Forecasting Market Regimes for Dynamic Asset Allocation: A Hybrid Approach Using Continuous Jump Models and Extreme Gradient Boosting

Abstract

This thesis examines whether a hybrid, two-step framework, combining asset-specific regime identification via a Continuous Jump Model with one-step-ahead regime forecasting using Extreme Gradient Boosting, can enhance dynamic asset allocation relative to a baseline strategy that does not incorporate regime information. Motivated by the limitations of classical regime-switching models such as Hidden Markov Models, the Continuous Jump Model provides persistent and interpretable regime estimates that avoid excessive switching and reflect heterogeneous market behavior. These unsupervised regime signals form the basis for supervised forecasting using Extreme Gradient Boosting, trained on engineered features including macro-financial indicators. The resulting regime probabilities are directly incorporated into mean-variance and minimum-variance portfolio optimization. Using daily data from twelve diversified asset classes over a 20-year period (2005–2024), we evaluate the performance of regime-aware strategies in terms of Sharpe ratio, maximum drawdown, volatility, and turnover. At both the asset and portfolio level, the CJM-XGB framework leads to economically and statistically significant improvements. In the mean-variance setting, the Sharpe ratio rises from 0.22 to 1.36, with the gain being statistically significant at the 5% level. We further introduce a regime-weighted covariance estimator to reflect conditional volatility dynamics, and show that it contributes to enhanced portfolio stability, though it remains heuristic. Robustness checks confirm that the strategy retains much of its performance even when signal execution is delayed. Altogether, this thesis provides empirical evidence that regime-aware allocation, built on probabilistic foundations and modern forecasting tools, can improve investment performance. The findings contribute to the literature on Regime-Based Asset Allocation presenting a practical and interpretable framework that integrates statistical modeling with machine learning to adapt to non-stationary financial environments.