Support Vector Regression in Financial Curve Construction: A Comparative Analysis of Benchmark-Implied Curves in Interest Rate and Commodity Markets
2026 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
Student thesis
Abstract [en]
This thesis evaluates whether Support Vector Regression (SVR) can improve financial curve construction by reducing functional instability without sacrificing market consistency. The study applies SVR as a post-processing smoother to benchmark-implied curves in both interest-rate and commodity markets. Three interest-rate curves, EUR-ESTR, USD-SOFR, and EURIBOR-6M-FWD, and one commodity forward curve, NYMEX-WTI, are examined. SVR models with polynomial, radial basis function, and sigmoid kernels are compared against traditional benchmark interpolation methods: Linear RT, Cubic Spline, and Hermite interpolation.
The empirical evaluation is based on two criteria: out-of-sample repricing accuracy, measured by mean squared error, and functional stability, measured by the total variance of daily curve changes. The results show that SVR does not provide a superior overall trade-off between market consistency and functional stability. The RBF kernel generally achieves the lowest repricing errors among the SVR models, but often at the cost of increased curve instability. Polynomial and sigmoid kernels can reduce total variance in several cases, but typically produce materially larger pricing errors. Overall, the benchmark interpolation methods remain more effective for the curves studied. The findings suggest that machine-learning methods for financial curve construction may need to incorporate financial pricing structure more directly in order to add practical value in a banking context.
Place, publisher, year, edition, pages
2026. , p. 81
Keywords [en]
Support Vector Regression, Financial Curve Construction, Yield Curve Modelling, Benchmark Interpolation, Interest Rate Curves, Commodity Forward Curves, Curve Smoothing, Market Consistency, Functional Stability, Out-of-Sample Repricing
National Category
Probability Theory and Statistics
Identifiers
URN: urn:nbn:se:liu:diva-225284ISRN: LIU-IEI-TEK-A--26/05486--SEOAI: oai:DiVA.org:liu-225284DiVA, id: diva2:2075505
External cooperation
SEB
Presentation
2026-06-15, A38, Olaus Magnus väg 37, 583 30 Linköping, Linköping, 16:53 (Swedish)
Supervisors
Examiners
2026-06-222026-06-182026-06-22Bibliographically approved