Retail Freight In Forecasting

Abstract

This applied data science project addresses the problem of forecasting inbound merchandise reception in a large-scale North American retail network under strict operational planning constraints. In this setting, workforce rosters must be defined two weeks in advance, while reliable information about inbound deliveries typically becomes available only a few days before execution, creating persistent misalignment between planned labor capacity and realized operational workload. The objective of the study is to design a reproducible and scalable forecasting methodology capable of estimating weekly inbound volumes at a two-week anticipation horizon and coherently distributing these forecasts into daily execution plans while preserving weekly totals. To achieve this, the project proposes a constraint-aware forecasting architecture that combines probabilistic weekly forecasting models—DeepAR and LightGBM—with a deterministic daily allocation mechanism. Weekly forecasts are generated using only information observable at least fourteen days in advance, including lagged summaries, calendar effects, movable events, and fixed store attributes, and are evaluated under a rolling-origin validation scheme using quantile-based metrics aligned with operational decision-making needs. Daily planning is formulated as a constrained disaggregation problem rather than an independent forecasting task, relying on empirically learned intra-week profiles, cluster-based shrinkage, and temporally weighted historical information to stabilize allocations without introducing additional uncertainty at the weekly level. The results show that probabilistic weekly forecasting is feasible and informative under severe information constraints, that DeepAR and LightGBM capture complementary structural and temporal patterns, and that coherent daily schedules can be derived despite high sparsity and intermittency in daily arrivals. This study does not introduce new forecasting algorithms; instead, it contributes to data science by demonstrating how existing probabilistic models can be systematically composed into a reproducible, decision-oriented architecture aligned with fixed planning horizons. The proposed methodology provides a solid foundation for future operational deployment in retail logistics and illustrates how forecasting theory can be translated into structured decision-support frameworks under real-world organizational constraints.

Date of Award	29 Jan 2026
Original language	English
Awarding Institution	Pontificia Universidad Javeriana