Reader question: What should a yield curve do after reliable market prices stop? If 20-, 30- or 50-year liabilities must still be discounted, how can an algorithm connect the last liquid market information to a long-run forward-rate anchor in a controlled, reproducible way?

The Smith–Wilson method builds discount factors as a long-run exponential base curve plus a weighted sum of Wilson kernel functions. The kernel weights are chosen so that observed liquid instruments are reproduced, while the forward curve converges asymptotically toward an Ultimate Forward Rate (UFR). A convergence parameter α controls how quickly the market-implied part fades toward that long-run anchor.

This article owns one precise computational job: Smith–Wilson interpolation/extrapolation of discount factors from liquid-market observations toward a prescribed ultimate forward rate. It does not own ordinary bootstrapping, Nelson–Siegel–Svensson cross-sectional smoothing, regulatory selection of the UFR itself, or the new post-review Solvency II extrapolation formula that becomes applicable in 2027.

This is public mathematical and computational education. It is not financial advice, an actuarial opinion or a recommendation to use Smith–Wilson for every pricing or accounting curve.

1. The market eventually becomes too thin

Yield curves are partly observed and partly constructed.

At short and medium maturities, sufficiently deep and liquid instruments can anchor discount factors closely to market prices. Farther out, quotes may become:

sparse;
illiquid;
stale;
instrument-specific;
too noisy to support a fully market-driven curve.

Yet long-dated insurance, pension and other liabilities still require discount factors.

The extrapolation question is therefore unavoidable: where does observed market information end, and what mathematical rule governs the tail?

2. Smith–Wilson separates a long-run base from market corrections

Let f_∞ be the continuously compounded ultimate forward-rate parameter.

Define the long-run base discount curve:

μ(t)=e^−f∞t.

Smith–Wilson represents the fitted zero-coupon price as:

P(t)=μ(t)+Σ_i=1^N ξ_iW(t,u_i),

where:

u_i are observed/calibration maturities;
W(t,u) is the Wilson kernel;
ξ_i are calibration weights.

The base says “eventually behave like the UFR.” The kernels perturb that base so the curve reproduces market information at liquid maturities.

3. The Wilson kernel

One standard continuous-compounding form is:

W(t,u)=e^−f∞(t+u)[α min(t,u) − e^{−α max(t,u)}sinh(α min(t,u))].

Equivalent algebraic forms appear in technical documentation and code libraries.

The kernel depends on:

the UFR through f_∞;
the convergence parameter α;
the observation maturity u;
the target maturity t.

Its shape allows observed-market corrections to decay as maturity moves far beyond the liquid region.

4. Zero-coupon calibration is linear

Suppose zero-coupon prices P_obs(u_i) are observed at N maturities.

Define matrix:

W_ij=W(u_i,u_j).

Define base vector:

μ_i=e^−f∞u_i.

The calibration equations are:

P_obs=μ+Wξ.

So:

Wξ=P_obs−μ.

Once ξ is solved, any target maturity t can be interpolated or extrapolated using the same kernel expansion.

Some implementations define ζ=−ξ and therefore write P=μ−Wζ. The sign convention is different; the fitted curve is the same if the equations are internally consistent.

5. Never invert W explicitly if you can solve the system

Textbooks often write:

ξ=W⁻¹(P_obs−μ).

Numerical code should normally solve the linear system:

Wξ=b

using a stable factorization rather than explicitly forming W⁻¹.

Useful diagnostics include:

condition number;
smallest singular/eigenvalue;
residual ||Wξ−b||;
sensitivity to small input-price perturbations.

A curve that reprices perfectly can still be numerically fragile if W is nearly singular.

6. General coupon instruments use a cash-flow matrix

If the observed instruments are swaps or coupon bonds rather than zero-coupon bonds, let C be the cash-flow matrix.

Each instrument price is a weighted sum of underlying zero-coupon prices:

M=Cp.

