EURO 2025 Leeds
Abstract Submission

650. A Hierarchical Approach to Forecasting Censored Demand in Lost-Sales Systems

Invited abstract in session TA-38: Forecasting, prediction and optimization 1, stream Data Science meets Optimization.

Tuesday, 8:30-10:00
Room: Michael Sadler LG19

Authors (first author is the speaker)

1. Diego José Pedregal Tercero
Organización de Empresas, University of Castilla-La Mancha
2. Juan Ramon Trapero Arenas
Business Administration, Universidad de Castilla-La Mancha. CIF: Q-1368009-E

Abstract

Accurate demand forecasting is critical for inventory management, especially in lost-sales systems where stockouts lead to unobservable demand. Traditional forecasting models relying solely on sales data systematically underestimate demand, resulting in biased forecasts and suboptimal inventory policies. To address this, we propose an extension of the Tobit Exponential Smoothing (TETS) model that integrates temporal hierarchies to enhance demand estimation across different time frequencies. This approach leverages state-space models to reconcile high-frequency data (e.g., hourly sales) with lower-frequency demand forecasts (e.g., daily replenishments). By capturing censoring constraints at multiple aggregation levels, our method significantly reduces forecasting bias and improves the estimation of demand variance—crucial for defining safety stock. Through simulated case studies and a real-world dataset from the M5 competition, we demonstrate that our model outperforms conventional exponential smoothing and Tobit-based methods, leading to lower lost sales and excess inventory. Additionally, we analyze the spiral-down effect, where naïve forecasting approaches exacerbate stockouts over time, and show how our method mitigates this issue. The results highlight the practical advantages of temporal hierarchies for demand forecasting, offering a robust, theoretically grounded framework for inventory optimization in supply chain management.

Keywords

Status: accepted

Back to the list of papers