Reader question: Market rates may be deep and liquid out to 20, 30 or 50 years, but an insurer or long-dated liability model can need discount factors beyond that. How can a curve algorithm extend beyond the last reliable market point without pretending that a 100-year rate was directly observed?
The Smith–Wilson method solves that problem by separating observable market information from an explicit long-run extrapolation assumption. It fits the present values of the chosen liquid input instruments, then lets the curve converge toward an ultimate forward rate through a smooth kernel controlled by a convergence-speed parameter called alpha.
The result is neither a pure market curve nor a free-form forecast. It is a rule-based hybrid: observed prices anchor the liquid part; the ultimate forward rate anchors the asymptotic long end; the Wilson function determines how the two are joined.
What this page owns — and what it does not
This page owns the computational transformation:
liquid market instruments + last liquid point + UFR + convergence rule → interpolated and extrapolated discount curve.
It does not replace yield-curve bootstrapping, which owns the general conversion of market instruments into discount factors and forwards; Nelson–Siegel–Svensson fitting, which owns a parametric cross-sectional yield model; or Kalman-filter yield-curve tracking, which owns dynamic latent-factor estimation through time.
This is term-structure mathematics and regulatory-methodology education. It is not an insurance valuation for a specific firm and not personalized financial advice.
The key boundary: market data stop somewhere
A curve engine should not treat every maturity identically when market liquidity disappears.
Let LLP denote the last liquid point. Before that point, selected market instruments are considered sufficiently deep, liquid and transparent for the methodology. Beyond it, direct market prices carry less authority or may not exist at all.
EIOPA’s current risk-free-rate technical documentation uses Smith–Wilson for interpolation where necessary and for extrapolation beyond the LLP. It explicitly separates the market-input section from the extrapolation controls.
Why the LLP matters mathematically
If a 100-year discount factor is not observed, the algorithm must decide how much of that value is driven by:
- the last observed market rates;
- the assumed ultimate forward rate;
- the speed of transition between them.
Moving the LLP changes the boundary between “market” and “model.” A 30-year LLP lets 30-year information directly anchor the fit; a 20-year LLP pushes maturities beyond 20 into the extrapolation regime.
This makes LLP sensitivity one of the first long-end diagnostics.
The ultimate forward rate
The ultimate forward rate is the asymptotic forward-rate level toward which the extrapolated forward curve is designed to converge.
It is not inferred solely from the last market quote. Under EIOPA’s Solvency II methodology it is derived separately under a stated long-run methodology and published by currency.
As a current example, EIOPA announced on 30 March 2026 that the euro UFR for 2027 remains 3.30%; the new values apply from 1 January 2027. EIOPA also stated that all relevant currencies except the Chinese renminbi remain unchanged for 2027, with the CNY UFR decreasing by 15 basis points to 4.05%.
Those values are time-sensitive policy inputs, not universal constants. A curve implementation should therefore version the UFR by effective date.
From annual rates to continuous intensities
EIOPA’s technical description introduces an annual interest factor:
R = 1 + r
and a continuous-time interest intensity:
ρ = ln(R).
If the ultimate annual forward rate is represented by UFR, an ultimate forward intensity can be written:
ω = ln(1 + UFR).
The asymptotic discount-factor backbone is then:
μ(t) = e−ωt.
The Wilson correction bends this simple asymptotic curve so it reproduces observed instrument prices before converging back toward the ultimate forward intensity.
The Wilson kernel
For maturities u and v, EIOPA defines the Wilson function in the form:
W(u,v) = e−ω(u+v) H(u,v)
with:
H(u,v) = α min(u,v) − e−α max(u,v)sinh[α min(u,v)].
Alpha controls the speed at which the influence of market nodes fades and the forward curve converges toward its ultimate level.
The function is smooth enough to generate a smooth present-value and yield curve at observed nodes, while still having a specific asymptotic structure.
Calibration in matrix form
Suppose observed zero-coupon prices are:
m = (m1, …, mn)
at maturities:
u = (u1, …, un).
Define:
μi = e−ωui.
Build the Wilson matrix:
Wij = W(ui, uj).
Using a convenient sign convention, solve:
Wζ = m − μ.
Then the fitted discount factor at any target maturity t is:
p(t) = e−ωt + W(t,u)ζ.
Some implementations define the calibration vector with the opposite sign and write the final expression with a minus sign. The mathematics is equivalent if the convention is consistent.
Why solving is better than explicitly inverting
The notation:
ζ = W−1(m−μ)
is mathematically compact but should not be read as an instruction to form a dense matrix inverse in production software.
A numerically safer implementation solves the linear system with a stable factorization. If the matrix is ill-conditioned, an explicit inverse magnifies floating-point error.
The diagnostic therefore includes the condition number of the Wilson matrix or an equivalent linear-solve stability measure.
Exact fit at observed maturities
A core Smith–Wilson property is that the present-value function can be fitted to agree with the observed input instrument prices at the chosen liquid maturities.
For zero-coupon inputs, the round-trip condition is simple:
p(ui) ≈ mi
within numerical tolerance.
