Yield Curve Forecasting using Deep Learning Nelson-Siegel Model

I propose a deep learning term structure forecasting model by introducing a Nelson-Siegel layer that generates smooth yield curve forecasts while explaining the stylized facts of the term structure of interest rates. It is flexible enough to easily incorporate time-varying or time-invariant decay parameters into the deep learning framework. In the empirical analysis using U.S. yield data, the proposed model outperforms the benchmark DNS model in terms of lower RMSEs for 1-, 3-, 6-, and 12-month out-of-sample forecasts with statistical significance at 1, 3, and 6 months. Interestingly, it prefers the time-invariant decay parameter over the time-varying one and produces relatively smoother and less volatile latent yield factor estimates, which is interpreted as related to an attempt to reduce overfitting for the unseen yield curve. Finally, some relevant hyperparameters need to vary with the forecast horizon rather than uniformly in order to improve forecasting performance