Beaten by a Coin Flip: Duffee and the Failure of Affine Forecasts

There is a particular kind of paper that a field needs and dreads: the one that takes the discipline's best machinery, points it at a simple task, and reports that it fails.

Gregory Duffee's 2002 paper is that paper for term structure modelling. He took the affine models that two decades of research had produced, the ones with the elegant closed forms, the ones fitted to data in dozens of published papers, and asked them to forecast where yields would be in a year.

They lost to a random walk. That is, they were beaten by the strategy of predicting that yields will be exactly where they are today.

The problem: a benchmark that should be easy to beat

Forecasting yields with a random walk means saying: "the ten-year Treasury is at 4 percent, so my forecast for the ten-year Treasury in twelve months is 4 percent." It uses no model, no theory, no data beyond the current quote. It is the laziest possible forecast.

It is also, in many financial markets, surprisingly hard to beat, which is exactly why it makes a good benchmark. But term structure models were supposed to have an edge here. They were built on the observation that interest rates mean-revert: when rates are unusually high they tend to fall, when unusually low they tend to rise. That is a genuine forecasting signal, it is the core assumption of every short rate model from Vasicek onward, and a model that knows about it should be able to outpredict a forecast that assumes nothing happens.

Duffee showed they could not. The standard affine models, the ones everyone was using, produced worse out-of-sample forecasts of Treasury yields than the random walk. Not marginally worse in some corner of the data. Worse.

The key idea via analogy: a thermostat wired to the wrong sensor

Duffee did not stop at the embarrassment. He diagnosed it, and the diagnosis is the paper's real contribution.

In these models, the compensation investors demand for bearing interest rate risk (the term premium) is not a free-floating quantity. It is hard-wired to be a multiple of the amount of risk. More volatility means proportionally more compensation. Less volatility, proportionally less. The two are chained together by construction, and this is exactly the "market price of risk is proportional to the risk" structure inherited all the way from Vasicek.

That sounds reasonable, even virtuous. It is a disaster for forecasting, for one reason: it means the risk premium can only move when volatility moves.

Here is the analogy. Imagine a thermostat whose only input is the humidity. It is a real sensor, humidity is a real thing, and sometimes humidity and temperature move together, so the thermostat is not useless. But you have wired the heating system so that it can only respond to humidity. On a cold, dry day, the thermostat sees nothing wrong and does nothing. The house freezes.

That is what the affine models were doing. The term premium in the real world swings around with the business cycle: investors demand a lot of extra yield for holding long bonds in some periods and very little in others, and those swings do not line up neatly with how volatile rates happen to be. (Fama and Bliss had shown the premium is predictable and cyclical; Cochrane and Piazzesi would later show it in even sharper relief.) But the models could not represent this. If volatility was steady, the model insisted the risk premium was steady too. Its only sensor was the wrong one.

Since forecasting future yields is essentially the job of separating "where the market expects rates to go" from "how much extra yield investors are demanding," a model that cannot get the risk premium right cannot get the forecast right. The model's rigidity, the very thing that made it tractable, was throttling it.

The fix: essentially affine models

Duffee's solution is to unchain the two. He introduces a broader class he calls essentially affine models, in which the compensation for risk is allowed to vary with the state of the world independently of how volatile that world currently is.

Rewire the thermostat so that it can respond to temperature as well as humidity. The heating can now come on during a cold, dry spell.

The crucial engineering constraint is that he does this without giving up tractability. The essentially affine class keeps everything that made affine models attractive: bond prices still come out in closed form, the model is still estimable, all the machinery still works. The only thing that changes is that the risk premium gets extra freedom to move.

And that freedom is worth a great deal. With it, the models start producing forecasts that improve on the random walk, particularly for the longer maturities where the term premium does most of its damage.

Why it mattered

• It is the standard indictment of the affine literature. Anyone claiming a term structure model is any good now has to confront Duffee's random-walk benchmark. "Beats the random walk out of sample" became a basic hygiene requirement.
• Essentially affine became the default. Almost every serious term structure model built after 2002, Ang and Piazzesi, Cochrane and Piazzesi's model work, the Gaussian models used inside central banks to decompose yields, uses the essentially affine specification of risk prices. It is not an alternative; it is the standard.
• It separated two jobs a model does, and showed they conflict. Fitting today's cross-section of yields is one task. Forecasting how they evolve is another. A model can be excellent at the first and useless at the second, and Duffee showed the affine class's structure had been quietly optimising for the first at the expense of the second.
• It connected the modelling literature to the predictability literature. Fama and Bliss, Campbell and Shiller had spent years documenting that bond risk premia vary over time and are forecastable. The models had, in effect, been ignoring them. Duffee's fix is what let the two literatures talk to each other.
• It made term premium estimation credible. Central banks routinely publish estimates of the term premium embedded in ten-year yields, used to judge whether long rates are low because the market expects weak growth or because investors are unusually happy to hold duration. Those estimates come from essentially affine models, and they would be much less trustworthy without Duffee's correction.

The honest limitations

• The extra freedom is freedom to overfit. Letting the risk premium respond to any state variable you like gives the model a lot of parameters and a lot of ways to fit historical wiggles that do not repeat. Later work has found term premium estimates from these models to be uncomfortably sensitive to specification and sample. The estimates are useful, but the error bars around them are wide, wider than the confident-looking central bank charts sometimes suggest.
• It improves on the random walk without being good. Beating a lazy benchmark is not the same as forecasting well. Yields remain extremely hard to predict, and the improvement is real but modest.
• It is still affine. Duffee relaxes the risk premium specification, not the underlying linear structure. Everything Dai and Singleton pointed out about the trade-off between stochastic volatility and correlation, everything nonlinear about how rates behave near zero, is untouched.
• The estimation problem persists. These models are notoriously hard to fit reliably, with flat and multi-peaked likelihood surfaces. Adding more risk-price parameters made that worse, not better, which is part of why Joslin, Singleton and Zhu had to write a paper on how to estimate them at all.
• The sample is a different world. The paper is fitted to US Treasury data from an era of positive rates and conventional policy. Whether the diagnosis and the fix behave the same way in a decade of zero rates and quantitative easing is a fair question the paper cannot answer.

The one-line takeaway

Duffee showed that the finest interest rate models in finance forecast yields worse than assuming nothing changes, traced it to a hidden wire forcing the risk premium to move only when volatility moves, and cut that wire, producing the essentially affine models that the field has used ever since.