Comparative study on supply chain demand forecasting based on traditional machine learning models—A Case Study Using the DataCo Dataset

School of Big Data and Supply Chain, Guangzhou College of Technology and Business, Guangdong, China

^* This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

Good demand forecasting can reduce operating costs and improve service levels. Once the traditional time series model encounters those slightly complex nonlinear laws, it will be more difficult to handle, so machine learning methods are used more and more. In this study, DataCo global supply chain data set was used to compare the three models of logistic regression, decision tree and random forest. First, the data was cleaned, and then the characteristics of time dimension and the statistics of rolling window (including the average value and standard deviation of seven and thirty days) were built. The delay feature was also introduced, and the two indicators of mean absolute error (MAE) and coefficient of determination (R²) were selected in the evaluation. The model completed the training on 80% of the data, and the remaining 20% was reserved for testing. The experimental results show that the random forest gets the minimum MAE and the maximum R², the decision tree is sandwiched in the middle, and logistic regression ranks last because of its linear characteristics. In addition, it can be seen that incorporating lag features and sliding window features benefits tree-based models significantly. Although the data set used is relatively old and there is no hyperparameter tuning, the experimental results show that the random forest has achieved a relatively stable balance between accurate prediction and easy deployment, which can provide a reference for enterprises when selecting demand forecasting tools.