End-to-End Censored Demand Modeling

Natchapol Lebkrut, Apithep Wongpiyachon

Keywords: Demand Forecasting, Censored Data, Transformers, Latent Variable Recovery, Missing Not At Random (MNAR)

TL;DR: We propose the End-to-End Censorship-Aware Transformer (ECAT), a framework that unifies latent demand recovery and multi-horizon forecasting using a joint Tobit-Likelihood loss to handle non-random stockout censoring in retail supply chains.

Abstract:

Accurate demand forecasting for perishable retail goods is frequently compromised by censored sales data, where stockouts lead to an underestimation of true latent demand. Existing methods often treat missing data as random or decouple imputation from forecasting, leading to significant error propagation and biased predictions in Missing Not At Random (MNAR) scenarios.

In this paper, we introduce the End-to-End Censorship-Aware Transformer (ECAT) framework, a novel differentiable architecture designed to unify latent demand recovery and censorship-robust forecasting. ECAT incorporates a shared multi-head transformer encoder and a specialized censorship-aware block that explicitly integrates stockout indicators to model the MNAR mechanism. We employ parallel probabilistic decoder heads optimized by a custom joint Tobit-Likelihood loss function, which applies survival functions during stockout periods to bound latent demand estimates.

We validate our approach using the FreshRetailNet-50K dataset and high-fidelity food chain simulations. Our preliminary results demonstrate that by optimizing demand recovery and forecasting simultaneously, ECAT reduces prediction bias and improves multi-horizon accuracy compared to state-of-the-art two-stage baselines.

Submission Number: 11