本研究提出一種創新的方法，用於識別多維時間序列中的變動點。其核心思想是透過多重隨機線性投影，將多維時間序列轉換為一組單維時間序列。接著，對每一條投影後的序列套用變動點偵測程序，從而得到一組潛在的結構變化位置。隨後，利用分群技術對這些候選位置進行分析，以同時估計變動點的數量與位置。為了評估所提出方法的有效性，本文進行了多項模擬研究。此外，亦進行了一項實證分析，資料涵蓋八個國家的股票市場指數、政府公債殖利率以及外匯市場，期間自 2006 年 2 月 7 日至 2024 年 12 月 30 日，以說明該方法在實務上的表現。數值結果顯示，相較於既有方法，所提出的方法在提升變動點偵測準確性的同時，亦具有較低的計算成本。實證結果亦與研究期間內的重要全球金融與經濟事件相吻合。;This study proposes an innovative approach for identifying change-points in multivariate time series. The key idea is to transform the multivariate series into a collection of univariate series via multiple random linear projections. Change-point detection procedures are then applied to each projected series， yielding a set of candidate locations for structural breaks. Clustering techniques are subsequently employed to this set to simultaneously estimate the number and locations of change-points. Several simulation studies are conducted to evaluate the effectiveness of the proposed method. An empirical analysis based on stock market indices， government bond yields， and foreign exchange markets from eight countries over the period February 7， 2006， to December 30， 2024， is further presented to illustrate its practical performance. The numerical results demonstrate that the proposed method achieves improved detection accuracy with lower computational cost compared to existing approaches. The empirical findings also align with major global financial and economic events occurring during the study period.
