Abstract

The Secured Overnight Financing Rate (SOFR) has emerged as the leading benchmark for U.S. dollar-denominated interest rate derivatives, replacing LIBOR due to its transparency and robustness. This thesis develops a comprehensive framework for modeling and forecasting the SOFR term structure using machine learning methods, with a particular focus on CME SOFR futures. We first apply the official CME methodology to construct a daily, piecewise-constant SOFR forward rate curve, incorporating policy-driven discontinuities at Federal Open Market Committee (FOMC) dates. The curve is then fitted to the dynamic Nelson–Siegel (DNS) model, extracting time series of level, slope, and curvature factors. To capture and predict the evolution of these factors, we implement recurrent neural networks (RNNs), including Long Short-Term Memory (LSTM) architectures, and integrate a Kalman filter for state-space estimation. The performance of the proposed model is evaluated through out-of-sample forecasts under various loss functions, including mean squared error (MSE) and mean absolute error (MAE), and benchmarks against curve persistence. Our results show that the machine learning approach provides robust short-term forecasts for the SOFR term structure. However, accurately modeling and forecasting abrupt changes around monetary policy announcements remains a challenge, highlighting an important direction for future research. This research offers a practical and robust modeling strategy for interest rate risk management, with direct applications in the pricing and risk assessment of SOFR-linked financial products.