Paper: arXiv 2505.07676

Authors: Nicolas Camenzind, Damir Filipovic

Abstract

We propose a framework for transfer learning of discount curves across different fixed-income product classes. Motivated by challenges in estimating discount curves from sparse or noisy data, we extend kernel ridge regression (KR) to a vector-valued setting, formulating a convex optimization problem in a vector-valued reproducing kernel Hilbert space (RKHS). Each component of the solution corresponds to the discount curve implied by a specific product class. We introduce an additional regularization term motivated by economic principles, promoting smoothness of spread curves between product classes, and show that it leads to a valid separable kernel structure. A main theoretical contribution is a decomposition of the vector-valued RKHS norm induced by separable kernels. We further provide a Gaussian process interpretation of vector-valued KR, enabling quantification of estimation uncertainty. Illustrative examples demonstrate that transfer learning significantly improves extrapolation performance and tightens confidence intervals compared to single-curve estimation.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper is mathematically dense, featuring advanced theory like vector-valued RKHS, separable kernels, and Gaussian process interpretations, with extensive LaTeX derivations. While it includes illustrative examples and promises empirical studies, it lacks code, backtests, or statistical metrics, focusing more on theoretical formulation than implementation-heavy validation.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop transfer learning<br>framework for discount curves across<br>fixed-income product classes"] --> B{"Data: <br>Historical market data for<br>multiple product classes"}
  B --> C["Methodology: <br>Vector-valued Kernel Ridge Regression"]
  C --> D["Formulation: <br>Convex optimization in vector-valued RKHS<br>with separable kernels"]
  D --> E["Computation: <br>Gaussian Process interpretation for<br>uncertainty quantification & confidence intervals"]
  E --> F["Outcome: <br>Transfer learning improves extrapolation &<br>tightens confidence intervals vs. single-curve"]
  F --> G["Key Contribution: <br>Theoretical decomposition of RKHS norm<br>+ economic regularization for spread smoothness"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a transfer learning framework that jointly estimates discount curves across fixed‑income product classes using vector‑valued kernel ridge regression.

Source sentence
We propose a framework for transfer learning of discount curves across different fixed‑income product classes. Motivated by challenges in estimating discount curves from sparse or noisy data, we extend kernel ridge regression (KR) to a vector‑valued setting, formulating a convex optimization problem in a vector‑valued

Assumes 5

  • The theoretical price of any fixed‑income product can be expressed as the sum of its discounted cash flows. model
    Source sentence
    Our framework requires only that the theoretical price of a fixed‑income product be expressed as the sum of its discounted cash flows.
  • Pricing errors are Gaussian with zero mean and a block‑diagonal covariance matrix. statistical
    Source sentence
    Σ = … for symmetric positive definite Ma × Ma‑matrices Σa.
  • When no quotes are available for a product class, the corresponding rows are omitted and the solution still depends on other classes via joint regularization. data
    Source sentence
    Theorem 2.1 remains valid even when no quotes are available for a given product class a, i.e., when Ma = 0. In this case, the corresponding rows in C and P are omitted, and we adopt the convention that P0 i=1 = 0.
  • Bid‑Ask spreads are tight, typically below 2 basis points, providing a benchmark for economic relevance. market
    Source sentence
    Overall, both product classes exhibit tight Bid–Ask spreads, typically below 2 basis points (bps).
Show 1 more
  • The prior mean function is a constant function equal to one. model
    Source sentence
    As prior we use the constant function p = 1.

Data

  • US government bonds · Bloomberg Finance L.P. market-prices commercial
    January 2020 through December 2024 · 610 fully taxable, non‑callable coupon bonds (excluding Treasury bills)
    Source sentence
    For US government bonds, we rely on daily end‑of‑day (mid) dirty prices
  • SOFR swaps · Bloomberg Finance L.P. market-prices commercial
    January 2020 through December 2024 · Daily end‑of‑day swap rates for tenors 1D to 8M
    Source sentence
    For SOFR swaps, we use daily end‑of‑day (mid) swap rates. The floating leg

Recognised sources: Bloomberg

Costs

Unclear - sources disagree The platform classifier and the paper-text extraction give different answers, so no value is asserted.

Platform classifier: Not applicable

Paper text extraction: Not modeled

Source sentence (the authors' words)
Overall, both product classes exhibit tight Bid–Ask spreads, typically below 2 basis points (bps).

