Bayesian inference for long memory term structure models

In this study, we propose a novel adaptation of the Dynamic Nelson-Siegel term structure model, incorporating long memory properties to enhance its forecasting accuracy. Our approach involves modelling the evolution of latent factors using fractional Gaussian noise processes, approximated by a weighted sum of independent first-order autoregressive components. The resulting formulation allows for a Gaussian Markov Random Field representation, facilitating the application of computationally efficient Bayesian techniques through Integrated Nested Laplace Approximations. Extensive simulation and empirical analysis demonstrate that integrating long memory significantly improves the model's forecasting performance, particularly for longer time horizons. By shedding light on the potential benefits of incorporating long memory concepts into traditional term structure models, our research highlights its utility in capturing intricate temporal dependencies and enhancing prediction precision.