1Describe a rates model whose state is the whole curve of instantaneous forward rates, and link it to bond prices and the short rate.
2Derive the HJM drift condition under ℚ with Itô's lemma, and say what the modeller chooses and what no arbitrage forces.
3Run the martingale test on a one-factor HJM model on the SONIA curve, with the drift and without it, and explain the gap.
4Explain why HJM with a general vol is not Markov in the short rate, and what that means for trees and simulation.
Why it matters
Module 5.3 built rates models from one number, the short rate, and derived the curve from it. Fitting today's curve took a time-dependent drift θ(t), worked out backwards (L5.3.05: The Hull-White model). Heath, Jarrow and Morton turned this round in 1992: start from the curve itself, the one thing the market tells you for certain, and let every point on it move. The price of that freedom is one condition, and it is the subject of this page: once you choose how the forwards move, no arbitrage leaves you no choice about where they drift. Every market model in this module is built on it.
Real curveIllustrative volsSterling SONIA, 31 December 2025
The curve is the level's one rates market: the SONIA OIS curve built earlier in this level from quotes derived from the Bank of England's fitted OIS curve at 31 December 2025 (L5.2.02: Bootstrapping a curve). Every volatility is illustrative: the level's caplet and swaption vols were set at plausible levels because no free source of sterling vols was found, so every parameter fitted to them is illustrative too. This page sets one vol by hand: a humped vol for the forwards, 75bp at the short end, peaking at 99.4bp at 1.52 years and falling towards 60bp. It is shaped like Module 5.2's stripped caplet vols but not fitted to them illustrative. For comparison it uses Module 5.3's house Hull-White, a = 1.6038% and σ = 92.17bp, fitted to the illustrative co-terminal swaptions (L5.3.09: Calibration to caps and swaptions) illustrative. Every simulated figure is given with its standard error, its number of paths and its seed, so it can be reproduced. Figures from earlier modules stand as booked; where this page reprices one, it shows both and says why they differ.
From the short rate to the whole curve
The instantaneous forward rateInstantaneous forward rate f(t, T): the rate, agreed at t, for borrowing over the instant just after T. Bond prices are the forwards added up: P(t, T) = exp(−∫ f(t, u) du from t to T), and the short rate is r(t) = f(t, t). f(t, T) is the rate agreed at t for borrowing over the instant just after T. Add the forwards up and you have every bond price, P(t, T) = exp(−∫ f(t, u) du), and the nearest forward is the short rate, r(t) = f(t, t). On the SONIA curve at 31 December 2025 the short rate is 3.725%; the forward dips to 3.353% at one year (the market expected cuts) and rises to 4.716% at ten years and 4.904% at twenty derived.
A short-rate model makes r the state and derives this curve from it. The HJM frameworkHJM framework: Heath, Jarrow and Morton's (1992) way to build a rates model. Start from today's forward curve, choose how each forward moves (its volatility), and let no arbitrage fix the drift. makes the curve itself the state. Each forward moves with its own shock: df(t, T) = α(t, T) dt + σ(t, T) dW. The volatility functionVolatility function σ(t, T): how much the forward for date T moves at time t. In a Gaussian HJM model it is a deterministic function, often of the time to maturity T − t alone. σ(t, T) says how much the forward for date T moves; it is the modeller's choice. Today's curve is the starting point, so the model fits it by construction: no θ(t) to solve for.
The drift is not a choice
Under the risk-neutral measure every bond counted in the bank account must be a martingale (L5.2.05: Risk-neutral pricing and numeraires). A bond is the exponential of minus a sum of forwards, so Itô's lemma (M-F03: Itô's lemma) gives its drift a convexity term, ½(∫σ)². The forwards must drift up by just enough to cancel it. That fixes the drift of every forward from the vols alone: the HJM drift conditionHJM drift condition: under the risk-neutral measure the drift of every forward is fixed by the vols, α(t, T) = σ(t, T)∫ σ(t, u) du from t to T. With several factors the products are summed., α(t, T) = σ(t, T)∫σ(t, u) du. The formal version derives it. Two vols on the SONIA curve illustrative:
The hump (illustrative)House Hull-White, σe^(−aτ)
The drift is a few basis points a year: small, but it compounds over years and across every forward a bond spans. Simulation adds one subtlety. The grid here has quarterly forwards and quarterly steps (h = 0.25), and Euler on the continuous α would leave the grid's bonds slightly off. The discrete driftDiscrete drift: the drift that makes every bond on a simulation grid an exact martingale, step by step. It tends to the continuous HJM drift as the step shrinks. μ makes every grid bond an exact martingale; it is a little below α at this step, and the two meet as h shrinks (Glasserman and Zhao 2000 note the adjustment is widely known in practice):
The hump on the quarterly grid at its first step (vols read at the middle of the step and the period); α(0, T) at the period's middle.
