Machine learning models for financial market volatility prediction: comparative economic forecasting accuracy

The ability to forecast volatility precisely plays an important role in risk management in financial markets. In this paper, we performed a comparative analysis of forecasting accuracy of machine learning approaches and the traditional GARCH model in the case of the Amman Stock Exchange in Jordan. Historical daily index data from 2010 to March 2025 were used to compare the forecasting performance of the LSTM algorithm, support vector machine algorithm, and random forest algorithm for one-day out-of-sample horizon (January-March 2025) to GARCH(1,1)-t model. According to the MSE and MAE measures, the LSTM model demonstrated the lowest forecasting errors and performed significantly better by the Diebold-Mariano test. In economic terms, this finding implies that more accurate volatility forecasts directly translate into better risk-adjusted portfolio performance and more efficient capital allocation. The LSTM model provided up to 30% MSE reduction compared to GARCH during the periods of high volatility regimes. As such, from the economic analysis, the LSTM forecast portfolio is observed to have a Sharpe ratio of 0.89 with a maximum drawdown of -4.15%, which is a better performance compared to the GARCH forecast portfolio, proving the economic benefit of using machine learning models for forecasting volatility in emerging markets. In summary, it can be seen that the use of machine learning, especially the LSTM method, does not only improve forecast accuracy but also has strong economic benefits. As such, the problem of choosing the best volatility forecasting model is well solved in the Jordanian financial market.

Hali tarjima qilinmagan