Term Structure Modelling and Forecasting with Neural Structured State-Space Models

Term Structure Modelling and Forecasting with Neural Structured State-Space Models

• Autor
• Pedro Amaral Amorim Oliveira
• Co-autores
• Marcio Poletti Laurini
• Resumo

• This paper proposes using Mamba-based neural Structured State-Space Models (SSMs) as a flexible alternative to traditional term structure models for yield curve forecasting. Unlike parametric approaches with fixed factor loadings and linear Gaussian dynamics, the SSM learns latent states and cross-maturity interactions directly from data, adapting to nonlinearities and regime dependence. Economic structure is incorporated through a B-spline layer that enforces smoothness and no-arbitrage restrictions embedding financial theory into the learning objective in a manner analogous to physics-informed neural networks. The framework is evaluated via Monte Carlo simulations across multiple structural DGPs and empirically on U.S. Treasury Constant Maturity yields from 2019 to 2025. Results show lower out-of-sample forecast errors relative to the Diebold-Li benchmark, with statistically significant gains particularly pronounced when underlying dynamics deviate from linear, time-invariant factor structures.

• Palavras-chave
• Term Structure, Forecasting, Structured State Space Models, Deep Learning, No-Arbitrage, Yield Curve
• Modalidade
• Comunicação oral
• Área Temática
• Econometria Financeira

Voltar Download