本文探討模式更新方式及資料轉換與否對季節性自我迴歸整合移動平均模式(SARIMA)及Holt-Winters模式(HW)在預測績效上的影響。在模式更新方面,本研究討論不更新、累積更新及逐月更新對模式在預測績效上的影響。在資料轉換方面,本研究討論將資料進行Box-Cox轉換後對模式在預測績效上的影響。本研究也進一步比較SARIMA與HW模式於研究課題預測績效上的差異。實證結果發現,資料轉換及模式更新對於SARIMA及HW模式於預測績效上的提升皆有貢獻。此外,SARIMA模式較HW模式在研究課題上具有較佳的預測能力,主要的原因在於SARIMA能擷取到較多時間上的影響因素。一般時間序列模式較少談論有關模式更新及資料轉換的議題,本研究提供使用此兩策略實證上之經驗。
The purpose of this study is to present the effects of update fashion and data transformation on short-term railway passenger demand forecasting by applying Seasonal AutoRegressive Integrated Moving Average (SARIMA) and Holt-Winters (HW) models. On update fashion, this study utilizes three concepts: do nothing, rolling window data learning method, and moving window data learning method. On data transformation, this study applies Box-Cox transformation. We observe the differences in predicting performance after applying these modeling strategies. In addition, this study also checks whether SARIMA is preferred to HW or vice versa. There are three major findings. First, data transformation is found to be beneficial to both SARIMA and HW. Second, moving window data learning method is a useful and economic fashion to implement update. Third, SARIMA is found to outperform HW in the study because SARIMA is more robust for capturing various temporal features. Although time series models are popular in the literature, topics of update fashion and data transformation are seldom discussed. This study renders empirical findings of utilizing these two modeling approaches.