The theme of the conference is “Artificial intelligence in the analysis of economic narratives, forecasting, and risk assessment.” In their paper, Sicco and his co-authors propose a self-driving neural network factor model for modeling yield curves. Their method offers a flexible and adaptive, yet interpretable, approach with better forecasting performance than static and less flexible models.
This research focuses on developing an alternative to the well-known Nelson-Siegel model, a three-factor model for the yield curve. In their paper, the authors model the factor loadings using neural networks with observation-driven parameters. Through carefully designed constraints, the output factors of this self-driving neural network factor model remain interpretable as the level, slope and curvature of the yield curve. The neural network and the self-driving dynamics lead to better forecast performance for U.S. government bond yields at horizons of one to twelve months ahead, relative to a variety of benchmark models, including the Nelson-Siegel model.
At the upcoming conference, Sicco will present these findings in Session 5 at 14:30: Time-Variation and State Dependence, a session dedicated to how research on time-varying models can improve economic and financial forecasting. Their contribution stands out for introducing novel time series machine-learning techniques to enhance yield curve forecasting.
The ECB Conference on Forecasting Techniques is one of Europe’s leading platforms for advances in economic forecasting, with a strong focus this year on the transformative role of artificial intelligence in economic analysis. A livestream, working paper and presentation slides will become available at the following link:
https://www.ecb.europa.eu/press/conferences/html/20260323_13th_conference_on_forecasting_techniques.en.html