近幾年來，非線性時間數列分析一直是時間數列及計量經濟學者所熱衷的研究主題之一，而非線性時間數列結構改變的研究也越來越受到重視。其中的門檻自迴歸模式，雖具有線性模式所不能配適的特性，但模式建構的問題，一直是其在發展應用上的瓶頸。本研究擬以門檻自迴歸模式建構的流程並結合遺傳演算法的最佳化搜尋技術，架構出時間數列遺傳演算法，藉此演算法則及程序，全域性地搜尋最佳的門檻自迴歸模式。 / Non-linear time series analysis is a research topic which the schalors of time series and econometrics are intent on, and the research of structure change of non-linear time series is attentive. Threshold autoregressive model (TAR model) of non-linear time series has some characters which linear model fail to fit while the problem of how to find an appropriate threshold value is still attracted many researchers attention. In this paper, we present about searching the parameters for a TAR model by genetic algorithms.

時間序列是用來預測未來趨勢的一種重要技術，然而在實務上建構時間序列模型時，參數很難有效估計。原因可能來自於時間序列本身的模糊性質，而導致參數的不確定性使得預測結果產生極大誤差。如果將參數模糊化引進時間序列的模型中，往往過於複雜。本論文提出相對移動率為新的模糊時間序列建構方法，讓原本具有模糊性質的時間序列經由反模糊化(defuzzification)後，以點估計的方式估計起始中心點，經由適當的修正調整為較佳的中心點以及半徑，建立有效的區間時間序列。並將相對移動率引進門檻自廻規模型中，取代原有之門檻值設定，並建立區間時間序列。最後，我們使用台灣加權股價指數為例，以本論文所提出之方法進行區間預測及效率評估。 / The time series is an important technology that is used to predict future trends, however in the real world, parameter is difficult to estimate effectively when we construct a time series model due to the of the fuzzy property of the times series data. The estimated parameters in the time series will cause a big error due to the uncertainty of fuzzy data. It is too complex to introduce the fuzzy parameters into the time series model. In this thesis, we propose relative moving ratio as a new criteria in constructing procedure of an interval time series. We defuzzify a fuzzy data and use point estimation to obtain an initial center, then we adjust the center and radius making it more appropriately. The resulting center and radius is then become an interval time series that can be use to forecast an interval data. We also apply relative moving ratio in threshold autoregressive models by replacing the threshold in constructing interval time series. Finally, in empirical studies chapter, we use Taiwan weighted Stock Index as examples to evaluate the performance of the proposed two methods in building the interval time series.