利用 Radius Neighbors Regressor 模型預測台灣股市加權指數並賦予強弱指標

Abstract

股票投資是現代人在累積資產上不可或缺的工具，雖然投資理財有賺有賠，但是若能夠找到一套良好的交易策略，以及善用各種分析工具，達到長期穩定獲利也是一件可以期盼的事情。本論文使用Radius Neighbor Regressor之機器學習方法並結合個人之股票交易經驗，在特定時間窗口之大盤強弱指標以及大盤中長期的多空頭判斷上取得了良好的結果。在實作上，我們使用Radius Neighbor Regressor與個人交易經驗所挑選出的特徵值作為產生強弱指標的依據。資料收集樣本時間為2013/1/25~2023/12/21，總共2668個交易日。主要資料來源取自XQ全球贏家之資料庫，並且使用所經加權後的強弱指標分別在幾種預測時間的長短進行比較與分析。從實驗結果驗證，我們發現使用Radius Neighbor Regressor搭配個人交易經驗所挑選出的特徵值，在以60個交易日預測後20個交易日的結果準確率高達73%，且在傳統多空頭的分析上也得到了良好的結果。另外，還證明了在特徵值選擇上以個人交易經驗做選擇的優勢，最後也彌補了單純使用Radius Neighbor Regressor機器學習方法的缺點，得出最佳的一種大盤強弱指標之模型。

Stock investment is an indispensable tool for modern people to accumulate wealth. Although investment and financial management come with risks, finding a good trading strategy and utilizing various analytical tools can lead to long-term stable profits. This thesis uses the Radius Neighbor Regressor machine learning method combined with personal stock trading experience to achieve good results in determining the strength and weakness indicators of the market and the long-term bullish or bearish trends within a specific time window.In practice, we use the Radius Neighbor Regressor and the features selected based on personal trading experience to generate strength and weakness indicators. The data collection period spans from January 25, 2013 to December 21, 2023, covering a total of 2668 trading days, with the main data source being the XQ Global Winner database. We then use the weighted strength and weakness indicators to conduct comparisons and analyses over various prediction periods.The experimental results confirm that using the Radius Neighbor Regressor combined with the features selected based on personal trading experience achieves an accuracy rate of up to 73% when predicting the results for the 20 trading days following a 60-day prediction period. This approach also yielded good results in traditional bullish and bearish analyses, demonstrating the advantage of selecting features based on personal trading experience. Ultimately, this method addresses the shortcomings of solely using the Radius Neighbor Regressor machine learning method, resulting in the optimal model for market strength and weakness indicators.