Corporate Bond Yield Curve Modeling: A Rating-Based Regime-Switching Generalized CIR Approach

Battle Count: 78

The paper is primarily a pricing and risk decomposition framework rather than a trading strategy paper. However, it is highly relevant for fixed-income quantitative trading in China's bond market: (1) the regime-switching framework identifies persistent rate and credit environments useful for tactical duration and credit positioning; (2) the spread decomposition separates discounting from credit compensation, enabling cleaner relative-value trades; (3) the filtered regime probabilities provide real-time state estimates for risk management; (4) the model-implied yield curves can be used for curve-fitting and arbitrage detection. The practical relevance is strongest for institutional fixed-income desks and risk managers rather than high-frequency trading.

The implementation involves: (1) a 4-dimensional nonlinear state-space model with two separate Markov regime chains; (2) regime-switching Unscented Kalman Filter with sigma-point propagation through nonlinear pricing maps; (3) block-recursive two-step estimation with generated regressors; (4) matrix-exponential pricing for rating migration with spectral decomposition; (5) backward recursion on an observation grid for regime-conditional bond prices; (6) price-level mixing to avoid Jensen-type distortions; (7) numerical optimization of a high-dimensional parameter vector with robust standard errors. The closed-form GCIR affine transforms (Appendix A) and the RS-UKF algorithm (Appendix D) are well-documented, but the full pipeline requires significant numerical expertise and careful handling of admissibility constraints.

The paper provides detailed mathematical formulations, algorithmic descriptions (Appendix D for RS-UKF), and closed-form affine solutions (Appendix A). However, the primary data source (ChinaBond weekly zero-coupon curves) requires institutional access. The rating migration data comes from CSCI Pengyuan Credit Rating disclosures. No code repository is mentioned. The model involves complex nonlinear filtering with multiple parameter dimensions, making exact reproduction challenging without access to the same data pipeline.