A Hybrid FIGARCH–LSTM Early Warning System for Volatility Regime Transitions in a Frontier Market: Evidence from Kenya

Abstract

Frontier financial markets face a diagnostic gap in forecasting volatility: linear and single-regime GARCH fails to capture breaks, spillovers, and regime transitions. Despite the importance of these markets, there is a gap in the literature: lack of a Kenya-specific, regime-sensitive Early Warning System (EWS) that can integrate long-memory filtering of volatility with nonlinear classification models. Therefore, policymakers lack signals to anticipate systemic stress. This study constructed a multi-stage pipeline using 6703 daily observations of the NSE 20 Share Index, USD/KES exchange rate, and Brent spot prices (1997–2024). K-Means clustering, FIGARCH filtering, LSTM–XGBoost ensemble classification, a logistic threshold model, and VaR/ES back-testing were used to generate alarm signals and validate risk-management performance. Across model estimation and evaluation phases, time horizons were assessed, showing that one-day signals were reactive and ten-day forecasts diluted precision, while five-day predictions achieved the strongest balance. Across model iterations, accuracy improved from 0.86 in the initial FIGARCH–LSTM pipeline, to 0.92 in the unbalanced classification model, and 0.94 in the weighted baseline, culminating in 0.98 with the final hybrid LSTM–XGBoost ensemble. The ensemble model delivered robust detection across Calm (F1 = 0.99), Moderate (F1 = 0.82), and Stress (F1 = 0.69) regimes. Stress thresholds were developed to translate regime signals into actionable alarms. This study provides an empirical application of a hybrid framework, combining long memory modelling of volatility with the adaptability of deep learning models to deliver Basel-compliant tail-risk alarms and a state-dependent policy matrix for regulators.