The Smith–Wilson calibration can be written using the cash-flow exposures to the kernel functions. Schematically, market values are matched by solving a linear system involving:

CWC^T

and the difference between observed instrument prices and base-curve prices.

This means the cash-flow engine—dates, coupons, day counts and settlement conventions—is part of the curve algorithm, not a preprocessing footnote.

7. Exact calibration at the chosen liquid instruments

When the system is well posed and solved accurately, Smith–Wilson is designed to reproduce the calibration instruments exactly or to numerical tolerance.

That is different from Nelson–Siegel–Svensson, which usually accepts residual fitting error in exchange for a small global parameter set.

Exact fit is useful, but it creates a strong dependency on the selected input instruments. A single bad input can propagate through the entire kernel expansion.

8. What the UFR means mathematically

The UFR is the long-run forward-rate anchor.

As maturity becomes very large, the market-correction kernel terms decay and the fitted instantaneous forward rate converges toward f_∞.

If a UFR is quoted in annual effective terms UFR_eff, a continuous-compounding implementation commonly converts it as:

f_∞=ln(1+UFR_eff).

Using an annual-effective UFR directly inside a continuous-exponential formula without conversion creates a systematic tail error.

9. The UFR is an input, not an output of the kernel fit

The market observations determine ξ.

The UFR is prescribed externally under the methodology being used.

Therefore the far tail is partly model/policy driven by construction.

This distinction matters for long liabilities: a 100-year discount factor may be far more sensitive to the UFR than to the last few basis points of a 10-year market quote.

10. Alpha controls convergence speed

The parameter α>0 controls how rapidly forward rates approach the UFR.

Broadly:

larger α: faster convergence toward the UFR;
smaller α: market deviations persist farther into the extrapolated tail.

α is therefore not a harmless smoothing knob. It changes the allocation of influence between observable market rates and the ultimate anchor.

11. Convergence criteria turn alpha into an algorithm

Under the long-standing EIOPA Smith–Wilson framework, α has been chosen to make the forward rate sufficiently close to the UFR by a specified convergence point, subject to methodology rules and a lower bound.

Historical EIOPA technical documentation describes a 1-basis-point convergence tolerance and a convergence point based on the last liquid point plus a long extrapolation horizon.

A generic calibration algorithm is:

choose UFR and market input set;
choose candidate α;
solve the Smith–Wilson weights;
calculate forward rate at the convergence test maturity;
measure its distance from the UFR;
search for the smallest permitted α satisfying the tolerance.

Exact numerical rules are jurisdiction/methodology specific and can change over time.

12. Current regulatory context matters

As of August 2026, EIOPA still publishes monthly Solvency II risk-free term structures under the currently applicable framework and its RFR documentation. EIOPA’s published UFR for the euro applicable from 1 January 2026 is 3.30%.

However, EIOPA has also published revised RFR technical documentation for the Solvency II review that becomes applicable from 30 January 2027. That revised framework changes the extrapolation methodology and provides a new worked spreadsheet implementation.

Therefore Smith–Wilson remains important mathematics and remains relevant to current/legacy curves, but a public article should not imply that today’s exact regulatory implementation will remain unchanged indefinitely.

13. Last Liquid Point is a model boundary

The Last Liquid Point (LLP), or in revised frameworks related smoothing-point concepts, identifies where the methodology stops treating all farther maturities as fully reliable direct market anchors.

Moving the LLP changes:

which observations determine ξ;
where extrapolation begins;
how much of the long curve is market-driven;
long-liability present values.

Falsifier: stress the LLP/input-set selection and report liability/curve sensitivity. If a one-tenor eligibility change materially moves the tail, that dependency belongs in model risk.

14. A real example of input-selection sensitivity

EIOPA updated its RFR technical documentation in 2025 after the liquidity assessment of the 19-year euro swap tenor changed. Removing that tenor altered which point sat immediately before the 20-year euro liquid boundary and was intended to reduce unwanted volatility in the extrapolated part of the curve.

