Authors: Assistant Professor Dr. Ranjeet Kumar Ambast, Aditya Vikram
Abstract: The foreign exchange (Forex) market is characterised by high liquidity, pronounced volatility, and non-linear dynamics. This makes accurate prediction of currency fluctuation and investment risk exceptionally difficult for traditional econometric models (Bollerslev, 1986; Cont, 2001). This paper synthesises secondary evidence on the comparative predictive performance of two Artificial Intelligence (AI) approaches Long Short-Term Memory (LSTM) deep learning networks (Hochreiter & Schmidhuber, 1997) and Random Forest (RF) ensemble learning (Breiman, 2001) against the traditional Generalised Autoregressive Conditional Heteroskedasticity (GARCH) family of models (Bollerslev, 1986), drawing on published studies that use historical USD, EUR, GBP, JPY, BRL, and ZAR exchange-rate data spanning approximately 2012–2025. The review evaluates model performance across stable and volatile market regimes for directional accuracy through Value-at-Risk (VaR) estimation, and risk-adjusted investment signals. Findings from the reviewed literature indicate that AI-based LSTM models achieve superior performance for short-horizon volatility forecasts. It’s particularly in capturing sudden shifts in implied volatility (Kraus & Feuerriegel, 2024). Random Forest models tend to deliver the highest directional accuracy and the lowest point-prediction error across several currency pairs and cryptocurrency markets (Milionis & Konstantinou, 2024; Ndlovu, 2025). Hybrid GARCH-LSTM architectures further improve predictive accuracy, with an APARCH-LSTM specification reported to achieve a coefficient of determination (R²) of 95.53% for USD/BRL volatility forecasting (Hottz, 2025), while hybrid models improve Value-at-Risk estimation accuracy by up to 10% during periods of elevated volatility relative to standalone GARCH or LSTM specifications (Nsengiyumva et al., 2025). The synthesis further shows that risk-adjusted performance measured through the Sharpe ratio, Calmar ratio, and maximum drawdown favours hybrid deep-learning architectures for multi-asset portfolios and tree-based ensembles such as XGBoost for equity-index applications (Saly-Kaufmann et al., 2026; Singh & Praveen, 2025). The paper concludes that no single AI technique dominates across all currency pairs, forecast horizons, and objectives, and that model choice should be aligned with the specific trading, risk-management, or regulatory objective at hand.