Forward, first step
Discrete μ, bp a year
Continuous α, bp a year
[1.00, 1.25]
0.775
0.996
[2.00, 2.25]
1.753
1.955
[5.00, 5.25]
3.397
3.510
[10.00, 10.25]
4.835
4.933
The martingale test
L5.2.05: Risk-neutral pricing and numeraires checked a short-rate model by simulation: the bond at T counted in the bank account must average today's P(0, U). The same martingale testMartingale test: a check that a simulated price counted in the numeraire averages to today's value. For a bond under ℚ: E[P(T, U)/B(T)] = P(0, U). on the HJM grid, with T = 5 and U = 10, 20,000 paths in antithetic pairs, seed 55,001 (M-D06: Monte Carlo simulation):
20,000 paths (10,000 antithetic pairs), seed 55,001, quarterly steps. The two-factor vols are Module 5.3's G2++ fit with the fast factor's mean reversion held at 1 (L5.3.10: Two-factor models).
E[P(5, 10)/B(5)]
Estimate ± se
z from P(0, 10)
P(0, 10) on the curve
0.670590
Hump, with the HJM drift
0.670663 ± 0.000096
+0.76
Hump, drift switched off
0.677213 ± 0.000097
+68.4
Two factors (G2++ with a held at 1, as HJM vols), with the drift
0.670641 ± 0.000097
+0.53
With the drift the test passes for one factor and for two. Without it the estimate is 68 standard errors high, and not by chance: in a Gaussian HJM model the drift is a number fixed in advance, so switching it off multiplies every path's discounted bond by the same factor, ev/2 = 1.009766, where v is the variance of ln(P(5, 10)/B(5)). The worked example computes it.
Is the short rate enough?
In Hull-White, knowing r(5) tells you the whole curve at 5 years (L5.3.05: The Hull-White model): the model is MarkovMarkov: a model whose future depends on the past only through a small current state. A short-rate model is Markov in r if knowing r(t) is enough to know the whole curve at t. in r, which is what lets a recombining tree price it (L5.3.06: Trinomial trees). Under a general HJM vol it need not be. Regress ln P(5, 10) on r(5) across 4,000 simulated paths derived:
4,000 paths, seed 55,001. An R² of 1 means the bond is an exact function of the regressors.
Regression of ln P(5, 10)
R²
Hull-White's vol, on r(5)
1.000000
The hump, on r(5)
0.995521
The hump, on the forwards at 5, 7 and 10 years
1.000000
Under the hump, two paths with the same short rate can have different curves. A short-rate tree has no room for that: a model that is not Markov in a few states needs a non-recombining treeNon-recombining tree: a tree in which an up move then a down move does not land on the same node as down then up, so the number of nodes doubles at each step. A model that is not Markov in a few states needs one., whose nodes double at each step, or simulation. This particular hump, (a + bτ)e−cτ + d, moves the curve in only three shapes of maturity, e−cT, Te−cT and a constant, so three forwards pin the curve down. A general vol needs all of it. That is why HJM models, and the market models built on them, are priced by Monte Carlo (M-D06: Monte Carlo simulation). L5.5.02: Hull-White as an HJM model shows the vol for which one number is enough.
Sort itChosen, forced or read?
In an HJM model under ℚ, sort each ingredient by where it comes from.
