Optimization of Prices and Promotions for an FMCG Distributor: Double Machine Learning and Decision-Focused Learning

Abstract

Pricing and discount management remain the most powerful and fastest-paying lever for raising profit in the fast-moving consumer goods (FMCG) sector, yet estimating the demand response to price from observational data is hampered by a systematic bias: discounts are not assigned at random but in response to seasonality and the promotional calendar, which by themselves raise demand. This paper proposes and empirically validates a three-stage pipeline for a distributor of juices and nectars (Samarkand region, Uzbekistan): (i) causal estimation of price elasticity over 85,720 order lines by double (debiased) machine learning (DML); (ii) per-SKU profit optimization under margin and price-change constraints; and (iii) decision-focused learning (Smart Predict-then-Optimize, SPO+), in which the predictive model is tuned to minimize decision loss rather than prediction error. To assess the causal inference rigorously, the data are equipped with a known ground-truth elasticity for every SKU. On these data, DML recovers the true elasticity with a mean absolute error of 0.35 against 0.85 for naive regression; the sign of the effect is correct for all 85 SKUs, and 91% of the confidence intervals cover the true value. The resulting elasticities, under ±15% price and ≥10% margin constraints, yield a 10.9% gain in weekly profit through bidirectional price changes. SPO+ exhibits the effect characteristic of decisionȔfocused learning: lower decision regret (0.54% vs. 0.58%) at a higher elasticity-prediction error. The work makes a methodological contributionȔcoupling causal elasticity estimation with decision-focused optimizationȔand carries applied and regional novelty for FMCG analytics in Central Asia.