How Hull–White Calibration Algorithms Fit Short-Rate Dynamics to Yield Curves and Swaption Prices: θ(t), Mean Reversion, Volatility, Repricing Tests and Identifiability

Reader question: If today’s yield curve is already observable, why does an interest-rate model need calibration at all, and how can one short-rate process be made consistent with both today’s curve and option prices?

This article owns the Hull–White one-factor calibration problem: an observed discount curve + an interest-rate option calibration set → a time-dependent drift function, mean-reversion parameter and volatility specification that can reproduce the chosen market observables closely enough for the intended pricing task.

It does not own callable-bond OAS, SOFR-futures convexity adjustment or generic swaption quoting. Those are separate downstream jobs already covered in the Bukit Timah Tutor finance-and-banking-algorithms lane. Here the object is the model-fitting machinery itself.

This is mathematical education about stochastic modelling. It does not recommend trades, securities or financial products.

The model in one equation

A common one-factor Hull–White specification under a risk-neutral pricing measure is:

dr_t = [θ(t) − a r_t]dt + σ dW_t.

Equivalent notation is often written as:

dr_t = a[m(t) − r_t]dt + σ dW_t.

The notation varies across books and implementations. The underlying roles are:

• r_t: instantaneous short rate;
• a: mean-reversion speed;
• σ: short-rate volatility, possibly generalized to be time dependent;
• θ(t) or m(t): a deterministic time-varying drift component chosen so the model is consistent with the initial term structure;
• W_t: Brownian motion under the pricing measure.

Why Hull–White is not just Vasicek with nicer notation

A constant-parameter Vasicek model generates only the term structures consistent with its fixed parameters. Hull and White’s extension introduces a time-dependent drift so the model can be fitted to the market term structure observed at time zero.

That is a structural distinction:

the initial yield curve is not predicted by the model; it is imposed as a calibration boundary.

The stochastic model then specifies how rates can evolve away from that starting surface.

Stage 1: bootstrap the market discount curve

Before Hull–White calibration begins, the model needs discount factors or zero rates across maturities. Those are normally built from market instruments through a curve-construction process.

The calibration should therefore inherit the same:

• valuation date;
• business-day calendars;
• day-count conventions;
• instrument definitions;
• collateral and discounting assumptions;
• interpolation method;
• market-data timestamps.

A perfectly coded Hull–White engine can still be wrong if its input curve is inconsistent with the products it later prices.

Stage 2: fit the deterministic drift to today’s curve

For constant a and σ, the model has an analytic relationship between the initial forward curve and the required time-dependent drift.

One common representation uses the instantaneous forward rate:

f(0,t) = −∂ ln P(0,t) / ∂t.

The drift function then includes the slope of that forward curve, the mean-reversion term and a volatility correction. The exact formula depends on notation, but the purpose is invariant:

P_model(0,T) = P_market(0,T)

for the maturities represented by the calibrated initial curve.

This produces the first hard verification test: feed today’s market curve into the model, rebuild zero-coupon bond prices, and confirm they are recovered to numerical tolerance.

The affine bond-price structure

Hull–White remains tractable because zero-coupon bond prices can be written in exponential-affine form:

P(t,T) = A(t,T) exp[−B(t,T)r_t].

For constant mean reversion:

B(t,T) = [1 − e^−a(T−t)] / a.

If a is small, B(t,T) approaches approximately T−t. If a is large, distant cash flows become less sensitive to today’s short rate because shocks decay more rapidly.

For example, if a = 5% per year and T−t = 5 years:

B ≈ (1 − e^−0.25)/0.05 ≈ 4.42.

The bond-price coefficient A(t,T) carries the initial curve and volatility adjustment needed to preserve no-arbitrage consistency.

Stage 3: calibrate mean reversion and volatility to options

Matching the initial yield curve does not determine the stochastic dynamics uniquely. Many combinations of mean reversion and volatility can fit today’s bond prices.

Option prices provide additional information because they depend on the distribution of future rates.

A calibration set might contain:

• ATM European swaptions;
• caps or floors;
• selected expiries and tenors;
• normal, lognormal or shifted-lognormal market quotes translated consistently into prices.