This is a useful public lesson: even when the extrapolation kernel is unchanged, input liquidity governance changes the tail.

15. Discount factor, spot rate and forward rate are different diagnostics

From fitted discount factor P(t), a continuous spot yield is:

y(t)=−ln P(t)/t.

The instantaneous forward rate is:

f(t)=−d ln P(t)/dt.

A spot curve can look smooth while its derivative reveals tail kinks or oscillations.

Always inspect forward rates as well as spot rates when validating an extrapolation method whose target is an ultimate forward rate.

16. Exact market fit does not guarantee pleasant forwards

Kernel interpolation can pass exactly through liquid inputs while producing forward shapes between or after those inputs that are:

steep;
locally non-monotone;
sensitive to sparse anchors;
numerically noisy after rounded spot-rate differentiation.

An exact repricing test is necessary but not sufficient.

17. Negative rates do not automatically invalidate Smith–Wilson

The method operates on discount factors and an exponential base. It can accommodate observed negative spot rates under consistent compounding conventions.

A discount factor above one is possible in a negative-rate environment.

Therefore a simplistic rule “P(t) must always decrease” is not universally valid.

The correct checks are consistency with observed instruments, compounding, cash-flow prices and the methodology’s admissibility conditions.

18. Long-liability sensitivity

For a liability cash flow L_T at long maturity T:

PV=L_TP(T).

If P(T) is predominantly driven by the UFR/α extrapolation, then long-liability PV can be highly sensitive to those assumptions.

A model audit should therefore calculate:

ΔPV under UFR shocks;
ΔPV under α changes;
ΔPV under LLP/input-set changes;
key-maturity contributions beyond the liquid region.

This does not mean the method is wrong; it reveals where valuation authority has moved from market observation to extrapolation assumptions.

19. UFR sensitivity is nonlinear in maturity

A small UFR change has little impact close to the liquid boundary because market-correction kernels still dominate.

Far into the tail, the UFR dominates more strongly.

Therefore the same one-basis-point UFR shock produces different relative discount-factor effects at 25, 50, 75 and 100 years.

Falsifier: do not summarize UFR risk with one scalar sensitivity measured at one maturity.

20. Alpha sensitivity is also maturity dependent

Changing α mostly changes the transition path from market curve to UFR.

Very far out, curves with different α converge toward the same UFR; close to liquid inputs, exact fit constrains them.

The largest differences often occur in the extrapolation corridor between those regions.

This makes α stress plots across maturity more informative than one terminal number.

21. Matrix conditioning

The Wilson matrix can become difficult when:

many maturities are extremely close;
α is extreme;
cash-flow instruments are nearly redundant;
units/compounding create very different numerical scales.

Large ξ coefficients that cancel each other can produce an apparently smooth curve but severe input sensitivity.

Falsifier: perturb market prices by realistic bid–ask amounts. If long extrapolated rates move disproportionately, the calibration is poorly conditioned.

22. Use log-sensible numerical formulas

At very long maturity, e^−f∞t and Wilson kernels can become very small.

Numerical implementations should test:

underflow/overflow;
stable hyperbolic-function evaluation;
large maturity units;
continuous/effective rate conversion;
precision of forward derivatives.

Printing rounded spot rates and then differentiating them can create artificial forward-rate jaggedness, a point also noted in EIOPA technical Q&A material.

23. Instrument-price reconciliation

After calibration, reprice every input instrument from the constructed discount curve.

For a coupon instrument:

P_model=Σ C_jP(t_j).

Compare with:

P_market.

Residuals beyond numerical tolerance indicate:

cash-flow convention error;
matrix/calibration error;
wrong compounding;
wrong settlement/accrual;
incorrect UFR/α implementation.

24. Smith–Wilson versus ordinary bootstrapping

Yield-curve bootstrapping algorithms own construction of discount factors from liquid instruments.

Smith–Wilson becomes especially important when the job includes interpolation/extrapolation beyond the reliable market horizon toward an externally specified long-run rate.