For swaps or coupon bonds, the calibration uses their cash-flow matrix, and the discount curve must reproduce the input instrument present values rather than merely match a displayed spot rate.
EIOPA’s current technical documentation explicitly describes the input as instrument market prices, payment dates and a cash-flow matrix.
Coupon instruments turn the problem into cash-flow fitting
Let C be a cash-flow matrix and p the vector of discount factors at all cash-flow dates. Instrument model prices are:
V = CTp.
The Smith–Wilson calibration chooses its coefficients so:
V ≈ Vmarket.
This matters because a par swap quote or coupon-bond yield is not itself a zero-coupon discount factor. The cash-flow engine and market convention must be correct before the extrapolation can be trusted.
Alpha: the convergence-speed parameter
Small alpha makes market influence persist further into the long end. Larger alpha makes the forward curve converge toward the UFR faster.
Under EIOPA’s current technical documentation, alpha is chosen as the lowest value that reaches the convergence tolerance by the convergence point, subject to a lower bound of 0.05. The stated convergence tolerance is 1 basis point.
This is a useful conceptual distinction:
UFR chooses where the forward rate goes; alpha helps choose how fast it gets there.
The convergence point
EIOPA’s current technical documentation states the convergence point as:
max(LLP + 40 years, 60 years).
For the euro it describes a 20-year LLP and a 40-year convergence period, so the forward rate is intended to be close to its ultimate level around the 60-year maturity region under that framework.
Specific currencies can have justified exceptions. This is another reason the curve’s control inputs must be versioned by currency rather than hard-coded globally.
Interpolation and extrapolation happen in one mathematical object
Smith–Wilson is often discussed as a long-end extrapolator, but the same kernel can also fill missing maturities between observed nodes.
The target maturity vector can therefore contain:
- market maturities used in calibration;
- intermediate maturities with no direct quote;
- maturities far beyond the LLP.
The interpretation changes across those regions even though the same present-value function produces the numbers.
Discount factors first, yields second
The most fundamental output is the discount factor:
p(t).
A continuously compounded zero rate can then be derived as:
y(t) = −ln[p(t)] / t.
A continuously compounded instantaneous forward rate is:
f(t) = −d ln[p(t)]/dt.
The UFR convergence condition is most naturally understood through the forward-rate behaviour, not by merely checking that long-dated spot rates “look close” to the UFR.
Why the long-dated spot rate need not equal the UFR quickly
A forward rate describes a marginal future interval. A long-dated spot rate averages discounting over the whole path from today to maturity.
Even if the instantaneous forward rate is already close to the UFR at a distant maturity, the corresponding spot rate can still reflect earlier lower or higher rates.
This is a useful falsifier of a common implementation mistake: checking convergence by comparing the 60-year zero rate directly with the UFR can test the wrong quantity.
Negative rates are mathematically possible
EIOPA’s technical documentation explicitly allows negative annual interest rates provided the interest factor remains positive:
1 + r > 0.
The continuous intensity:
ln(1+r)
therefore remains defined as long as that condition holds.
A curve engine that assumes every discount factor must be below one can fail in negative-rate environments. Positive discount factors are essential; discount factors greater than one can be consistent with negative rates.
Positive discount factors are a stronger invariant than monotonicity
Under strictly positive rates, discount factors usually decrease with maturity. But if rates are negative over part of the curve, discount factors can increase.
The hard mathematical requirement for standard present-value use is:
p(t) > 0.
Monotonicity should be interpreted in the context of the rate regime, not used as a universal rejection rule.
Long-end sensitivity can be large even when input changes are small
Long-duration liabilities amplify discount-curve differences.
If a liability cash flow occurs at 70 years, a few basis points of long-end discount-rate change can produce a material present-value difference because duration is enormous.
Therefore validation should include not only yield residuals but also present-value sensitivity of representative long cash flows to:
- UFR;
- alpha;
- LLP;
- market-node perturbations.
Inputs and outputs
A Smith–Wilson engine can require:
- valuation date;
- market instrument types;
- market prices or adjusted par rates;
- cash-flow dates and amounts;
- day-count and payment conventions;
- selected liquid maturities;
- LLP;
- UFR by currency/effective date;
- convergence point;
- convergence tolerance;
- alpha policy or fixed alpha;
- target maturities.
Outputs should include discount factors, spot rates, forward rates, calibration residuals, alpha, condition diagnostics, convergence diagnostics and the exact methodology/version metadata.
Evidence polarity: what supports confidence?
Evidence for a curve includes exact repricing of calibration instruments, positive discount factors, smooth yields at observed nodes, forward convergence within the stated tolerance, stable linear solves, reproducibility from published inputs and sensible present-value sensitivity under small perturbations.
Evidence against confidence includes unexplained instrument repricing errors, large condition numbers, negative discount factors, oscillatory forwards near the LLP, alpha pinned unexpectedly at a bound, extreme long-end sensitivity to one market node, or disagreement with an independent implementation using the same control inputs.
