A Comparison of Deep Learning and Econometric Models for Forecasting Downside Risk in the South African Gold Market

Purpose: The purpose of this study is to compare the performance of econometrics and deep-learning-based methodologies for forecasting downside risk in the South African gold market across different forecasting time horizons.
Design/Methodology/Approach: We compare an existing Markov-Switching Generalised Autoregressive Conditional Heteroscedasticity (MS-GARCH) model, a Deep Neural Network (DNN), and a Long Short-Term Memory (LSTM) neural network, all using daily historical data for the South African Rand Gold Price from January 1, 2012, to July 31, 2025. We use Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Forecast Error (MFE) to assess the forecasting performance of our models. To further assess each model's ability to protect against downside risk, we then employ Maximum Drawdown, Sortino Ratio, and Marginal Expected Shortfall, all evaluated within the context of a Rolling Window forecasting.
Findings: Our findings indicate a trade-off between a model's ability to provide accurate predictions and its ability to mitigate potential losses. Specifically, the LSTM produced more accurate short-term predictions. However, due to temporal complexity, the LSTM exhibits larger maximum drawdowns and generally poorer risk-adjusted return performance over longer forecasting horizons. Conversely, the MS-GARCH model demonstrated much greater reliability in terms of preserving wealth and protecting against downside risk, especially over longer forecasting horizons. Additionally, the DNN performed poorly relative to the other two models across.
Originality: This study is unique in providing a detailed comparative analysis of econometric methodologies and deep learning-based methodologies for predicting the losses in emerging commodity markets.