時序資料分析在金融、製程監控與氣候預測等領域中越顯重要，然而，現實中的資料常出現分佈變動，導致傳統模型難以穩定應用，影響估計與選擇的準確性。特別在「時序分佈轉移」（temporal distribution shift）情境下，資料結構變化不可預測，對模型的適應能力形成挑戰。

針對此一挑戰，本文提出一套具備自適應能力的模型選擇與預測方法 —— 自適應資料驅動視窗（Adaptive Detection and Rolling Window, ADRW），該方法可根據資料於不同時間點的統計特性，自動調整模型建構範圍，並無需預先訓練候選模型，即可即時進行估計與預測。相較於現有的自適應滾動視窗法（ARW），ADRW 強調資料驅動的靈活性與方法簡便性，適用於未知或快速變動的環境中。

此外，ADRW 方法亦可應用於時間序列的預測問題。模擬結果顯示，相較於傳統自迴歸模型（AR model），ADRW 在不同情境下均具備更佳的預測表現，展現出其廣泛的適用性與優越性。

理論分析與模擬實驗均顯示，本方法能有效對應資料非平穩性所帶來的估計誤差，並在時序分佈轉移情境下，展現良好的穩健性與預測效能，為處理變動環境中的模型選擇與預測提供一套具體且可行的解決方案。

Time series data often experience temporal distribution shifts, making traditional models unstable and reducing estimation and selection accuracy. To address this, we propose the Adaptive Detection and Rolling Window (ADRW) method, which dynamically adjusts model construction based on data characteristics without requiring prior model training. ADRW enables real-time estimation and prediction, offering greater flexibility and simplicity compared to existing Adaptive Rolling Window (ARW) approaches.

ADRW combines change-point detection with adaptive model selection, allowing it to respond effectively to both abrupt and gradual distributional changes. Furthermore, the method can be extended to time series forecasting tasks, offering consistent performance without the need for pre-specified models.

Simulation results show that ADRW outperforms ARW and traditional Autoregressive (AR) models, particularly under non-stationary conditions. The method demonstrates robust estimation and forecasting performance across varying data scenarios, providing a practical solution for dynamic environments.