Demand forecasting and inventory improvement in supply chain management using hybrid boosting ensemble techniques

Abstract

Effective supply chain management in retail depends heavily on accurate demand forecasting and inventory improvement to ensure balanced stock levels and minimize costs. This study proposes a novel framework for a single echelon inventory management problem, that integrates a hybrid boosting ensemble model applied to two case studies (with and without handling uncertainty) with the Order-Up-to-Level inventory policy to enhance forecast precision and operational efficiency within the retail environment of Rossmann stores. Orders are aggregated for multiple items to reduce ordering costs, improving overall cost efficiency. The objective is to determine key inventory parameters, including reorder points, safety stock, and total inventory costs. The methodology follows a two-phase approach. First, a hybrid ensemble of LightGBM, CatBoost, and XGBoost is developed using multivariate data from the Rossmann dataset, which includes sales figures, store attributes, time-related variables, and promotional factors. Prior to training, fuzzy generalization is applied to handle uncertainty in the data. The ensemble operates sequentially: LightGBM provides initial predictions, CatBoost reduces residual errors, and XGBoost performs final refinements. Model tuning is conducted through grid search combined with 5-fold cross-validation. These forecasts are then incorporated into the Order-Up-To-Level inventory policy to determine improved ordering decisions. The framework is benchmarked against traditional and singlemodel approaches, which often struggle with demand volatility and integration with inventory systems. The hybrid ensemble outperforms these methods by leveraging the strengths of each model—computational efficiency from LightGBM, categorical handling by CatBoost, and regularization from XGBoost. Empirical validation using data from 1,115 Rossmann stores over 942 days demonstrates substantial improvements in key performance metrics such as Root Mean Squared Error, Mean Absolute Error, and Pred(x=10%) compared to baseline models such as ARIMA, Moving Average, Simple Exponential Smoothing, and single-model machine learning approaches (LightGBM, XGBoost, and CatBoost). The proposed system results in lower forecast errors, reduced inventory costs, fewer stockouts and overstock incidents, and higher service levels. In summary, the integration of hybrid boosting, fuzzy generalization, and the Order-Up-To-Level policy provides a robust, data-driven solution to the challenges of demand uncertainty and inventory control in retail supply chain management. The proposed framework offers tangible improvements in operational performance, highlighting its potential for broader adoption in real-world retail applications.

Date	11 Aug 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Yvan Beauregard (Supervisor)