The model parameters are then chosen to minimize a loss function such as:

L(a,σ) = Σ_i w_i[V_model,i(a,σ) − V_market,i]².

The weights w_i matter. Equal price errors, equal volatility errors and vega-weighted errors are not the same objective.

Calibration is an inverse problem

Forward pricing asks:

parameters → price.

Calibration reverses the arrow:

market prices → parameters.

Inverse problems can be ill-conditioned. Two materially different parameter pairs may generate very similar prices across a limited set of instruments.

This is the identifiability problem.

Why mean reversion and volatility can trade off

A higher volatility tends to widen future rate distributions and increase option values. Faster mean reversion dampens the persistence of shocks, especially over longer horizons.

Over a narrow calibration set, increasing σ while also increasing a can sometimes preserve similar option values. The objective function can contain a long shallow valley rather than one sharp minimum.

A solver may report convergence even though the parameters are weakly identified.

Therefore a production calibration should not only report the minimum loss. It should inspect:

• parameter sensitivity;
• multiple starting points;
• local curvature of the objective;
• stability across neighboring valuation dates;
• out-of-sample instruments.

A conceptual calibration example

Suppose the engine starts with:

• a = 0.03;
• σ = 0.008;
• a set of ten ATM swaption prices.

The model reprices short-expiry swaptions too cheaply and long-expiry swaptions too richly.

The optimizer tries a new parameter pair. Each iteration performs the same loop:

1. derive the drift function consistent with the initial curve;
2. price the calibration instruments under the candidate parameters;
3. compute residuals;
4. update parameters subject to bounds;
5. stop only when convergence and residual conditions are satisfied.

A converged pair is not automatically accepted. It must pass model-risk diagnostics.

Tree construction for early-exercise products

European options can often be handled analytically or semi-analytically in one-factor Hull–White. Early-exercise products require state-by-state decisions over time.

A recombining trinomial tree is a common numerical representation. At each time step, the short-rate state can move up, stay near the middle, or move down. Transition probabilities are chosen so the discrete tree matches the local drift and variance of the continuous process while preserving a manageable recombining structure.

Backward induction then works from maturity toward today:

continuation value = discounted expected next-node value.

At an exercise date:

option value = max(exercise value, continuation value)

for the relevant holder or issuer decision.

This is the computational bridge to callable-bond OAS algorithms.

Inputs and outputs

A production-shaped Hull–White calibration can require:

• discount curve;
• projection curve where relevant;
• valuation date and calendars;
• swaption or cap/floor market quotes;
• quote convention;
• instrument schedules;
• parameter bounds;
• initial guesses;
• objective-function definition;
• calibration weights;
• solver tolerance;
• tree or PDE grid settings for numerical products.

Outputs can include:

• mean-reversion parameter;
• volatility parameter or term structure;
• time-dependent drift function;
• model option prices;
• market-model residuals;
• curve-repricing residuals;
• solver convergence status;
• parameter-sensitivity diagnostics;
• tree-convergence diagnostics.

Evidence polarity

Evidence for confidence includes exact recovery of the input discount curve to numerical tolerance, small and pattern-free option-pricing residuals, stable parameters under reasonable starting guesses, convergence of tree prices as the time grid is refined, and independent agreement between analytic and numerical prices for products where both are available.

Evidence against confidence includes parameters stuck on bounds, large residual clusters by expiry or tenor, violent day-to-day parameter jumps without comparable market changes, materially different solutions from different starting points, failure to recover the initial curve, or numerical prices that do not converge as the grid is refined.

Counterexample: exact curve fit does not mean correct dynamics

The drift function is designed to fit the initial curve. That is a boundary condition, not proof that the model describes future yield-curve evolution well.

A model can reproduce every initial discount factor and still generate unrealistic volatility or correlation patterns.

Counterexample: one factor cannot create independent curve shocks

With one Brownian driver, instantaneous shocks across maturities are driven by the same underlying random factor. Their loadings differ by maturity, but the model cannot independently shock the short end upward while simultaneously giving the long end an unrelated stochastic shock.

Real yield curves exhibit level, slope and curvature variation. This is why yield-curve PCA often finds several economically meaningful factors and why two-factor models exist.

