Downside Risk of the South African Mining Index: Leveraging the Power of Long Short-Term Memory and Explainable AI

Abstract

With a primary focus on downside risk (i.e., losses), this study uses a combination of deep learning and probability distributions to the South African mining index. We model the downside risk by combining probability _distribution, SHAP, and long-short-term memory. In contrast, Expected shortfall (ES) and value at risk (VaR) are used in the downside risk assessment. To illustrate the varied levels of model performance across various distributions (normal, student t, skew-normal, generalised hyperbolic, and Laplace), these risk metrics are backtested at 90%, 95%, and 99% confidence intervals. The findings show that normalcy assumptions are inadequate for accounting for high losses. In terms of forecasting power, the generalised hyperbolic distribution performs better than all the other distributions, especially over longer forecasting horizons and greater confidence levels. Findings from this study further show that the Russian-Ukraine war, ZAR/USD exchange rate and COVID-19 have the highest impact on the downside risk.