Deep Learning-Based Trading Strategy for Index ETFs Using LSTM and GARCH Under Market Uncertainty

Abstract

Tumelo Ranoto, Faezeh Ranjbar and Joe Wayne Byers

This study deduces a trading system that combines econometric volatility modeling and deep learning for the purpose of enhancing index Exchange-Traded Fund (ETF) forecasts. It employs a series of Generalized Autoregressive Conditional Heteroskedasticity (GARCH)-family models: GARCH(1,1), Exponential GARCH (EGARCH), Glosten– Jagannathan– Runkle GARCH (GJR-GARCH) , Asymmetric Power ARCH (APARCH), and Integrated GARCH (IGARCH) to estimate volatility processes in SPDR S&P 500 ETF Trust (SPY) returns and model them. These estimates are transformed into normalized and volatility-adjusted returns and used to provide insights towards sharpening predictive signals. Hidden Markov Models (HMMs) analyze raw and adjusted returns with the aim of detecting changes in market behavior and discriminating between various volatility regimes. Also, technical analysis indicators such as the Relative Strength Index (RSI) and Moving Average Convergence Divergence (MACD) augment input features, which are indicative of the market. A Long Short-Term Memory (LSTM) network is trained on sequentially constructed inputs to forecast near-term price movements, classifying outcomes into three: down, neutral, or up. Class weighting is applied by the model to compensate for skewed distributions, with a stable accuracy prediction. Backtesting results highlight the advantage of merging volatility regimes, technical signals, and GARCH-derived features in favor of a rigorous, evidence-based approach to capturing market factors as well as risk factors.

PDF