Discount Curve Estimation Standalone to Transfer Learning

This thesis studies the estimation of risk-free discount curves, a central problem in financial mathematics with wide-ranging applications in asset pricing, risk management, and insurance regulation. The three chapters develop a framework for constructing robust and economically consistent discount curves, with a particular focus on transfer learning across fixed-income product classes to address data sparsity, and on the challenges of extrapolation at long maturities.

Chapter 1 develops a framework for estimating the Swiss discount curve based on kernel ridge regression in a reproducing kernel Hilbert space. An extensive empirical study using Swiss government bond data shows that the method outperforms established benchmark models both in- and out-of-sample across maturities and over time. A Bayesian interpretation based on Gaussian processes enables uncertainty quantification. The chapter also discusses how the Swiss National Bank policy rate can be incorporated and outlines fundamental challenges of extrapolation.

Chapter 2 is theoretical in nature and extends this framework to a transfer learning setting in which related discount curves are estimated jointly across different fixed-income product classes. The problem is formulated as a vector-valued kernel ridge regression, yielding a convex optimization problem with a closed-form solution. An economically motivated spread regularization penalizes excessive deviations between curves. The framework admits a multivariate Gaussian process interpretation, enabling uncertainty quantification. Empirical evidence from US government bond and interest rate swap markets shows that transfer learning delivers economically significant improvements.

Chapter 3 applies the proposed methods to construct long-dated risk-free discount curves through the joint estimation of government bonds and overnight-indexed interest rate swaps for CHF, EUR, GBP, and USD. In regulatory and practitioner-relevant settings, it examines data quality and standalone estimation performance across product classes, and then extends the analysis to transfer learning by systematically assessing extrapolation using large-scale masking experiments. The results demonstrate that transfer learning yields economically significant improvements across product classes.