Financial markets rarely behave uniformly over time, instead exhibiting periods of different market conditions. These structural shifts pose significant challenges for traditional forecasting models, particularly due to heteroskedasticity and non-linearity. While market regime frameworks seek to address this complexity, many existing approaches are constrained by univariate inputs, rigid transitions, or lack predictive capability. This paper proposes a two-phase framework that overcomes these limitations by integrating unsupervised and supervised machine learning techniques. In the first phase, a Gaussian Mixture Models (GMM) is used to detect latent market regimes from a rich set of macro-financial indicators, capturing overlapping probabilistic states with interpretable economic characteristics. We identify six distinct regimes—including high-inflation, crisis, and expansionary environments—each associated with unique risk-return profiles. In the second phase, we train an XGBoost-based Gradient Boosting Machine (GBM) classifier on historical regime labels to forecast future regime states. The predictive model achieves high out-of-sample classification accuracy (92.2%) and outperforms the market in a simple long-short trading strategy based on anticipated regime shifts.