OAS is the spread a bond offers over the risk-free rate, adjusted for embedded optionality. Getting to OAS requires knowing the risk-free rate at every relevant cash flow date. Getting that requires interpolating the yield curve between the benchmark tenor points you actually observe. Most systems use cubic spline interpolation. For the vast majority of bonds, it is accurate enough. For bonds with unusual coupon structures or maturities that fall in sparse regions of the curve, cubic splines introduce systematic errors that propagate through the fair value estimate. We have spent time on this problem and want to be direct about what we found.
The Standard Interpolation Setup and Where It Breaks
The standard USD risk-free curve is built from on-the-run Treasury yields at key benchmark maturities: 3-month, 6-month, 1-year, 2-year, 3-year, 5-year, 7-year, 10-year, 20-year, and 30-year. Cubic spline interpolation fits a smooth piecewise cubic polynomial through these knot points to produce a continuous curve.
Cubic splines impose a specific mathematical smoothness constraint: the first and second derivatives of the fitted curve must be continuous across each knot. This is a reasonable assumption for most of the curve. Where it breaks down is in regions where the underlying rate structure has local features that the smooth spline cannot capture, or where knot point spacing is wide relative to the maturity of the bond being priced.
The clearest example is the 7-year to 10-year region. The on-the-run benchmarks at 7 and 10 years are directly observed. Bonds maturing at 8 or 9 years are priced off the interpolated segment between these two knots. In a normal, positively sloped curve environment, the interpolation is smooth and the error is small. During curve inversions, periods of unusual curve steepness or flatness at specific segments, or when the 7-year and 10-year on-the-run yields diverge significantly from their theoretical smooth-curve values due to on-the-run richness effects, the interpolated values at 8 and 9 years carry non-trivial errors relative to what a well-traded bond in that maturity range would actually yield.
Quantifying the Error in Practice
Rather than stating this in the abstract, we can describe the magnitude of the problem in terms we observed during our internal model development (validation against TRACE data, 2023-2025 window, covering approximately 1,400 IG CUSIPs with maturities between 7 and 10 years).
For bonds maturing in the 7.5 to 9.5-year range, the OAS difference between using standard cubic spline interpolation versus our corrected curve was typically under 2bp in normal market conditions. Not zero, but small enough that it would not drive a meaningful pricing decision on its own.
During the periods in our validation window where the 7-to-10 curve segment showed unusual convexity (specifically, two episodes in 2023-2024 where the segment flattened sharply relative to adjacent tenors), the interpolation error for bonds in this maturity range widened to a range of 5 to 12bp. For context, a typical IG corporate bond's fair value OAS might be in the range of 100 to 200bp, so a 5 to 12bp systematic error represents roughly a 3 to 8% relative distortion. That is large enough to matter when a desk is quoting at 2bp spreads or comparing two bonds as relative value.
Bonds with Non-Standard Cash Flow Schedules
The interpolation problem is more concentrated for bonds with cash flows that fall at non-standard calendar dates. Most corporate bonds pay semi-annual coupons at reasonably standard intervals, and their cash flows map onto the interpolated curve at points where interpolation error is small.
Bonds with unusual first coupon periods (long or short initial coupon), bonds with accrued interest from prior periods carried through corporate actions, or bonds from foreign issuers structured to pay on non-US-standard schedules can have cash flows landing at interpolated curve points that are particularly sensitive to interpolation method. For these bonds, the choice of interpolation method can have disproportionate impact on the computed OAS.
Callable corporate bonds compound the problem. The OAS calculation for a callable bond requires discounting cash flows over multiple scenario paths, each of which depends on the interpolated rates at a grid of future dates. If the interpolation error is systematic in a particular curve region, it compounds across the Monte Carlo paths rather than averaging out. In our internal testing, the OAS error for callable IG bonds with exercise dates in the 7 to 9-year range was larger in absolute terms than the equivalent error for bullet bonds with the same maturity, consistent with the path-compounding mechanism.
Our Correction Approach
We moved away from standard natural cubic splines as the sole interpolation method early in model development. The current pipeline uses a regime-conditional interpolation framework that selects or blends methods based on local curve geometry.
For curve segments that are smooth and well-behaved (positive slope, stable curvature, no large rate moves in the prior session), natural cubic splines remain the baseline. They are computationally efficient and accurate in normal conditions.
For segments showing unusual local convexity or high recent volatility, we switch to a monotone cubic Hermite interpolation, which preserves local monotonicity within each segment and avoids the oscillation artifacts that natural splines can produce when knot values are not smooth. The tradeoff is that Hermite interpolation is less smooth globally, but for bond pricing purposes, the local accuracy in the relevant maturity segment matters more than global smoothness.
For the 20-to-30 year region, where on-the-run Treasury supply dynamics and pension demand create persistent distortions at the long end, we supplement the spline with a Nelson-Siegel-Svensson level-slope-curvature parametric fit as a sanity check. When the two methods diverge by more than a threshold (currently 3bp equivalent in yield), we flag the affected maturities for manual review rather than automatically preferring one over the other.
The Practical Implication for Fair Value Comparison
The reason this matters for a desk using a fair value signal is that the same OAS error is present in whatever pricing tool the desk is comparing against, if that tool also uses standard cubic splines. When two systems produce slightly different OAS estimates for the same bond, part of that difference may be attributable to different interpolation methods rather than different views on the bond's credit spread. Understanding which part is which helps avoid misinterpreting model divergence as a trading signal when it is actually a methodological artifact.
We surface this distinction explicitly when bonds in our coverage fall into maturity ranges or curve environments where interpolation method choice has meaningful impact. The signal output includes a flag for bonds where the cross-method OAS difference exceeds our threshold, along with the magnitude of that difference. A desk seeing that flag knows to treat the OAS estimate with somewhat wider uncertainty bounds, not because our model is less confident in the credit, but because the risk-free rate anchor itself is less precisely determined for that maturity.
This is not a failure of the pricing approach. It is an honest accounting of a structural limitation in fixed-income valuation that exists regardless of what pricing tool you use. We would rather surface it explicitly than have a desk discover the discrepancy when comparing our output to another system and conclude that one of us must be wrong.