Self-driving neural networks for term structure modeling

We propose a factor model with time-varying loadings for term structure modeling and fore casting. While maintaining the interpretation of the factors as level, slope, and curvature through explicit identification restrictions, we allow the loadings to take flexible shapes by specifying them as neural networks that evolve over time using a "self-driving" updating scheme based on past forecast errors, with gradient scaling to improve robustness. Using an empirically calibrated simulation study and an application to U.S. Treasury yields across 24 maturities, we show that flexible and dynamic factor loadings improve forecasting performance relative to standard benchmarks, including Nelson-Siegel models and the random walk. The gains are strongest at medium maturities and shorter forecast horizons, highlighting the importance of capturing curvature dynamics. In-sample results further illustrate how time-varying loadings provide insight into changes in yield curve shape beyond traditional parametric specifications.