Counterexample: Gaussian rates can go negative

The one-factor Hull–White process is Gaussian. Negative short rates are therefore mathematically possible.

That was once treated as an obvious defect; subsequent negative-rate regimes showed that negative rates themselves are not impossible. The deeper issue is whether the Gaussian tails and symmetric shock structure remain appropriate for the product and market being modelled.

Counterexample: fitting ATM options does not fit the smile

A constant-volatility one-factor Gaussian model can fit a selected set of ATM option prices but cannot generally reproduce a full strike-dependent volatility smile.

If smile-sensitive products matter, the calibration residuals across strikes become a falsifier. More flexible local-volatility, stochastic-volatility, SABR or market-model approaches may be needed.

Counterexample: numerical differentiation can amplify curve noise

The drift function depends on information from the initial forward curve, which in turn can require differentiating an interpolated discount curve.

Small irregularities in market quotes or interpolation can become larger irregularities after differentiation. A calibration can therefore inherit noise from curve construction.

Diagnostics should compare alternative interpolation schemes and inspect the smoothness of implied forwards and θ(t).

Failure diagnostics

• Curve-repricing test: reconstruct P(0,T) and compare with input discount factors.
• Instrument-repricing test: compare each model option price with its calibration target.
• Residual-shape test: plot errors by expiry, tenor and strike.
• Multi-start test: rerun calibration from several plausible parameter guesses.
• Boundary test: flag solutions at parameter limits.
• Objective-surface test: inspect whether many parameter pairs produce nearly identical losses.
• Analytic-versus-tree test: compare products that have both valuation routes.
• Grid-convergence test: halve the time step and confirm prices stabilize.
• Day-over-day stability test: separate genuine market movement from calibration noise.
• Out-of-sample test: price liquid options excluded from the calibration set.

Alternatives answer different questions

Black–Derman–Toy uses a different tree-based short-rate structure.

CIR introduces square-root volatility and, under suitable conditions, a non-negative short rate.

G2++ / two-factor Gaussian models add a second stochastic factor so curve movements are less constrained.

HJM models forward-rate dynamics directly and imposes a no-arbitrage drift restriction. See the HJM article in this batch.

SABR focuses on volatility-smile dynamics rather than short-rate mean reversion. See SABR calibration algorithms.

How this connects to the surrounding knowledge estate

The initial curve can be represented or studied through Nelson–Siegel–Svensson curve fitting or Kalman-filter curve dynamics. Swaption market targets connect to Black-76 and Bachelier swaption algorithms. Short-rate dynamics also enter SOFR-futures convexity adjustments. Early exercise connects directly to callable-bond OAS and Bermudan-style tree valuation.

What would falsify confidence?

Confidence should be withdrawn if the model cannot recover the initial curve, if calibration residuals show systematic structure, if materially different parameter sets fit equally well but produce materially different downstream values, if tree prices fail numerical convergence, if out-of-sample option prices are persistently poor, or if the product depends on smile or multi-factor dynamics the model cannot represent.

Verification and update triggers

Preserve the market-data snapshot, curve build, calibration instruments, quote conventions, objective function, weights, parameter bounds, starting guesses, solver version and residuals. Recalibrate or revalidate after material curve shifts, volatility-regime changes, quote-convention changes, benchmark transitions, curve-construction changes, solver-library upgrades, persistent out-of-sample mispricing or material downstream hedge errors.

Primary and high-quality references

• John Hull and Alan White, Pricing Interest-Rate-Derivative Securities, Review of Financial Studies, 1990.
• John Hull and Alan White, One-Factor Interest-Rate Models and the Valuation of Interest-Rate Derivative Securities, Journal of Financial and Quantitative Analysis, 1993.
• John Hull and Alan White, Numerical Procedures for Implementing Term Structure Models I: Single-Factor Models, Journal of Derivatives, 1994 bibliographic record.
• MathWorks, Hull–White calibration using caps, for a documented implementation of calibration mechanics.
• Federal Reserve Board and OCC, SR 11-7 — Guidance on Model Risk Management, for independent validation, monitoring and limitations.

Educational boundary: This article explains stochastic interest-rate calibration and numerical diagnostics. It does not recommend a trading strategy or determine whether any instrument should be bought or sold.