Understanding how individuals make economic choices underpins much of applied microeconomics, yet standard models often rely on restrictive assumptions about preferences and homogeneity. While tractable, discrete choice models such as multinomial logit and random coefficient logit impose functional forms that may obscure real behavioral variation. This paper introduces a Mixture of Experts (MoE) framework as a flexible, nonparametric alternative for modeling individual choice behavior. MoE uses probabilistic gating functions to assign observations to latent subpopulations, each governed by its own expert model, enabling endogenous segmentation and localized functional responses without predefined utility specifications. Using high-frequency transaction data, we show that MoE improves out-of-sample predictive accuracy by up to 15 percentage points over benchmark models and identifies four distinct consumer segments with sharply differing price elasticities and attribute sensitivities. These results uncover nonlinear responses to discounts and heterogeneous substitution patterns. Our findings have implications for welfare estimation, pricing strategy, and policy design—particularly where regulatory interventions depend on understanding behavioral heterogeneity and distributional impacts. By bridging machine learning and economic theory, this study advances structural modeling tools for applied microeconomic.
Diego Vallarino (Thu,) studied this question.