FORECASTING INVENTORY MANAGEMENT USING MACHINE LEARNING

Abstract

This independent research aims to forecast sales volume using time series analysis and to study factors affecting inventory management through machine learning techniques, following the CRISP-DM methodology from business understanding through deployment. The data used in this study consists of historical sales records from January 2022 to June 2025, obtained from VFT Fastening Co., Ltd., focusing on the highest-selling product type IF4048. The variables include daily sales volume, inventory stock, Thai and Korean public holidays, number of defective products, transportation data, supplier purchase orders, and product lead time. Sales forecasting results using five models — ARMA, ARIMA, SARIMA, LSTM, and Prophet found that the SARIMA model achieved the lowest AIC value of 397.523, indicating the most appropriate statistical model structure. However, the LSTM model yielded the lowest error values, with MAE of 222,441 and RMSE of 243,546, demonstrating the highest forecasting accuracy in terms of error reduction against actual data. For inventory management factor analysis using three algorithms in WEKA, Decision Tree (J48) achieved the highest accuracy at 99.1218%, followed by Support Vector Machine (SMO) at 95.4995% and Naïve Bayes at 82.4369% respectively. The key factors identified by J48 as influencing inventory risk were stock quantity (STOCK), daily sales volume (UNITS_SOLD), and number of defective products (DEFECT). In conclusion, LSTM is the most suitable model for sales forecasting in terms of predictive accuracy, while Decision Tree (J48) is the most appropriate model for supporting inventory management decisions, reducing risks from both stockouts and overstock situations, and enhancing the company's inventory planning efficiency in the future.

Keywords: Inventory Management, Time Series Analysis, Machine Learning, Decision Tree, LSTM, SARIMA