A model that seeks to estimate an interest rate curve should have two desirable capabilities in addition to the usual characteristics required from any function-estimation model: it should incorporate the bid-ask spreads of the securities from which the curve is extracted and restrict the curve shape. The goal of this article is to estimate interest rate curves by using Support Vector Regression (SVR), a method derived from the Statistical Learning Theory developed by Vapnik (1995). The motivation is that SVR features these extra capabilities at a low estimation cost. The SVR is specified by a loss function, a kernel function and a smoothing parameter. SVR models the daily U.S. dollar interest rate swap curves, from 1997 to 2001. As expected from a priori and sensibility analyses, the SVR equipped with the kernel generating a spline with an infinite number of nodes was the best performing SVR. Comparing this SVR with other models, it achieved the best cross-validation interpolation performance in controlling the bias-variance trade-off and generating the lowest error considering the desired accuracy fixed by the bid-ask spreads.Bid-ask spread, Interest rate curves, Interest rate swaps, Support Vector Regression,