本文使用三狀態馬可夫轉換模型探討臺灣股價指數日報酬的可測性與隨機性。實證結果現臺灣股價指數日報酬並不具有隨機漫步行為的現象，且不能拒絕股價指數報酬存在多頭、盤整與空頭等三個狀態，顯示馬可夫轉換模型存在之必要性。在預測績效方面，馬可夫轉換模型在樣本內預測時明顯優於其他模型，但在樣本外預測方面，馬可夫轉換模型則無明顯優於其他線性模型，僅優於隨機漫步模型，此表示馬可夫轉換模型仍得以解釋市場上的部份資訊。

This paper investigates the forecasting performance and the behavior of random walk to stock index returns using three-states Markov switching model. The empirical results find out the stock returns do not follow the random walk process, and the hypothesis of three states-bull, flat and bear market can not be rejected. It indicates that stock index returns are better specified by Markov switching model. As to the forecasting performance in sample, we find that the Markov switching model is better than alternative models. However, in out-of-sample forecast, it does not perform better than other linear models, but outperform than random walk model. It represents that part of information can be caught by Markov switching model.