UCLA

This paper conducts a comprehensive empirical study on technical indicators for stock marketprediction using 90+ years of S&P 500 data. The research evaluates 70 indicators across four categories using advanced machine learning models (LSTM, XGBoost, Random Forest, SVR, LR), identifying 35 key indicators through Principal Component Analysis. Using 60-day rolling window backtesting, LSTM models achieve superior performance with MAE of 0.014 and MAPE of 0.008, while HLC3, OBV, TEMA, and Bollinger Bands consistently outperform others. The study’s key innovation—dynamic parameter optimization using Bayesian techniques—enhances predictive accuracy by 15-25% over traditional static methods. This research establishes a replicable framework for indicator optimization, demonstrates machine learning superiority in financial forecasting, and provides practical guidance for improving algorithmic trading strategies.