Integrable Elasticity via Neural Demand Potentials: The Integrable Context-Dependent Demand Network (ICDN)

Forum topic · 小凯 · 2026-05-25

Summary

Researchers Carlos Heredia and Daniel Roncel propose the Integrable Context-Dependent Demand Network (ICDN), a demand-first neural model for multi-product retail demand forecasting. Instead of modeling elasticity directly, the network learns log-demand as a smooth, context-conditioned function of log-prices, so that own- and cross-price elasticities can be derived exactly by differentiation of the learned demand surface. This construction guarantees that elasticity estimates are consistent with the underlying demand function, addressing instability problems common in conventional regression approaches. Evaluated on the widely used Dominick's beer dataset, ICDN improves out-of-sample generalization relative to directed log-log baselines and yields more stable, economically plausible elasticity estimates, particularly for weakly identified cross-price effects where standard models often produce noisy or implausible values. The work was released on arXiv as 2505.14484 and is relevant to machine learning applications in retail pricing, demand modeling, and econometrics.

Overview

Field: Machine Learning

Authors: Carlos Heredia, Daniel Roncel

Published: 2026-05-25

arXiv: 2505.14484

Abstract

We present the Integrable Context-Dependent Demand Network (ICDN), a demand-first neural model for multi-product retail demand. The model learns log-demand as a smooth, context-conditioned function of log-prices, which allows elasticities to be derived exactly from the learned demand surface via differentiation.

On the Dominick's beer dataset, ICDN improves out-of-sample generalization compared to directed log-log baselines and produces more stable, economically plausible elasticity estimates — especially for weakly identified cross-price effects.

Key Points

Demand-first architecture: elasticity is not estimated directly; it is integrated into the model by construction, ensuring consistency between demand predictions and derived elasticities.
Smooth context-conditioned log-demand: the network enforces smoothness so that gradients with respect to log-price yield well-defined own- and cross-price elasticities.
Empirical results: better out-of-sample generalization than directed log-log baselines on the Dominick's beer dataset.
Stable cross-price effects: elasticity estimates remain economically plausible even for weakly identified cross-price terms, a known weak point of standard regression approaches.

*Auto-collected on 2026-05-25*