A comparative study of LSTM and GARCH in predicting stock prices volatility and estimating value-at-risk

Predicting stock prices is a coveted skill among investors due to its potential for profitability. In this context, stock volatility emerges as a critical predictor of future prices. This study utilizes two forecasting approaches: the Long Short-Term Memory (LSTM) neural network model and the GARCH (Generalized Autoregressive Conditional Heteroskedasticity) econometric model to identify which approach is more suitable for predicting stock price volatility. These models are applied to forecast the volatility of Indonesian stocks across various industries, with the goal of determining which model yields more accurate predictions. To assess the models’ accuracy, performance metrics including Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE) are employed. Additionally, this research leverages the calculated volatility estimates to determine the Value-at-Risk (VaR) for these stocks, providing a comprehensive risk assessment tool. This dual application of forecasting and risk analysis aims to help investors in deciding between neural network or econometric model for estimating risk that comes with stock invesment. This study found that LSTM produces a more accurate rolling volatility prediction compared to GARCH, and that a lower Value-at-Risk threshold from GARCH does not necessarily mean that it is the better model as it might overestimate the risks involved in stock invesment.