We consider HJM forward rate models and dynamic factor models of the term structure for forecasting yield curves and the prices of coupon bonds using filter-based recurrent neural networks. We propose machine learning models that integrate the Kalman filter, extended Kalman filter, or the particle filter with arbitrage-free free regularization using deep neural networks. We consider the impact of the arbitrage penalty on different time horizons using a data set of U.S.treasuries and corporate bond issues. We analyze the predictive performance and the results of out-of-sample tests to show the efficiency and performance of our models. Our results show that arbitrage-free regularization improves model performance most significantly at the short end of the yield curve. This is joint work with Xiang Gao.

Bio: Cody Hyndman is Associate Professor and Chair of the Department of Mathematics and Statistics, Concordia University. He is the director of the Mathematical and Computational Finance program at Concordia and a co-founder of the NSERC CREATE program on Machine Learning in Quantitative Finance and Business Analytics (FIN-ML). His research interests include: mathematical and computational finance; probability and stochastic analysis; filtering and control; term-structure models; energy and commodity markets; and insurance mathematics. His recent research on machine learning algorithms and applications in finance has appeared in "Risks" and the "Journal of Machine Learning Research".