Transfer Learning of Discount Curves between Bonds and Swaps: An Empirical Study

This paper studies the joint estimation of risk-free discount curves based on government bonds and overnight-indexed interest rate swaps. Using daily data for CHF, EUR, GBP, and USD, we examine data quality, goodness of fit, and the role of transfer learning in extrapolating discount curves beyond the maturity range covered by government bonds. Motivated by regulatory requirements such as the Swiss Solvency Test, we build on kernel ridge regression and its transfer learning extension. We first estimate bond and swap discount curves separately and assess fitting errors, then apply transfer learning using a systematic masking experiment that mimics long-end data scarcity. Transfer learning substantially improves the extrapolation of government bond discount curves while leaving the fit at observed maturities essentially unchanged, whereas standalone estimation remains optimal for interest rate swap discount curves. The results support a transparent and data-driven approach to discount curve construction in practice.