Regime-Aware Portfolio Management via Retrieval-Augmented LLM-Guided Expert Switching

Financial markets are highly non-stationary, and deep learning or reinforcement learning based portfolio strategies often fail under regime shifts. In this work, a novel retrieval-augmented, regime-aware framework that dynamically selects expert strategies based on similarity to historical market regimes is proposed. A dual-stream Transformer-based VAE encodes technical and market-wide features for regime retrieval. An LLM agent estimates uncertainty and guides expert selection. Portfolio allocation considers returns, risk, and uncertainty for interpretable, adaptive decisions. Experiments on cryptocurrency markets show a 1.19% ROI, 20.80 Sharpe ratio, and 0.47% maximum drawdown, outperforming all baselines. The results confirm that similarity-driven expert selection and uncertainty-aware reasoning improve robustness and returns in volatile markets.