Model the Whole Curve at Once: Heath, Jarrow and Morton

Every interest rate model before 1992 had the same architecture. Pick a short rate. Say how it moves. Derive everything else from it.

Heath, Jarrow and Morton asked a different question, and it turned out to be the right one: why start with the short rate at all? The short rate is a single point on a curve. If you want to model the curve, model the curve.

The paper that resulted is the most general statement of arbitrage-free interest rate modelling anyone has written, and it contains a result so clean that it reorganised the entire field.

The problem: the short rate is a strange place to start

The short rate models, Vasicek, CIR, Ho-Lee, Hull-White, Black-Derman-Toy, all pull the same move: assume something about the rate for the next instant, then let the whole yield curve emerge as a consequence.

This has two costs.

It is indirect. You do not observe the short rate's future path. You observe the yield curve. Yet these models make you specify the thing you cannot see in order to derive the thing you can. Every one of them then needs a calibration step to force its derived curve to match the observed one.

It is limiting. Because everything descends from one number, one factor short rate models produce curves whose points are all perfectly correlated. The curve shifts as a rigid body. Adding richness means adding factors, and adding factors to a short rate model gets awkward fast. There was no natural way to say "I want the short end to be very volatile and the long end to be calm, and I want them to be about 60 percent correlated." You had to hope some parameter choice happened to imply it.

The key idea via analogy: describe the shape of the wave, not the height of one buoy

HJM's move is to take the entire forward rate curve as the object being modelled. The forward curve is the market's schedule of rates for borrowing over each future instant: the rate for lending from year 3 to year 3-and-a-bit, from year 7 to year 7-and-a-bit, and so on. It is directly extractable from bond prices, and it contains exactly the same information as the yield curve.

Instead of tracking one buoy bobbing on the sea (the short rate) and trying to infer the whole wave, HJM describe how the entire wave deforms. Today's shape is given (there it is, on the screen). What you must specify is how each point on it wiggles: how volatile the 2-year forward is, how volatile the 10-year forward is, and how their wiggles are correlated with each other.

Then comes the theorem, and this is the heart of the paper.

Every random process has two parts: a random part (the volatility, how much it jiggles) and a systematic part (the drift, where it is heading on average). HJM prove that in a world with no arbitrage, you do not get to choose the drift of the forward curve. It is completely determined by the volatilities you already chose.

This is startling. It says the two things are not independent at all. Once you have told the model how much each forward rate shakes, the requirement that no one can construct a risk-free money machine locks in exactly where all those forward rates must drift. There is a formula. It falls out of the no-arbitrage condition, and it is now known simply as the HJM drift condition.

Think of a set of interlocking gears. You are free to choose the size of the teeth (the volatilities). But once you have, you cannot also choose which way each gear turns: the mesh determines that. Try to force a gear the other way and the mechanism jams, and a jammed mechanism, in finance, means someone is standing there printing money.

There is a second gift. Because the whole construction is built on the observed forward curve, today's yield curve fits by construction. There is no calibration step to force the model to agree with the market on bond prices; it agrees automatically. And the market prices of risk, that awkward free parameter that haunted Vasicek and had to be assumed, simply drop out of the valuation formulas. They are irrelevant for pricing derivatives.

Why it mattered

• It is the general theory, and it showed everything else was a special case. Vasicek, Ho-Lee, Hull-White: every one of them turns out to be an HJM model with a particular choice of forward rate volatility. HJM did not add another model to the pile. They drew the map that contains the pile.
• The drift condition is the deepest result in the literature. "No arbitrage determines the drift from the volatility" is the fixed income equivalent of the risk-neutral pricing insight in Black-Scholes, and it is what every subsequent framework, including the LIBOR market model, is built on.
• It made multi-factor modelling natural. Want the curve to twist as well as shift? Specify a second volatility function with a different shape across maturities. The framework accommodates any number of factors without contortion, which is exactly what a real curve, with its level, slope and curvature movements, requires.
• It freed you from the short rate. Modelling what you observe rather than what you must infer is a better default, and after HJM, "which forward curve volatilities do I believe in?" became the honest form of the modelling question.
• It set up the market models. Brace, Gatarek and Musiela's LIBOR market model, the one that finally made model prices consistent with the Black formula traders were already using, is built directly inside the HJM framework.

The honest limitations

HJM is a framework, not a model, and its generality is both its virtue and its cost.

• It usually is not Markovian, which makes it slow. This is the practical headache. In a short rate model, the current short rate is a complete summary of the world: to price anything, you only need to know today's number. In a general HJM model, the future of the curve can depend on the entire path it took to get here. You cannot summarise the state in a couple of numbers, which means you cannot use a recombining tree. You are pushed toward Monte Carlo simulation, which is far more expensive and, importantly, awkward for products with early exercise (Bermudan swaptions). Only for special choices of the volatility function does the model collapse back to something Markovian and tree-friendly, and those special cases are, unsurprisingly, the old short rate models.
• It tells you nothing about which volatilities to choose. The theory says "pick your volatility structure and I will hand you the drift." It does not say which structure is right. That is an empirical question the framework is silent on, and it is the question that actually determines whether your prices are any good.
• Instantaneous forward rates are not real. HJM models the rate for an infinitesimally short borrowing period starting at each future date. That object does not trade. Markets trade three-month LIBOR, six-month rates, swap rates: discrete, tradeable things. This mismatch between the model's fundamental object and the market's quoted objects is exactly the gap the LIBOR market model was built to close.
• Forward rates can go negative. In Gaussian versions, which are the tractable ones, nothing prevents it. (As with Vasicek, history has been kinder to this "flaw" than anyone expected.) Lognormal versions avoid negativity but can explode to infinity in finite time, a result that HJM themselves helped establish.
• Calibration is genuinely hard. With a rich, multi-factor volatility structure you have a lot of freedom and a lot of ways to overfit. The framework's flexibility hands you enough rope to build a model that matches every quoted price and behaves nonsensically anywhere off the grid.

The one-line takeaway

Heath, Jarrow and Morton modelled the entire forward curve at once and proved that once you have chosen how volatile each forward rate is, the no-arbitrage condition dictates every drift in the system, a result that turned all previous interest rate models into special cases of one framework.