Enhancing Regime Shift Detection Using Unstructured Data: A Study on the Treasury Market

Battle Count: 72

Highly relevant for quantitative trading in fixed-income and macro markets. Accurate regime shift detection is foundational for: (1) calibrating yield-curve models within stable regimes, (2) identifying when stock-bond correlations flip, (3) detecting when flight-to-quality channels break down, (4) timing duration bets around policy transitions, (5) adjusting risk models when term premium dynamics change. The same-day detection latency and detector-agnostic design make it practical for production systems. However, it is primarily a detection/classification tool rather than a direct trading signal generator.

Moderate complexity. The pipeline has four stages with clear interfaces. Stage A and C require LLM API calls with specific prompts. Stage B requires implementing a bootstrap likelihood-ratio VAR(1) test (non-trivial due to over-parameterisation at N=14). Stage D is simple clustering/deduplication. The detector-agnostic design means integrating any change-point detector is straightforward (just produce a candidate set). Main challenges: prompt engineering for LLM stages, proper bootstrap calibration, and handling the 14-variable panel preprocessing (z-scoring, multicollinearity reduction).

Code is available at https://github.com/mingxuan-yi/regime_shift. All hyperparameters are specified (W=90d, B=500 bootstrap reps, alpha=0.05, theta_A=0.6, theta_C=0.8, Delta=14 days). The 26-event anchor list construction is described. However, the LLM (Claude Sonnet 4.6) is a proprietary API model, and the exact prompt templates are only partially shown (condensed). The FOMC corpus is publicly available from federalreserve.gov. The 14-variable panel construction is described but exact data sources for all variables would need verification.