In this paper, we considered a demand forecasting problem for railroad car maintenance parts that a maintenance company in Korea has been trying to resolve. Due to sporadic demand nature of railroad car maintenance parts, the demand cannot be estimated by using normal approaches such as time-series analysis or multiple regression analysis. Therefore, we applied various machine learning methods and identified good estimators for them. Specifically, we classified maintenance parts into several clusters using K-means method based on average demand interval and coefficient of variation, which can be calculated for each item. Then, for each cluster, we identified a proper estimator by testing decision tree, random forest, and neural network. By adopting these results, we expect that we can improve the forecasting capability of the company and reduce the production lead time up to 40 to 60 days.
