The Hull-White model is a mean-reverting Gaussian short-rate model fitted to the current yield curve for valuing interest-rate derivatives.
The Hull-White model is a short-rate model that represents the instantaneous interest rate as a mean-reverting Gaussian process with a time-dependent drift chosen to fit the current yield curve. The one-factor version is widely used as a tractable framework for valuing interest-rate options, callable bonds, swaptions, caps, floors, and other rate-sensitive contracts.
The model describes risk-neutral rate dynamics for valuation. Fitting today’s term structure does not mean it predicts the future path of policy rates or bond yields.
A common risk-neutral specification is:
where:
| Term | Interpretation |
|---|---|
| (r_t) | Instantaneous short rate at time (t) |
| (\theta(t)) | Time-dependent drift function selected to fit the initial term structure |
| (a) | Mean-reversion speed |
| (\sigma) | Instantaneous short-rate volatility under the simplified constant-volatility form |
| (W_t) | Brownian motion under the pricing measure |
Some texts use different symbols or write the model as a mean-reverting factor plus a deterministic shift. Those forms can be mathematically equivalent under their definitions. The symbol (\theta(t)) in the equation above is not simply a constant long-run average rate.
A common implementation writes:
with:
The stochastic factor (x_t) is an Ornstein-Uhlenbeck process, while the deterministic shift (\phi(t)) is selected to reproduce the initial term structure under the chosen conventions. This form separates today’s curve from future random shocks.
Parameterizations are not interchangeable by symbol alone. One implementation’s (\theta(t)), (\phi(t)), volatility function, initial factor value, and curve inputs must be mapped algebraically before prices or parameters are compared.
The parameter (a) controls how quickly a short-rate shock decays in the model.
The two parameters interact when fitting option prices. Different combinations can produce similar prices for selected instruments while generating different exposures for other maturities or exercise structures.
For constant (a>0), the remaining effect of a factor shock decays approximately as (e^{-a\Delta t}). Its model half-life is:
For example, (a=0.10) per year gives a half-life of about 6.93 years, while (a=0.50) gives about 1.39 years. These values describe risk-neutral factor persistence under the model, not forecasts that observed yields will return to a policy target on those dates.
The conditional variance of the constant-volatility factor over a horizon (\Delta t) is:
As (a) approaches zero, the factor behaves more like a non-mean-reverting Gaussian rate process over the relevant horizon. Large (a) dampens distant effects, but (a) and (\sigma) must be interpreted together.
The time-dependent drift is selected so the model reproduces the initial prices of zero-coupon bonds from the chosen discount curve. Under a common constant-(a) form, model zero-coupon bond prices can be written:
with:
The function (A(t,T)) depends on the initial curve and model parameters. This affine structure contributes to the model’s analytical tractability.
Under the constant-volatility one-factor form, the magnitude of a zero-coupon bond’s instantaneous volatility is proportional to:
For a bond close to maturity, (B(t,T)) approaches zero. For a distant maturity, mean reversion limits the loading toward (1/a). This maturity pattern is one reason the mean-reversion parameter affects long-expiry and long-tenor option values.
The initial curve must still be built carefully. Instrument selection, interpolation, bootstrapping, day counts, collateral or discounting conventions, and market-data timestamps can change the fitted curve and derivative value.
Hull-White calibration commonly has two distinct layers:
| Layer | Main inputs | What is matched |
|---|---|---|
| Deterministic curve fit | Discount factors or zero-coupon prices | Today’s selected term structure |
| Stochastic-parameter fit | Cap, floor, or swaption prices or volatilities | Selected option evidence |
An exact curve fit does not establish that caplet or swaption prices are matched. Conversely, a low option-calibration error can depend on a curve whose instruments, interpolation, collateral basis, or timestamp differ from the trade’s valuation framework.
Modern rate valuation can use separate discount and projection curves. The one-factor short rate does not automatically model basis between them. Implementations may treat some curves deterministically, introduce additional factors, or apply basis adjustments. The chosen architecture must be documented rather than inferred from the Hull-White label.
| Feature | Hull-White one-factor | Vasicek model | Black-style market model |
|---|---|---|---|
| State variable | Instantaneous short rate | Instantaneous short rate | Forward rate, swap rate, or futures-related rate under selected formulation |
| Mean reversion | Yes | Yes | Not represented through one universal short-rate process |
| Initial curve fit | Time-dependent drift can fit the selected curve | Basic constant-parameter form generally does not fit an arbitrary current curve exactly | Current forward or swap rate is an explicit market input |
| Rate distribution | Gaussian in the basic form | Gaussian | Common lognormal or normal quote model, depending on convention |
| Negative rates | Possible | Possible | Depends on model and displacement or normal convention |
| Typical use | Consistent valuation across rate-dependent cash flows and exercise dates | Foundational term-structure modeling | Quoting and valuing selected caps, floors, or swaptions |
No row makes one model universally superior. The choice depends on product payoff, exercise rights, market quote conventions, curve framework, and required risk measures.
