時間序列預測常被應用在各種領域,如財務經濟和天文地理等。大部分時間序列預測模型為單一函數,難以準確預測趨勢走向變動較大的時間序列的未來值,本研究以基因表示規劃法(Gene Expression Programming)為基礎,發展多函數時間序列預測模型:主要做法是根據各函數在歷史資料的預測能力,決定各函數的權重,建構預測模型。分析之資料為太陽黑子活動量(Sunspots)時間序列、Furnas 時間序列及Venice Lagoon時間序列,實驗採時間序列分析(Time series analysis)和符號式迴歸(Symbolic Regression)兩種做法,結果證實本研究提出的方法,效果優於傳統基因表示規劃法及EGIPSYS提出的方法;另外研究分析顯示本方法計算成本高,建議應用本方法於中短期時間序列。
Time series prediction has been widely used in various fields such as finances, economy and physical phenomena. However, most prediction models only contain one single function. A high level of accuracy of dynamic time series prediction cannot be easily achieved. The purpose of this paper is to develop a system integrates multiple functions for time series prediction, named AWFM which is based on Gene Expression Programming. Its main idea is to allocate adaptive weight to each function according to the previous prediction accuracy. To examine the effectiveness of AWFM, Sunspots series, Furnas series and Venice Lagoon series have been applied, and two issues are focused on: Time Series Analysis and Symbolic Regression Problems. The result shows that AWFM has higher performance than basic Gene Expression Programming and the method of EGIPSYS. Moreover, the study reveals that our method has higher computational cost. Thus, applying AWFM on short and medium period of time series is recommended.