Counterexample: exact market fit does not validate the extrapolation
Two extrapolation methods can both reproduce every liquid instrument exactly and produce materially different 70-year discount factors.
Market-fit residual is therefore a test of the liquid calibration, not proof that the long-end assumption is economically correct.
Counterexample: changing the UFR does not only move the farthest maturity
The Wilson kernel links the asymptotic anchor to the entire fitted present-value function. Changing the UFR can alter maturities before the formal convergence point as the curve transitions toward the new level.
A UFR sensitivity should therefore be examined as a curve vector, not one endpoint.
Counterexample: a larger alpha is not automatically “better convergence”
A larger alpha reaches the UFR faster, but faster convergence can reduce the persistence of market information beyond the LLP.
If the methodology says choose the lowest alpha satisfying a convergence requirement, selecting an unnecessarily large alpha changes the intended trade-off.
Counterexample: Nelson–Siegel–Svensson and Smith–Wilson solve different jobs
NSS fits a low-dimensional parametric yield shape and can be excellent for smooth cross-sectional estimation. Smith–Wilson is designed around exact present-value fit plus a specified ultimate forward convergence structure.
One should not judge them only by in-sample RMSE. Their extrapolation assumptions and use cases are different.
Weak links in implementation
Wrong instrument conversion. Par swap rates are treated as zero rates.
Cash-flow matrix error. Coupon dates or notionals are wrong before calibration begins.
UFR version error. A future-effective UFR is applied too early or an old UFR persists after its effective date.
LLP mismatch. The wrong currency or methodology version supplies the market boundary.
Explicit matrix inversion. Numerical error is magnified unnecessarily.
Forward/spot confusion. Convergence is tested against the wrong rate object.
Hidden unit error. Percent, decimal and continuously compounded intensities are mixed.
Insufficient precision. Long maturities magnify floating-point and rounding errors.
Diagnostics: how to test the engine
- zero-coupon round trip: reproduce every observed zero-coupon price.
- swap round trip: reprice each calibration swap to its market par value.
- independent implementation: compare against EIOPA’s published tool or a separately coded Wilson-kernel solver.
- condition-number test: monitor the Wilson/cash-flow system for numerical instability.
- UFR convergence test: verify the forward rate is within the stated tolerance at the convergence point.
- alpha scan test: independently reproduce the minimum alpha satisfying the criterion.
- LLP sensitivity test: move the LLP where analytically permitted and measure long-end PV impact.
- UFR sensitivity test: bump the UFR and inspect the entire extrapolated curve.
- negative-rate test: use a valid negative-rate curve and require positive discount factors.
- smoothness test: inspect yields and forwards around observed nodes and the LLP.
What would falsify confidence?
Confidence should be withdrawn if the engine cannot reproduce calibration-instrument prices; if the forward-rate convergence test fails under the stated alpha/UFR/LLP inputs; if small numerical perturbations produce extreme long-end jumps; if an independent implementation disagrees materially; if discount factors become non-positive; or if the exact methodology version and UFR effective date cannot be reconstructed.
Alternatives
Nelson–Siegel–Svensson gives a compact level/slope/curvature parameterisation.
Spline extrapolation can preserve local smoothness but needs an explicit boundary rule beyond the final node.
Constant-forward extrapolation is simple and transparent but makes the final observed forward unusually influential.
Macro-anchored methods can use long-run inflation and real-rate assumptions more directly, but introduce their own parameter uncertainty.
The question is not which method is universally superior. It is which extrapolation assumption is appropriate, transparent, reproducible and validated for the liability horizon being valued.
How this connects to the surrounding knowledge estate
The liquid inputs first depend on term-structure construction. Smith–Wilson then adds a governed long-end extrapolation layer rather than replacing the underlying cash-flow mathematics. Its shape can be contrasted with NSS fitting and dynamic Kalman-filter factor tracking. Long-end discount factors then feed present-value, duration and liability-sensitivity calculations elsewhere in the fixed-income mathematics estate.
Verification and update triggers
Preserve the technical-documentation version, input-market snapshot, cash-flow matrix, LLP, UFR effective date, convergence point, tolerance, alpha, compounding convention and linear-solver implementation. Revalidate after EIOPA methodology updates, annual UFR announcements, DLT/LLP changes, curve-input changes, software-library migrations or repeated long-end discrepancies across independent implementations.
Primary and high-quality references
- EIOPA, Risk-free interest rate term structures, including current RFR technical documentation and monthly technical files.
- EIOPA, RFR Technical Documentation, effective for the 2026 methodology, including LLP, convergence, Wilson-function and Smith–Wilson fitting sections.
- EIOPA, EIOPA publishes the Ultimate Forward Rate for 2027, 30 March 2026.
- lifelib, SmithWilson model documentation, an open implementation reference.
- Open Source Modelling, Smith–Wilson Python implementation, useful for independent replication against EIOPA examples.
Educational boundary: This article explains long-end term-structure extrapolation. It does not calculate regulatory technical provisions for any specific insurer or provide personalized financial advice.