Every modeled bond and rate depends on the same Brownian shock in the basic one-factor framework. This creates a parsimonious, recombining state representation, but it also restricts how the yield curve can move.
The model can still be useful when the omitted risks are measured separately, the product is not materially sensitive to them, or reserves and alternative-model comparisons address the limitation.
Assume a European receiver swaption gives its holder the right at time (T) to enter a swap receiving a fixed rate (K=3.50%). At expiry, suppose:
For a simplified physically settled receiver swaption, the exercise value at (T) is:
Using the illustrative values:
The receiver right has value because receiving 3.50% is favorable relative to the 2.90% market swap rate in this simplified state. If the market swap rate were above the strike, the holder could let the option expire.
Before expiry, the Hull-White model values the discounted payoff across possible future rate states, using the fitted curve and calibrated rate dynamics. The $270,000 calculation is an expiry payoff illustration, not a model price and not a forecast.
Implementation commonly separates two tasks:
A price-error calibration can be represented schematically as:
The parameter vector (\mathbf{p}) can contain (a), constant or time-dependent volatility parameters, and any approved extensions. The weight (w_i) determines which instruments and errors matter most.
Calibration choices include:
A calibration can fit liquid at-the-money instruments while mispricing strikes or exercise patterns not included in the objective. Residuals should be reviewed across the relevant expiry-tenor grid.
| Calibration evidence | Question to answer |
|---|---|
| Price and quote residuals | Where does the model miss by expiry, tenor, and strike? |
| Bid-ask comparison | Is the model value outside executable market evidence? |
| Starting-value reruns | Does the optimizer converge to the same parameter region? |
| Parameter stability | Do (a) and (\sigma) jump without a corresponding market change? |
| Out-of-sample instruments | Does the fit extend beyond the instruments used? |
| Hedge or P&L attribution | Do model sensitivities explain subsequent value changes? |
Different combinations of mean reversion and volatility can fit a limited instrument set similarly. Parameter stability and out-of-sample behavior should therefore be evaluated separately from the optimizer’s total error.
Interest-rate option markets can quote normal volatility, lognormal volatility, shifted-lognormal volatility, or premium. A calibration engine ultimately needs prices consistent with the curve, annuity, settlement, and quote convention.
Comparing a normal-volatility error directly with a lognormal-volatility error is not meaningful without conversion. A one-basis-point volatility difference also does not represent the same currency-value error for every instrument.
The product’s reference rate, fallback language, collateral currency, multi-curve framework, exercise settlement, and day-count conventions must be represented separately from the short-rate equation.
A recombining short-rate tree is a common implementation for products with early or repeated exercise. At each node, the engine:
Grid spacing and time steps affect accuracy. The tree should reproduce the input curve and calibration instruments within documented tolerances. Monte Carlo methods may be used for exposure profiles or more complex multi-factor structures, although early exercise then requires an additional approximation method.
| Method | Typical use | Main implementation control |
|---|---|---|
| Analytical bond-option formula | European claims reducible to bond options | Curve, bond volatility, and formula convention |
| Recombining trinomial tree | Callable bonds and Bermudan exercise | State range, transition probabilities, and time grid |
| Finite-difference PDE | Early exercise and state-dependent features | Boundary conditions, grid, and convergence |
| Monte Carlo simulation | Exposure profiles and path-dependent cash flows | Time step, regression or exercise rule, and sampling error |
For an exercisable product, value at an exercise date compares immediate exercise with continuation:
The comparison is straightforward to state but model-dependent to calculate. Continuation value reflects future rates, cash flows, notice periods, settlement, accrued interest, call schedules, and any path-dependent contract terms.
A Bermudan swaption or callable bond can be especially sensitive to mean reversion because (a) affects how shocks propagate between exercise dates and across maturities. Two calibrations with similar European swaption errors can therefore produce different exercise boundaries and callable values.
Hull-White risk is broader than one parallel-rate delta. A useful report can include:
Frozen-parameter Greeks hold calibrated parameters fixed while curves move. Recalibrated scenarios allow parameters to change with market quotes. Those approaches answer different questions and can produce different P&L explanations.
For an actual valuation, use current curves, option quotes, trade terms, model documentation, calibration records, validation findings, and executable market evidence. This article is for financial education only and is not personalized investment, derivatives, valuation, legal, accounting, or tax advice.