Feature Expansion Effect Approach for Improving Stock Price Prediction Performance

The purpose of paper is to synthesize various techniques used in stock price prediction research and to present an appropriate methodology. For stock price prediction, a comprehensive stock price analysis server was built in connection with the Yahoo Finance API to predict stock prices of Korea's KOSPI Index and the U.S. S&P 500 Index the next day, 5 days, and 10 days later. To find the optimal modeling method, PCA, CNN-LSTM, and LightGBM algorithms were applied, and a multi-view sliding window technique, technical indicators processing stock prices, and macroeconomic indicators were added for a comprehensive analysis. CNN-LSTM showed higher performance than LightGBM, and when only basic indicators were used, the performance of PCA deteriorated, when technical indicators were combined, performance improved, and when macroeconomic indicators were also combined, performance improved significantly. Depending on the use of macroeconomic indicators, KOSPI Index forecasts showed high performance in stock price forecasts 5 to 10 days later, excluding next-day forecasts, but S&P 500 Index showed low performance in all. The optimal window size of the multi-view sliding window was different depending on the objective variable, algorithm, etc. In conclusion, stock indices in Korea and the United States showed differences in performance improvement depending on country characteristics. These results are expected to be used not only for companies that expect high investment returns through accurate stock price predictions, but also in various fields that attempt predictions using deep learning models.