conformally Calibrated Ambiguity Sets

Decision-focused learning often relies on point predictions or treats learned distributions as exact, leaving downstream decisions vulnerable to model error. We propose a distributionally robust decision-focused learning framework that predicts a conditional distribution and conformally calibrates a Wasserstein ambiguity set around it. The resulting distributionally robust optimization problem is differentiated end-to-end, aligning distributional prediction with robust downstream decision quality. The calibrated ambiguity set contains the latent conditional distribution with finite-sample marginal coverage, yielding an out-of-sample decision certificate under exhchangeability assumption. For losses affine in the uncertain parameter, the robust layer admits a tractable regularized formulation. Experiments on portfolio allocation demonstrate improved downstream performance over two-stage and stochastic decision-focused baselines.