本論文將探討長短期記憶模型(Long Short-Term Memory)對股票市場每日收盤價預測之成效，研究目標對象為國內成分證券ETF中的數支股票。研究方法為將time step設定為一天，以每日的股市資訊為特徵值做預測，較接近現實的股票市場狀況。本研究選擇以元大台灣50(0050)、元大中型100(0051)、元大高股息(0056)及富邦台50(006208)四支熱門的ETF股票為分析對象。對於特徵點的選擇則是以技術面和籌碼面資料處理後篩選輸入。本論文會先設定超參數的維數及範圍，反覆測試並調整，以增進LSTM模型對個股之預測能力。為能呈現更理想的結果，後又在模型建置上結合超參數優化的功能。相較於傳統的網格搜索 (Grid Search)的方式，貝氏最佳化(Bayesian Optimization)運算效益更高。最後觀察到LSTM模型與貝氏參數最佳化，對於ETF股票收盤價之預測表現皆有一定的成效。;This Paper is discussing about the result that trying Long Short-Term Memory to predict closing prices of stocks.The research objects we chose ETFs tracking Taiwan stocks.Research method we set time step as one day. And we record market information as features to predict, make prediction result closer to reality market.In this research we chose four popular ETFs as research objects, including 0050, 0051, 0056, 006208. The choices of feature we picked some data of technical analysis and chip analysis, and input to model training after data pre-processing.In this research we will set dimension and space of hyper-parameters, repeat testing and adjustment to improve prediction ability on the stock.For getting better result, we combined hyperparameters tuning function on model structure. Bayesian Optimization have higher performance rather than traditional Grid Search. Finally, we can learn that LSTM and Bayesian Optimization will boost prediction result of ETFs closing prices.
