科技金融在人工智慧應用中是一個熱門的主題，過去雖已有許多研究 使用股票歷史數據及技術指標進行個股漲或跌的預測，但是要如何將預 測結果結合投資組合的配置仍然是個問題。因此本研究中，我們將財經 新聞與技術指標納入為漲跌預測中，以獲得更好的預測效果。為了使動 態資組合優化，我們進一步預測股票漲跌幅排名，透過簡單的策略投資 排名前面的股票，使預測模型達到比台灣 50ETF 及中 100ETF 更高的 投資報酬率。本研究分成隔日股票漲跌的二元分類、隔週/隔月的股票漲 跌幅排名的預測、以及投資組合策略應用回測三個部份。 首先應在股票趨勢的預測中，以股票及期貨市場上常用的技術指標 作為模型的輸入特徵，比較有無加入新聞標題及新聞情緒對於隔日股票 漲跌預測結果的影響。實驗結果顯示，技術指標加入新聞情緒之預測 效果較加入新聞標題來的好，其效果提升約 4%;與基準方法 Random Rorest 相比，其效果提升約 5%。其次我們重新定義問題改以預測隔 週/隔月的股票漲跌幅排名，以減少頻繁交易，在使用技術指標基準下， 比較有無加入新聞情緒時，對於預測股票漲跌幅排名的影響。實驗結果 顯示，在加入新聞情緒時，其損失函數的損失 (loss) 較沒加入新聞情緒 時低。最後我們選擇股票排名預測排名前 K 名的股票形成動態資組合配 置。實驗結果顯使用投資組合配置策略的投資報酬率普遍比台灣 50ETF 好;尤其是模型在加入新聞情緒時，投資報酬率平均可提高約 23%。;FinTech is a popular application in artificial intelligence applications. Although many studies have been conducted to predict the movement of stock price based on historical data and technical indicators, how to com- bine the prediction with portfolio allocation remains a problem. In this re- search, we consider financial news and technical indexes in the stock price movement prediction to achieve better performance. To support adaptive portfolio optimization, this study further proposes a new task to predict the ranking of stocks return. With a simple strategy of investing top ranked stocks, the prediction model achieves a high return on investment (ROI) than Taiwan 50 ETF and Taiwan mid-cap 100 ETF. The thesis is divided into three parts: binary classification of stock movement on the next trad- ing day, the ranking prediction of stock returns in the next week/month and an application of portfolios trading strategy based on backtesting. For stock trend prediction, we use stock technical indicators as the model’s input features, and compare the performance with or without news headlines or news sentiment on the next trading day stock movement pre- diction. Experiments show that adding news sentiment has better per- formance than adding news headline about 4%. Comparing it with the baseline (Random Forest), performance has been improved around 5%. Secondly, we formulate a new problem to predict stock rank for the next week/month based on return of investment to reduce frequent trading. Base on technical indicators, we compare the performance of stock rank prediction with or without news headlines and news sentiment. While the model with news headlines achieves lower loss, the model with news sen- timent has better balance in stock rank prediction. Finally, we use the next week/month stock rank prediction to select the top ranked k stocks for adaptive portfolio allocation. Experiments show that the portfolio al- location strategy achieve higher investment return than Taiwan 50ETF, especially when using the model with news sentiment, the average ROI performance can be improved around 23%.
