一個基於基因演算法的自我回歸整合移動平均模型

Abstract

時間序列預測在做未來決策時扮演著重要的角色，而自我回歸整合移動平均 (Autoregressive Integrated Moving Average)是其中一個被廣泛使用的時間序列預測模型，但該模型仍有修正的空間，使預測能達到更佳的準確度。本篇論文提出一個基於基因演算法(Genetic Algorithm)的自我回歸整合移動平均模型，該模型與傳統自我回歸整合移動平均模型的差別為有對時間序列進行平滑(smoothing)處理以達到雜訊降低(noise reduction)的效果，並期望透過這樣的設計，可以提升預測準確度。我們採用許多研究使用的日經225股票之歷史資料來驗證，在實驗中，本研究提議之方法的平均絕對誤差百分比(Mean Absolute Percentage Error)相較於傳統自我回歸整合移動平均模型降低了34.19%，由實驗結果可證實本研究所提議之方法可以獲得較佳的預測準確度。

Time series forecasting plays an important role in making future decision. The autoregressive integrated moving average (ARIMA) is one of most popular time series forecasting models, but it still can be improved to obtain better accuracy. This thesis proposes an ARIMA model based on genetic algorithm. Compared to the traditional ARIMA model, the proposed model reduces noise among time series through data smoothing, and we hope this design can increase forecasting accuracy. The Nikkei 225 stock index which is used in many previous studies, is selected for experiments. Experimental results indicate that the MAPE of the proposed model is reduced by 34.19% as compared to the MAPE of the ARIMA, and it can verify that the proposed approach achieves better accuracy.