MACHINE LEARNING-BASED FORECASTING OF INCOMING CUSTOMER ORDERS: A CASE STUDY ON HEAVY-DUTY FILTER PRODUCTS

Abstract

Accurate demand forecasting plays a critical role in inventory management and production planning, particularly in the manufacturing of heavy-duty filter products, where customer order patterns are influenced by seasonal and external factors. This study proposes a machine learning–based forecasting framework to predict incoming customer orders using structured sales data. The analysis focuses on the 2021–2025 period, during which demand behavior was observed to be more stable compared to previous years. The study evaluates the effects of feature engineering and feature selection on forecast accuracy, with special attention to time-based, seasonal, and external economic variables. Random Forest, LightGBM, and Linear Regression models were developed and their performance was assessed using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), MAPE, and R2 score. Experimental results show that the inclusion of seasonal and exogenous features significantly enhances model performance, particularly for item types with pronounced cyclical trends. Furthermore, the study conducts hypothesis testing to assess the statistical impact of engineered features and model types on forecast accuracy. The findings demonstrate the applicability of machine learning in industrial forecasting scenarios, offering a replicable and scalable approach to data-driven decision making in procurement and inventory planning.