Risk-Averse Two-Stage Stochastic Programming Using Hybrid Scenario Generation

Retail chains need to define orders before knowing the demand, and the quality
of these decisions in Stochastic Two-Stage Programming (2SSP) depends critically on the
fidelity of scenarios that represent uncertainty. This work proposes a data-
strategy driven that combines (i) a nonlinear ML estimator for the conditional average of demand and (ii)
a Gaussian copulation on the residues, preserving the joint dependence between items, applied
to a multi-product newsvendor with a medium objective–CVaRα (Rockafellar-Uryasev). The method is
compared against Sample Average Approximation (SAA) and independent triangular distributions
(Specialist). In out-of-sample evaluation on real Walmart data (200 items: 10 stores × 20
departments), the hybrid generation dominates both references with statistical significance in
Wilcoxon tests, reducing the total cost and raising the rate of attendance.