大多數人在進行股票交易的時候,經常會運用一些指標做為決策參考,而且習慣觀察相同的類型,再根據當中的變化決定買進或賣出;但有些時候會發生判斷失靈的狀況,也就是說某些時候在系統性的因素外,指標與股價走勢會突然偏離正常的理解,這些背離的情況造成了預測失準及投資損失。本研究認為股價在各波段中應該隱藏著不同的關鍵指標,相同的指標在不同時間具有不同程度的權重,這樣的關係主導著股價變化,所以若能各波段中找出代表的關鍵指標,再利用動態的方式進行訓練或許能讓模型更接近真實。因此本研究提出「動態特徵選擇與學習 (Dynamic Feature Selection and Learning, DFSL)」方法。利用動態的方式找出各波段的關鍵指標,再動態的對各波段進行訓練。根據本研究實驗結果得到結果:DFSL 在預測未來股價漲跌的正確率優於靜態特徵選擇與學習方式。;Most people often use some metrics as a reference when making stock transactions, and they are accustomed to observing the same type, and then decide to buy or sell according to the changes in it. But sometimes there will be a situation of judgment failure, that is to say, in some cases, outside of non-systematic factors, the trend of metrics and stock prices will suddenly deviate from normal understanding. These deviations have caused misprediction and investment losses. This study believes that different key metrics should be hidden in the stock price in each band. The same indicator has different degrees of weight and meaning at different times. This relationship dominates the change in the stock price. So if we can find the representative key metrics in each band, and then use dynamic training to make the model closer to the real. Therefore, this paper proposes a method of “Dynamic Feature Selection and Learning (DFSL)”. Use a dynamic method to find the key metrics of each band, and then dynamically train each band. According to the experimental results of this research, the result is that DFSL is better than static feature selection and learning methods in predicting future stock price changes.