Predicting S&P 500 index movements using advanced neural network architecture

This thesis develops a hybrid deep learning framework for the short-term directional forecasting of the S&P 500 Index. The methodology integrates advanced econometric techniques to address non-linearity and conditional volatility in financial time series. Specifically, a Generalized Autoregressive Conditional Heteroskedasticity (GARCH(1,1)) model captures and forecasts market volatility. Feature selection is rigorously performed via LASSO Cross Validation (LASSO-CV), yielding an optimal and sparse set of 8 indicators. The core predictive mechanism employs an optimized Bidirectional Long Short-Term Memory (Bi-LSTM) network. Empirical results demonstrate significant predictive capability, showing an average Directional Accuracy (DA) of 57.06%. Furthermore, a comprehensive one-step-ahead rolling backtest simulation validates the model's generalization capacity. The findings confirm that combining econometric rigor with Bi-LSTM architectures offers a statistically robust methodology for generating actionable short-term market forecasts.