Abstract

Agile inventory management requires timely, accurate, and stable short-term forecasts to support rapid yet reliable replenishment decisions. In retail distribution networks, particularly those with multiple branches and a large number of stock keeping units (SKUs), forecasting errors and excessive forecast volatility can lead to operational inefficiencies, frequent replanning, and increased costs. This study investigates the use of time-series machine learning to enhance agile replenishment planning by improving next-month readiness in a multi-branch bookstore context. Using a publicly available dataset of monthly inventory movements, the replenishment problem is formulated as a one-step-ahead forecasting task at the branch–SKU level. Inbound inventory transfers are treated as a proxy for replenishment demand, reflecting practical planning conditions in the absence of direct sales or inventory-on-hand data. A classical Exponential Smoothing (ETS) model is employed as a baseline and compared against a global Light Gradient Boosting Machine (LightGBM) regression model. The machine-learning approach integrates lagged demand features, seasonal indicators, pricing and margin information, and branch-level context, and is trained using a strictly time-aware validation strategy. Experimental results show that the LightGBM model consistently outperforms the ETS baseline in terms of forecasting accuracy, achieving lower Mean Absolute Error and Root Mean Squared Error on a held-out test set. In addition, an agility-oriented stability metric reveals that machine-learning-based forecasts are substantially smoother than those produced by the classical model, indicating reduced sensitivity to short-term fluctuations. Visual analysis further confirms that the global machine-learning model generalizes better across SKUs and branches, producing predictions that are more tightly aligned with actual replenishment quantities. The findings demonstrate that time-series machine learning can effectively support agile replenishment planning by balancing predictive accuracy and planning stability. The study also highlights a practical trade-off between responsiveness and smoothness, suggesting that machine-learning forecasts are well suited as a baseline planning tool, complemented by classical methods for exception handling. Overall, this work provides a reproducible and data-driven framework for integrating machine learning into agile inventory decision-making.

Keywords: Agile Inventory Management, Replenishment Forecasting, Time-Series Machine Learning, Lightgbm, Retail Supply Chain

How to Cite:

Mashel, J. & Purba, I. B., (2026) “Adaptive Inventory Replenishment Forecasting Using Time-Series Machine Learning for Multi-Branch Bookstore Operations”, Agile Management 1(4), 242-257. doi: https://doi.org/10.63913/am.v1i4.139