Arbitrage-Free Bond and Yield Curve Forecasting with Neural Filters under HJM Constraints

Battle Count: 71

Highly relevant for fixed-income quantitative trading. The framework directly addresses bond pricing and yield curve forecasting with arbitrage consistency, which is critical for relative-value trading, curve trades, and risk management in fixed income. The bid-ask hit rate metric is directly actionable for execution. The 5-day horizon improvements at short maturities are particularly relevant for short-term trading strategies. However, the paper focuses on forecasting accuracy rather than trading signal generation or portfolio construction, and the computational cost of PF may limit real-time applications.

Very high complexity. Requires implementing three different filter types (KF, EKF, PF) as differentiable components within a deep learning framework. The architecture combines CLSTM for cross-sectional compression, LSTM for temporal dynamics, dense layers for parameter output, residual LSTM for noise modeling, and differentiable filter blocks. Particle filter with EKF-assisted importance sampling and systematic resampling adds significant complexity. Multivariate generalized Gaussian noise modeling, AER penalty computation over tenor grids, and end-to-end backpropagation through filter blocks require careful numerical implementation. Runtime varies dramatically: KF in minutes, EKF in seconds per epoch, PF in hours with 300 particles.

The paper provides detailed mathematical formulations, algorithm descriptions (including PF in Appendix A), LSTM/CLSTM architecture details (Appendices B and C), and mentions TensorFlow implementation. However, no GitHub repository is provided, proprietary Treasury data is used alongside FINRA-TRACE, and hyperparameter settings are partially described. The AER penalty formulation and filter equations are fully specified, enabling theoretical reproduction but practical reproduction requires access to the same data and implementation details.