Forecasting Market Regimes in an Emerging European Market: A Machine Learning Framework for the Romanian BET Index

Financial markets do not evolve in a stable or uniform way. They often move through different phases, shaped by changes in risk, returns, volatility, and investor behaviour. These shifts are especially important in emerging markets, where lower liquidity, exposure to global shocks, and domestic vulnerabilities can make transitions from calm periods to stress episodes faster and more severe. This paper examines whether such regimes can be identified and forecast in the Romanian equity market, using the BET index as the main reference. The analysis covers the period from November 2017 to November 2025. In the first stage, a three-state Gaussian Hidden Markov Model is used to identify latent market regimes. The estimated states are then interpreted as calm, stress, and panic regimes based on their conditional realised volatility. In the second stage, the regimes are forecast one day ahead using an XGBoost classifier trained on macro-financial variables, including global risk indicators, cross-asset returns, sovereign yield curve measures, commodity prices, and domestic macroeconomic indicators. The results show strong overall predictive performance. Calm periods are identified with high reliability, while stress regimes are detected with meaningful sensitivity. Panic regimes, however, remain harder to classify because they occur less frequently, which highlights the difficulty of forecasting extreme market conditions in emerging economies.