Reproduce “Robust Yield Curve Estimation for Mortgage Bonds Using Neural Networks” on OpenResearch

About this paper

Sina Molavipour, Alireza M. Javid, Cassie Ye, Björn Löfdahl, Mikhail Nechaev

Researchers from SEB Group developed a neural network framework to robustly and stably estimate yield curves, particularly for small, illiquid markets like Swedish mortgage bonds. Their model uses a custom loss function that balances market accuracy with desired curve smoothness and alignment to economic benchmarks, showing improved stability and robustness against data noise compared to traditional methods.

Key insights

• Even simple neural network architectures can effectively model complex financial curves and enhance robustness in data-scarce environments when guided by a thoughtfully designed, domain-specific loss function.
• Explicitly incorporating financial domain knowledge, such as smoothness constraints and alignment with economic benchmarks, into the model's objective function is crucial for generating stable and economically plausible yield curves for practical applications.
• There is an inherent trade-off between achieving precise market pricing accuracy for individual bonds and ensuring desired qualitative properties like overall curve smoothness and stability; the proposed framework allows practitioners to manage this balance.

Summary by alphaXiv.