基於機器學習之需求預測：資通訊產業之實務應用

Abstract

良好的存貨管理政策需要準確的需求預測。本研究利用機器學習之觀念，自動找出時間序列資料最佳需求預測模型的流程，此流程分別是找出離群值並能夠以三種不同方法處理之，處理完畢後挑選各個候選模型的最佳參數，候選模型包含現行方法、改良版現行方法、線性迴歸模型、正規化線性迴歸模型、移動平均法、指數平滑法和整合移動平均自迴歸模型，並在最後根據驗證誤差挑選最佳需求預測模型。本研究將上述模型訓練與預測方法整合，並實作一決策支援系統，應用在臺灣一家資通訊產業之領導廠商。經由實際資料驗證，本研究使用之預測方法確實可以提昇資通訊零件耗用之預測準確度。本研究亦就預測方法的實務應用做設計與討論。

Demand forecasting enhances the policy-making of inventory management. We propose a procedure of automatically finding the best model for time series data using the concept of machine learning in this research. The procedure is as the following. Firstly we use three different ways to deal with outliers. Secondly, we choose the best parameters for candidate models, which include the current method, improved current method, linear regression, regularized linear regression, moving average, exponential smoothing, and ARIMA. Finally, we choose the best demand forecasting model among all candidate models based on their validation errors. We implement this procedure as a decision support system for a leading ICT company in Taiwan. Through the real data, it is verified that our procedure can improve the forecasting accuracy. We also discuss the predicting method and practical application and design.