本文以即時預測股市狀態為研究主題，透過區分全樣本認定與遞迴認定兩種認定流程來探討在不同前提假設下對於模型的影響，本研究發現在使用非參數認定法進行狀態認定時，若採用遞迴認定，則PS法則會發生認定翻轉問題與認定延遲問題，實證結果則顯示PS法則在改善認定翻轉問題前難以應用於即時預測領域，而由於LT法則只有認定延遲問題，因此可以藉由改善遞迴預測流程來達到初步解決認定延遲問題。除此之外，為解決高維資料下的過度適配與多重共線性問題，本文在變數篩選方法採用階層式分群法(Hierarchical Clustering)與LASSO迴歸來篩選重要變數，研究結果顯示此方法的變數篩選能力十分優秀。同時本文也測試不同模型組合下的預測能力表現，包含動態結構、相依性結構與股市狀態結構，研究結果發現在三種結構同時存在的模型有最佳的預測能力。;This paper takes real-time prediction of stock market status as the research topic, discussing the impact on the model under different assumptions by distinguishing between the two identification processes of full sample identification and recursive identification. We find that under the recursive identification using the PS rule, the problem of identification inversion and identification delay will occur, while under the LT rule, there is only the problem of identification delay under the recursive identification. The empirical results show that the PS rule is difficult to apply to the real-time prediction field before solving the identification inversion problem, and the LT rule can solve the identification delay problem by improving the recursive prediction process. In order to solve the problem of overfitting and multicollinearity under High-Dimensional data, we use Hierarchical Clustering and LASSO regression to select key variables. The results also show that this method performs very well. In addition, we also test the performance of forecasting ability under different model settings, including dynamic structure, dependency structure and stock market state structure. The result shows that the model with all structures has the best forecasting ability.
