Stripping Discount Curves Across Currencies: Transfer Learning from US Treasuries

We provide precise international yield curves that constitute the new reference database for the government bond markets of six major economies, estimated with a robust and flexible machine learning method regularized by economic principles. We first show that the kernel ridge regression (KR) method of Filipovic, Pelger, and Ye (2024), the most precise method for US Treasuries, generalizes to all international bond markets and uniformly dominates the Nelson– Siegel–Svensson (NSS) benchmark used by many central banks. When foreign exchange (FX) markets provide additional information, we further improve upon KR with a novel transfer learning extension, KR Transfer (KR-T). KR-T converts US Treasuries into target-currency “synthetic bonds” using FX spot and forward rates and estimates the discount curves jointly under a weak covered interest parity (CIP) condition, which penalizes only the curvature of the cross-currency spread and leaves its level and slope unconstrained. Under block cross-validation that mimics real-world maturity gaps, KR-T reduces out of-sample yield errors by up to 51% relative to KR and by 71–86% relative to NSS, while matching KR where local data are sufficient. The gains concentrate at maturities where FX forwards are observed, identifying FX forward availability as the economic channel of transfer learning. KR-T also delivers more stable forward rates and more plausible yield curve dynamics.