The two layers can be combined: a liquid curve can be bootstrapped first, then Smith–Wilson can be applied to selected liquid discount-factor anchors.

25. Smith–Wilson versus Nelson–Siegel–Svensson

Nelson–Siegel–Svensson uses a small number of global exponential factors and generally minimizes fitting error.

Smith–Wilson uses as many kernel weights as needed to reproduce selected market instruments, plus explicit UFR and α tail controls.

NSS is parsimonious global smoothing; Smith–Wilson is kernel interpolation/extrapolation with an ultimate-forward anchor.

26. Smith–Wilson versus Hagan–West

Hagan–West monotone-convex interpolation owns local shape-preserving interpolation between market forward-rate knots.

Smith–Wilson is global: one observed anchor can influence many target maturities through the kernel functions, especially in the tail.

27. Smith–Wilson is not a forecasting model

The fact that the far-forward curve converges to a UFR does not mean the method predicts that future realized short rates will equal that path.

It is a valuation/extrapolation rule.

Falsifier: keep forecast evaluation separate from discount-curve construction. Historical forecast error cannot by itself validate or invalidate a prescribed extrapolation method.

28. Smith–Wilson is not automatically a complete stochastic no-arbitrage model

A deterministic curve constructed today can price deterministic cash flows consistently.

That does not specify how the entire curve evolves tomorrow under a risk-neutral measure.

Derivative pricing that requires stochastic term-structure dynamics still needs a model such as HJM, Hull–White or another arbitrage-consistent framework calibrated to the initial curve.

29. Inputs and outputs

Inputs can include:

liquid instrument prices/rates;
cash-flow schedules;
calibration maturities;
LLP/eligible-input rule;
UFR and compounding convention;
α or α-selection rule;
target maturities;
day-count/settlement conventions;
linear-solver tolerance;
forward-rate differentiation method.

Outputs can include:

kernel-weight vector ξ;
fitted discount factors;
spot rates;
forward rates;
input repricing residuals;
α convergence error;
matrix condition diagnostics;
UFR/α/LLP sensitivity;
long-liability PV sensitivity;
extrapolation confidence/governance flags.

30. Evidence polarity

Evidence for confidence includes:

all calibration instruments reprice to tolerance;
matrix conditioning is acceptable;
forward curve approaches the prescribed UFR under the required criterion;
results are stable to realistic quote perturbations;
compounding conversions reconcile independently;
alternative implementations reproduce the same curve;
long-tail sensitivities are understood and within governance expectations;
input-liquidity changes produce explainable, monitored effects.

Evidence against confidence includes:

large ξ coefficients with cancellation;
ill-conditioned Wilson/cash-flow matrix;
forward convergence criterion fails;
small market-input changes create large tail moves;
spot-rate rounding causes unstable forward diagnostics;
UFR effective/continuous conversion is inconsistent;
long-liability PV is highly assumption-sensitive but reported as market-observed;
regulatory methodology changes are ignored.

31. Counterexample: wrong UFR compounding

The methodology supplies a 3.30% annual-effective UFR, but code inserts 0.033 directly as f_∞ in e^−f∞t.

The correct continuous equivalent would be:

ln(1.033).

The difference is small per year but compounds over very long maturities.

Falsifier: unit-test effective/continuous conversions at 30, 60 and 100 years.

32. Counterexample: perfect fit, unstable tail

Two adjacent input discount factors are almost redundant and W is poorly conditioned.

Both are fitted exactly, but changing one by 0.1 basis point moves the 70-year forward rate materially.

Falsifier: perturb inputs within bid–ask/noise ranges and inspect tail sensitivity before accepting exact fit as evidence of robustness.

33. Counterexample: alpha chosen too small

A very small α preserves market-shape deviations far into the future and fails the required UFR convergence tolerance by the prescribed convergence point.

Falsifier: calculate the forward-rate gap to UFR at the formal test maturity, not merely the spot-rate gap.