Finds 3

  • Transfer learning significantly improves extrapolation performance in a masking experiment while leaving well‑identified regions unaffected. performance
    Source sentence
    An extensive masking experiment demonstrates that transfer learning US government bonds with SOFR swaps improves extrapolation while leaving well‑identified regions unaffected.
  • Transfer learning tightens confidence intervals compared to single‑curve estimation. statistical
    Source sentence
    Illustrative examples show how transfer learning tightens confidence intervals compared to single‑curve estimation.
  • The method is robust to hyper‑parameter variation; LOOCV RMSE heatmaps show stable performance across a wide grid. theoretical
    Source sentence
    Figure 4 displays the aggregated RMSE in bps across hyperparameter combinations … These heatmaps underscore the robustness of the method to hyperparameter variation.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (70)

  1. Shrinking the Term Structure (2022) — link
  2. Transfer Learning Across Fixed-Income Product Classes (2025) — link
  3. The T reasury– SOFR Swap Spread Puzzle Explained (2024) — link
  4. Arbitrage Theory in Continuous Time (2009) — link
  5. Reconstructing the yield curve (2021) — link
  6. The Information in Long-Maturity Forward Rates (1987) — [link](http://www.jstor.org/stable/1814539 http://www.jstor.org/stable/1814539?seq=1{&}cid=pdf-reference{#}references{_}tab{_}contents http://about.jstor.org/terms)
  7. Estimating and Interpreting Forward Interest Rates: Sweden 1992 - 1994 (1994) — link
  8. Parsimonious Modeling of Yield Curves (1987)
and 62 more
  1. The U.S. Treasury yield curve: 1961 to the present (2007) — link
  2. Kernel methods for vector output
  3. Kernels for Vector-Valued Functions: A Review (2012) — link
  4. Exact Smooth Term-Structure Estimation
  5. Stripping the Discount Curve — a Robust Machine Learning Approach (2024) — link
  6. Stripping the Discount Curve - a Robust Machine Learning Approach (2022) — link
  7. An Introduction to the Theory of Reproducing Kernel Hilbert Spaces (2016)
  8. Kernels for Vector-Valued Functions: a Review (2012)
  9. Multi-output learning via spectral filtering (2012)
  10. Graphical Multi-Task Learning (2008)
  11. Multivariate Geostatistics (2003)
  12. On learning vector-valued functions (2005)
  13. Kernels for multi-task learning (2004)
  14. Clustered Multi-Task Learning: A Convex Formulation (2004)
  15. Multivariate Gaussian and Student-t process regression for multi-output prediction (2019) — link
  16. Matrix Analysis (2012)
  17. Operator-valued Kernels for Learning from Functional Response Data (2016) — link
  18. Parsimonious Modeling of Yield Curves (1987) — link
  19. Estimating and Interpreting Forward Interest Rates: Sweden 1992 - 1994 (1994)
  20. The U . S . Treasury yield curve: 1961 to the present (2007) — link
  21. The Information in Long-Maturity Forward Rates (1987)
  22. Reconstructing the yield curve (2021) — link
  23. Fitting yield curves with long term constraints (2001)
  24. Stripping the Swiss discount curve using kernel ridge regression (2024) — link
  25. Multitask Learning (1997) — link
  26. DeCAF: A Deep Convolutional Activation Feature for Generic Visual Recognition (2014) — link
  27. End-to-end multitask learning, from protein language to protein features without alignments (2019) — link
  28. Joint Learning of International Yield Curves (2023)
  29. Kernels and Regularization on Graphs (2003) — link
  30. Gaussian processes for machine learning (2005)
  31. Technical documentation of the methodology to derive EIOPA ’s risk-free interest rate term structures (2021)
  32. Delegated Regulation (EU) 2015/35 (2015)
  33. Modeling Term Structures of Defaultable Bonds (1999)
  34. Beyond LIBOR: a primer on the new benchmark rates (2019) — link
  35. Linear-rational term structure models (2017)
  36. Secured O vernight F inancing R ate (SOFR)
  37. Swiss R eference R ates (SARON)
  38. Life after L ibor: A new era of reference interest rates
  39. LCH clears first €STR swaps
  40. Euro short-term rate (€STR)
  41. Economic B ulletin, I ssue 4 (2022)
  42. Sterling O vernight I ndex A verage (SONIA)
  43. SONIA key features and policies
  44. Benchmark R eform – USD M arket
  45. Transition to RFR s R eview: F ull Y ear 2020 and the F ourth Q uarter of 2020 (2020)
  46. Transition to RFR s R eview: F ull Y ear 2021 and the F ourth Q uarter of 2021 (2021)
  47. Transition to RFR s R eview: F ull Y ear 2022 and the F ourth Q uarter of 2022 (2022)
  48. Transition to RFR s R eview: F ull Y ear 2023 and the F ourth Q uarter of 2023 (2023)
  49. NEX G roup
  50. Foreign exchange swaps and cross-currency swaps (2023) — link
  51. Role of cross currency swap markets in funding and investment decisions (2019) — link
  52. The widening of cross-currency basis: When increased FX swap demand meets limits of arbitrage (2024) — link
  53. Beyond Libor: the Impact of SOFR on Rates, Bonds and Loans (2021) — link
  54. A Quick Guide to the Transition to Risk-Free Rates in the International Bond Market (2020)
  55. Mechanics of Cross Currency Swaps (2017)
  56. Learning partial differential equations in reproducing kernel Hilbert spaces (2023)
  57. A review of recent advances and strategies in transfer learning (2025)
  58. A survey on transfer learning (2010)
  59. A survey of transfer learning (2016)
  60. Pricing derivatives on financial securities subject to credit risk (1995)
  61. Teaching Economics to the Machines (2023) — link
  62. Financial Machine Learning (2023) — link

0 of 70 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.