Today's forward curve f(0, T)
The vol function σ(t, T)
The drift α(t, T) of each forward under ℚ
How many Brownian factors drive the curve
Today's short rate r(0)
The discrete drift on a simulation grid
The formal version
The HJM drift condition
Under the risk-neutral measure ℚ, with K independent Brownian motions (Heath, Jarrow and Morton 1992; the grid version as in Glasserman and Zhao 2000):
P(t,T)=exp(−∫tTf(t,u)du),r(t)=f(t,t),B(t)=exp(∫0tr(s)ds)df(t,T)=α(t,T)dt+k=1∑Kσk(t,T)dWk(t)B(t)P(t,T) is a Q-martingale for every T⟺α(t,T)=k=1∑Kσk(t,T)∫tTσk(t,u)dugrid: fl(ti+1)=fl(ti)+μlh+hk∑σk,lZk,μl=2hk∑(Sk,l2−Sk,l−12),Sk,l=j=i+1∑lσk,jno drift: EQ[B(T)P(T,U)]=P(0,U)ev/2,v=Var(lnB(T)P(T,U))
f(t, T) is the instantaneous forward for T seen at t; σ_k(t, T) the deterministic normal vol of factor k; W_k independent Brownian motions under ℚ; B the bank account. On the grid, f_l is the forward over [t_l, t_l + h), t_i = ih, Z_k independent standard normals, and σ_k,l is read at the middle of the step and of the period. The hump is σ(τ) = (a + bτ)e−cτ + d with a = 0.0015, b = 0.005, c = 0.55, d = 0.006 (Rebonato's abcd form, used here in normal-vol units). The last line holds for any Gaussian HJM model: without the drift, the log of the discounted bond is normal with mean ln P(0, U).
Try it
Pick a vol and run the test. Each change simulates the grid twice on the same random numbers, once with the HJM drift and once without, using at most 2,000 paths so that it stays fast. The headline run above used 20,000.
Try itTest the drift on the curve
Volatility of f(t, T)
Test date T
Bond maturity U − T
Paths (antithetic pairs)
Solid: the vol σ(0, T) in bp a year. Dashed: the drift α(0, T) the vol forces, drawn to the same height at 9.0bp a year.
P(0, 10) on the curve0.670590
With the drift0.671401 ± 0.000467
z from P(0, U)+1.74
Without the drift0.677958 ± 0.000471
z from P(0, U)+15.6
Predicted: P(0, U)e^(v/2)0.677139
z from the prediction+1.74
R², ln P(T, U) on r(T)0.996351
R², on three forwards1.000000
With the drift the estimate sits within a couple of standard errors of the curve's P(0, U), whatever the vol. Without it every path is the same path times ev/2, so it misses by the predicted amount: try doubling the vol (the gap grows four-fold) or a 10-year bond. Then compare the R²s: Hull-White and Ho-Lee give exactly 1 on the short rate alone; the hump needs more than r, and three forwards are enough.
Worked example
One forward's drift, and what it buys
The hump on the SONIA curve, quarterly grid
Step 1 of 5
1
The vol ten years out
τσ(τ)=T−t=10.125=(a+bτ)e−cτ+d=(0.0015+0.005×10.125)e−0.55×10.125+0.006=61.99bp a year
The drift of 4.93bp a year looks negligible beside a vol of 62bp. It is not: it is what turns a convexity of 0.98% in the bond's average into zero.
Regulatory lens · looking ahead
In the UK the PRA's supervisory statement SS1/23, Model risk management principles for banks, took effect on 17 May 2024 (an updated version took effect on 23 April 2026). It applies to banks with internal-model approval for capital, and its principles cover all types of models used to inform business decisions, including those used for financial reporting: pricing models such as these. Two of its five principles are model development, implementation and use, and independent model validation. A test with a known answer, run with and without the piece being tested, as on this page, is one kind of evidence a developer or validator can show. Model risk in rate models is L5.5.10: Model risk in rate models; model validation as a discipline is M6.7: Model risk, validation and machine learning.
Spark Atoms
✦Spark AtomNo market price of risk+
Equilibrium short-rate models such as Vasicek and CIR start in the real world and need a market price of risk to reach ℚ (L5.4.02: Vasicek and CIR as affine models). Heath, Jarrow and Morton's abstract makes the opposite point: starting from today's forward curve and a family of possible movements, their valuation formulas do not depend explicitly on market prices of risk. The curve already has them in it.
✦Spark AtomWhen is the short rate Markov?+
Andrew Carverhill asked exactly this in a short paper in Mathematical Finance in 1994. His answer, for HJM models with deterministic vols: only when the vol structure meets a constraint, and he noted it matters for building efficient numerical methods. This page's R² of 0.9955 is the hump failing it; L5.5.02: Hull-White as an HJM model is a vol that meets it.
Key takeaways
HJM makes the whole forward curve the state: P(t, T) = exp(−∫f), r = f(t, t), and today's curve is fitted by construction.