34. Counterexample: alpha chosen too large

A very large α forces the curve rapidly toward the UFR immediately after the liquid region.

This can make long liabilities overly anchored to the external UFR and reduce the influence of observable market information faster than intended.

Falsifier: compare forward curves and liability PV across the permissible α range and document the methodology criterion.

35. Counterexample: stale liquid-point classification

A tenor previously considered liquid becomes thin and noisy but remains an input.

Because Smith–Wilson fits it exactly, its noisy price can distort the extrapolation corridor.

Falsifier: repeat the fit excluding that tenor and compare tail rates/PV. Liquidity governance must precede exact interpolation.

36. Counterexample: rounded spot curves used to validate forwards

A published spot curve is rounded to a few decimal places. Numerical differentiation of those rounded rates creates a jagged forward curve.

Falsifier: validate forwards from full-precision discount factors or analytical derivatives where available.

37. Counterexample: 2027 rules assumed to be identical to 2026

A system hard-codes the current Smith–Wilson extrapolation and assumes no regulatory change.

EIOPA has already published revised RFR documentation and an example workbook for the methodology that becomes applicable on 30 January 2027.

Falsifier: version the curve methodology by effective date and regression-test old/new frameworks separately. Public educational material should distinguish the Smith–Wilson method from whichever regulatory rule is currently in force.

38. Alternatives

Nelson–Siegel–Svensson: parsimonious global parametric curve.

Splines/monotone-convex interpolation: local market-driven interpolation with separate extrapolation rules.

Regulatory revised extrapolation formulas: methodology-specific blending of market forward information and UFR.

Macro-finance long-rate models: economic long-run modeling rather than prescribed kernel extrapolation.

The appropriate method depends on whether the job is derivative pricing, accounting, regulatory valuation, ALM or economic forecasting.

40. What would falsify confidence?

Confidence should be withdrawn if observed instruments fail to reprice; if independent implementations disagree; if the forward convergence rule fails; if input perturbations create unexplained tail instability; if matrix conditioning is severe; if long-liability values are dominated by undocumented UFR/α assumptions; or if the wrong methodology version is applied for the valuation date.

41. Verification and update triggers

Preserve the exact methodology version, input instruments, cash-flow matrix, UFR and compounding basis, α rule, LLP/input eligibility, Wilson-matrix diagnostics, calibration weights and full-precision curve outputs.

Revalidate when:

UFR changes;
liquid-point/input eligibility changes;
α/convergence rules change;
regulatory technical documentation changes;
new long-dated instruments become liquid;
matrix condition deteriorates;
long-liability sensitivity increases;
the curve is repurposed for stochastic derivative pricing.

42. Primary and high-quality references

European Insurance and Occupational Pensions Authority, Risk-Free Interest Rate Term Structures, current monthly publications and technical documentation.
EIOPA, Updated RFR Technical Documentation, changes effective from January 2026.
EIOPA, Ultimate Forward Rate for 2026, including the euro UFR of 3.30% applicable from 1 January 2026.
EIOPA Q&A 3564, 2027 Extrapolation Documentation and Example Workbook, confirming the revised methodology resources.
Note on the Smith–Wilson Interest Rate Curve, International Journal of Theoretical and Applied Finance, 2016, for the kernel/RKHS interpretation.
Bank of England Prudential Regulation Authority, Technical Information for Solvency II Firms, including 2026 Smith–Wilson extrapolation parameter files.

Educational boundary: Smith–Wilson makes the market-to-model boundary explicit. Its value is not that it discovers an unknowable 100-year market rate, but that it provides a transparent mathematical rule connecting liquid observations to a specified ultimate-forward assumption—and exposes which long-dated valuations depend on that assumption.

Small Group Tutorials

How Smith–Wilson Yield-Curve Algorithms Extrapolate Beyond Liquid Markets: Wilson Kernels, Ultimate Forward Rates, Alpha Convergence, Exact Calibration, Long-Liability Sensitivity and Method Change