| 摘要: | 本論文專注於利用長短期記憶網絡（LSTM）來預測S&P 500和Yahoo 100成分股的股價波動。傳統的股價預測方法主要依賴於單一公司的日常股價資訊，如開盤價、最高價、最低價、收盤價和成交量。與之不同的是，我們的研究擴展了資料範圍，包含了S&P 500和Yahoo 100的聯集，共508家公司的資料，從而提供了更全面的市場視角。
除了標準的每日股價資訊，本論文還納入了公司基本資訊（例如行業類別、全職員工數、公司成立年限、除息日和市值）以及當年的宏觀經濟資料（如基於GDP的衰退指標指數、GDPNow、實際個人消費支出和實際私人國內固定投資的即時預測）。此外，我們還融合了多種技術性指標，包括移動平均線（MA）、日收益率、價格波動、相對強弱指數（RSV）、K線、D線和J線，以增強預測的精確度。
研究結果顯示，相比僅依賴每日股價資訊（如開盤價、最高價、最低價、收盤價和成交量）的方法，結合多維度資料的方法在預測股價漲跌方面展現出更高的精確度。這一發現凸顯了在股價預測中考慮公司基本資訊、宏觀經濟資料和技術性指標的重要性。在不同時序切割方案的比較中，以前30日預測後5日的模型表現最佳，這可能是因為該時序範圍提供了足夠的資料來捕捉市場動態，同時避免了短期波動的隨機性和長期預測的不確定性。;This paper focuses on utilizing LSTM networks to predict stock price volatility of S&P 500 and Yahoo 100 component stocks. Traditional stock price prediction methods primarily rely on daily price data for individual companies, such as opening price, high price, low price, closing price, and volume. In contrast, our study expands the scope of data to encompass the union of S&P 500 and Yahoo 100, totaling data for 508 companies, thereby providing a more comprehensive market perspective.
In addition to standard daily price data, this paper incorporates fundamental company information (e.g., industry category, full-time employees, year of establishment, ex-dividend date, and market capitalization) as well as macroeconomic data for the year (such as GDP-based recession indicator index, GDPNow, actual personal consumption expenditures, and actual private domestic fixed investment real-time forecasts). Furthermore, we integrate various technical indicators, including MA, daily returns, price volatility, RSI, K-line, D-line, and J-line, to enhance prediction accuracy.
The research results indicate that the multidimensional approach, in contrast to methods relying solely on daily price data (such as opening price, high price, low price, closing price, and volume), demonstrates higher accuracy in predicting stock price movements. This finding underscores the importance of considering fundamental company information, macroeconomic data, and technical indicators in stock price prediction. In comparisons among different time-series segmentation schemes, the model that predicts the next 5 days based on the previous 30 days performs the best. This may be attributed to this time frame providing sufficient data to capture market dynamics while avoiding the randomness of short-term fluctuations and the uncertainty of long-term predictions. |
