Multi-SKU Retail Sales Forecasting With FP-Growth Association-Rule Features: A Comparative Study of Echo State Network and LSTM

Abstract

This study examines whether cross-product signals derived from market-basket transactions can enhance multi-SKU retail sales forecasting when embedded as auxiliary features in global sequence models. Retail transaction records from January 2024 to November 2025 were utilized. Frequent itemsets and association rules were mined using FP-Growth, and the resulting co-purchase structure was transformed into forecasting features. Daily sales were aggregated into weekly multivariate series and normalized using Min–Max scaling. Two feature-augmented forecasting architectures were evaluated: an Echo State Network (ESN+FP) and a Long Short-Term Memory network (LSTM+FP). Hyperparameters were tuned with Optuna (30 trials) using validation sMAPE, and per-SKU performance was assessed using MAE and sMAPE. Across 43 Food & Beverage SKUs, the lowest sMAPE was obtained by ESN+FP for 28 SKUs, whereas LSTM+FP performed best for 15 SKUs, indicating substantial heterogeneity in demand dynamics. Although slightly lower average errors were achieved by LSTM+FP, ESN+FP yielded superior results for several volatile yet repetitive series, while more stable trend patterns were better captured by LSTM+FP. An adaptive, per-SKU model selection strategy is therefore supported