The modeller chooses the vols σ(t, T); under ℚ the drift is forced: α(t, T) = σ(t, T)∫σ(t, u)du, from Itô's lemma and the martingale condition.
On a grid the discrete drift makes every bond an exact martingale; on the SONIA curve it is 4.835bp a year against α's 4.933 at ten years, and they meet as the step shrinks.
The martingale test passes with the drift (0.670663 ± 0.000096 against 0.670590) and fails without it by 68 standard errors, exactly the factor e^(v/2) = 1.0098.
A general HJM model is not Markov in r: R² of 0.9955 for the hump. It needs the whole curve (here three forwards suffice), so it is priced by simulation.
Reaction
Check what stuck
Question 1 · objective 2
In Ho-Lee every forward has the same vol, σ(t, T) = σ = 90bp a year. What is the HJM drift of the forward five years ahead, α(t, t + 5), in bp a year? (Two decimals.)
bp a year
Question 2 · objective 2
In an HJM model under the risk-neutral measure, what does the modeller choose and what is forced?
Question 3 · objective 3
Without the drift the test gives E[P(5, 10)/B(5)] = 0.677213, above P(0, 10) = 0.670590. Why above?
Question 4 · objective 4
Across 4,000 simulated paths, ln P(5, 10) regressed on the short rate r(5) has R² = 0.9955 under the humped vol, and exactly 1 under Hull-White's vol. What does that tell you?
Question 5 · objective 3
With the drift, 20,000 paths give 0.670663 ± 0.000096 against P(0, 10) = 0.670590. How many standard errors is the estimate from the target? (Two decimals.)
se
Glossary
Instantaneous forward rate
Instantaneous forward rate f(t, T): the rate, agreed at t, for borrowing over the instant just after T. Bond prices are the forwards added up: P(t, T) = exp(−∫ f(t, u) du from t to T), and the short rate is r(t) = f(t, t).
HJM framework
HJM framework: Heath, Jarrow and Morton's (1992) way to build a rates model. Start from today's forward curve, choose how each forward moves (its volatility), and let no arbitrage fix the drift.
Volatility function
Volatility function σ(t, T): how much the forward for date T moves at time t. In a Gaussian HJM model it is a deterministic function, often of the time to maturity T − t alone.
HJM drift condition
HJM drift condition: under the risk-neutral measure the drift of every forward is fixed by the vols, α(t, T) = σ(t, T)∫ σ(t, u) du from t to T. With several factors the products are summed.
Discrete drift
Discrete drift: the drift that makes every bond on a simulation grid an exact martingale, step by step. It tends to the continuous HJM drift as the step shrinks.
Martingale test
Martingale test: a check that a simulated price counted in the numeraire averages to today's value. For a bond under ℚ: E[P(T, U)/B(T)] = P(0, U).
Markov
Markov: a model whose future depends on the past only through a small current state. A short-rate model is Markov in r if knowing r(t) is enough to know the whole curve at t.
Non-recombining tree
Non-recombining tree: a tree in which an up move then a down move does not land on the same node as down then up, so the number of nodes doubles at each step. A model that is not Markov in a few states needs one.
D. Brigo and F. Mercurio, Interest Rate Models: Theory and Practice, 2nd ed., Springer, 2006, ch. 5: the HJM framework and when its short rate is Markov. Advanced.
L. B. G. Andersen and V. V. Piterbarg, Interest Rate Modeling, Volume II: Term Structure Models, Atlantic Financial Press, 2010: separable HJM models and their implementation. Advanced.
Origins
D. Heath, R. Jarrow and A. Morton, “Bond pricing and the term structure of interest rates: a new methodology for contingent claims valuation”, Econometrica 60(1), 77–105, 1992.
P. Glasserman and X. Zhao, “Arbitrage-free discretization of lognormal forward Libor and swap rate models”, Finance and Stochastics 4, 35–68, 2000.
A. Carverhill, “When is the short rate Markovian?”, Mathematical Finance 4(4), 305–312, 1994.
Prudential Regulation Authority, SS1/23 Model risk management principles for banks, May 2023 (effective 17 May 2024; updated version effective 23 April 2026).
Bank of England, UK OIS yield curve data (fitted curve at 31 December 2025), through Module 5.2's SONIA quotes.
Watch this lesson
Video version
Prefer to watch? The same lesson as a 14-minute video will sit here, made from this page's script.