以ETF建構指數追蹤組合之套利策略研究：動態模型與機器學習之應用

Abstract

傳統指數套利策略通常透過買賣一籃子現貨成分股來複製現貨部位，然而此方法往往伴隨較高的資金需求、交易成本與執行複雜度。為提升策略之實務可行性，本研究提出以一籃子指數股票型基金（Exchange-Traded Funds, ETFs）取代現貨成分股，作為臺灣指數套利策略中的現貨替代部位。

本研究採用滾動迴歸模型估計 ETF 動態投資組合權重，實證結果顯示，ETF 投資組合能有效追蹤目標指數，同時降低實際執行上的複雜性。在此架構下，進一步建構日內指數套利策略，並結合機器學習模型，針對期現價差偏離所產生之交易訊號進行篩選。

除傳統價差相關變數外，本研究亦納入財務報表特徵與三大法人交易相關變數作為解釋變數。實證結果顯示，機器學習模型能有效捕捉上述額外資訊來源，進而提升交易訊號之篩選效率。此外，不同盤中時段之穩健性檢定結果顯示，套利策略之有效性具有明顯的時間異質性，模型績效亦會隨市場狀態變化而有所差異。

整體而言，本研究結果顯示，以 ETF 作為現貨替代工具，並結合機器學習訊號篩選機制，能有效提升指數套利策略之實務可行性與交易績效。

Traditional index arbitrage strategies typically replicate the spot position by trading a basket of constituent stocks, which involves high capital requirements, transaction costs, and execution complexity. To improve practical feasibility, this study proposes using a portfolio of exchange-traded funds (ETFs) as a substitute for the spot position in Taiwan index arbitrage.

Dynamic ETF portfolio weights are estimated using rolling regression models, and the results show that ETF portfolios can effectively track the benchmark index while reducing implementation complexity. Based on this framework, the study develops an intraday index arbitrage strategy and incorporates machine learning models to filter candidate trading signals derived from futures–spot spread deviations.

In addition to traditional spread-related variables, this study incorporates financial statement characteristics and institutional trading variables as explanatory features. The empirical results show that machine learning models are able to capture these additional sources of information and improve trade selection efficiency. Finally, robustness checks across different intraday periods confirm that the effectiveness of trading strategies varies over time, and that model performance is closely linked to changing market conditions.

Overall, the results suggest that ETF-based spot substitution combined with machine learning signal filtering provides a practical and effective enhancement to index arbitrage.