Business Aligned Model Selection for Retail Demand Forecasting: A Multi-Criteria Evaluation of AI/ML versus Traditional Methods

Abstract

This study develops and tests a business-aligned framework for selecting retail demand-forecasting models by evaluating AI/ML against traditional methods using multiple, operationally relevant criteria. Motivated by the practical over-reliance on single accuracy metrics and the need to reflect inventory and planning realities, the research investigates whether AI/ML delivers superior performance and how multi-criteria decision-making can improve model choice in a long-lead, seasonal retail context.

A pragmatic mixed-method design combines a systematic literature review with an empirical case study using confidential sales, inventory, and lead-time data from Kingfisher PLC (UK DIY/home-improvement retail). The dataset covers ten stores over January 2022–December 2023 and is aggregated to a monthly cadence. After rigorous cleaning and ABC (Pareto) segmentation, two models are benchmarked using rolling-origin expanding-window validation: Linear Regression (baseline) and Random Forest (AI/ML). Performance is assessed using accuracy metrics (MAPE, MAE, RMSE) and forecast stability (Coefficient of Variation, CV), reported for Total and A/B/C segments.

Results show Random Forest consistently outperforms Linear Regression on all accuracy metrics across all segments, with the largest gains on A-class SKUs. While Linear Regression is marginally more stable for Total/A/B (lower CV), a hierarchical AHP model-selection framework ranks Random Forest highest, and sensitivity tests (equal weights; stability-heavy) preserve the same preference. The study contributes a replicable, transparent multi-criteria evaluation approach that links forecast performance to business impact (inventory, service, and working capital) and supports phased, ABC-aware adoption of AI/ML forecasting in practice.