Forecasting a Curve by Forecasting Three Numbers: Diebold and Li

Nelson and Siegel's formula was a snapshot tool. Give it today's bond prices and it draws today's curve. Ask it what the curve will look like next year and it has nothing to say, because it is not a model of dynamics at all.

Francis Diebold and Canlin Li's contribution is to notice that this is an entirely fixable problem, and that fixing it produces a forecasting model that beats most of the sophisticated competition.

The problem: forecasting a curve is forecasting a hundred things at once

Suppose you want to predict the yield curve twelve months from now.

The naive framing is horrible. A yield curve is, in principle, a yield at every maturity: the 3-month, the 6-month, the 1-year, the 2-year, all the way out to thirty. That is dozens of numbers you would have to forecast, and they are all wildly correlated with each other, and you have limited history. Any statistical model with that many outputs will drown in its own parameters and fit noise.

Worse, if you forecast each maturity independently you can easily produce a "curve" that is not a curve at all: a jagged, implausible thing where the 5-year yield sits above the 6-year and below the 7-year for no reason. Nothing in the setup forces the answer to look like a yield curve.

And the benchmark you must beat is brutally hard: the random walk, which simply says "next year's curve will be this year's curve." Yields are extremely persistent. Predicting no change is a very strong forecast, and the term structure literature was littered with sophisticated models that could not beat it, a point Duffee had made devastatingly a few years earlier.

The key idea via analogy: forecast the sliders, not the sound

Diebold and Li's insight is a two-step move, and it is elegant enough that it seems obvious once stated.

Step one: compress. Take Nelson and Siegel's formula and reinterpret it. Nelson and Siegel had three components that mix to draw a curve. Diebold and Li point out what those components actually are: they are level, slope and curvature. The first shifts the whole curve, the second tilts it, the third humps it. These are exactly the three factors Litterman and Scheinkman had found by pure statistics.

They also make one crucial simplifying move: they fix the decay parameter at a sensible constant. This is the awkward nonlinear knob that makes Nelson-Siegel estimation fiddly. Nail it down, and the whole thing becomes a simple linear regression. You can now fit the curve for any given month in a fraction of a second.

So: run that regression on every month of history. Now instead of a mountain of yields, you have three time series: the level over time, the slope over time, the curvature over time. Decades of yield curves have been compressed into three wiggly lines.

Step two: forecast the three lines. Fit a simple time series model to each one, an autoregression, which is essentially "next month's level is a bit like this month's level, plus a shock." That is all. Three modest little forecasts.

Then reassemble. Feed your forecasted level, slope and curvature back into the Nelson-Siegel formula, and it draws you a complete, smooth, plausible yield curve for twelve months from now.

The analogy is music. Rather than trying to predict the sound wave at every instant, you predict where the three equaliser sliders will be, and let the equaliser rebuild the sound. Because a real yield curve is essentially a mixture of level, slope and curvature, forecasting those three sliders is forecasting the curve, and you have reduced an impossible problem to three easy ones.

And the reassembled curve is guaranteed to be well behaved, because Nelson-Siegel can only draw well-behaved curves. You get coherence for free.

Why it mattered

• It beat the random walk, which almost nothing did. Diebold and Li report that at longer horizons, particularly twelve months ahead, their forecasts outperform a range of competitors across maturities. Given how many sophisticated no-arbitrage models had been humiliated by the random walk benchmark, a simple, transparent model that actually wins was a significant result.
• It made the yield curve a time series object. Before this, curve fitting and curve forecasting were separate literatures. Diebold and Li fused them, and the "dynamic Nelson-Siegel" model became a standard tool at central banks, in asset management and in academic work.
• It is transparent and cheap. No latent factors, no nightmare likelihood surfaces, no optimiser that lands somewhere different each run. A regression and three autoregressions. You can build it in an afternoon and explain it to a risk committee.
• It fused two independent discoveries. Nelson-Siegel's three components (found by asking what shapes curves take) and Litterman-Scheinkman's three factors (found by asking what actually moves bond returns) turn out to be the same objects. Diebold and Li made that identification explicit and then put it to work.
• It launched a family. Dynamic Nelson-Siegel with time-varying loadings, with macro variables, with stochastic volatility, and, importantly, the arbitrage-free Nelson-Siegel models (Christensen, Diebold and Rudebusch) that repair the model's biggest theoretical flaw.

The honest limitations

• It is not arbitrage-free. This is the standing criticism, and it is legitimate. The model has no financial theory in it. The fitted and forecast curves can, in principle, imply prices that a trader could arbitrage against. For forecasting yields, this may not matter much. For pricing derivatives off the model, it is disqualifying.
• Fixing the decay parameter is a fudge. It makes everything easy, but it is a choice, and different choices move where the curvature factor peaks, which changes the fitted factors, which changes the forecasts. The convenience is bought with an arbitrary assumption.
• Two-step estimation is statistically inefficient. Fit the factors, then forecast the factors, treating the first-step estimates as if they were observed data with no uncertainty. They are not. The standard errors are understated and the procedure is not the efficient one; later versions estimate everything in a single state-space framework.
• Beating the random walk is a low bar, and even then it is only sometimes. The advantage is real but modest, and it is much stronger at long horizons than short ones. At one month ahead, predicting no change is still extremely hard to improve on.
• It cannot predict turning points. The three factors are modelled as persistent autoregressions, which means the model is fundamentally saying "things will keep drifting the way they have been." It will miss regime changes, policy shocks and inversions, which are exactly the moments a yield curve forecast would be most valuable.
• Estimated on a particular era. The original work uses US data from a period of positive rates and conventional policy. In a world of zero rates, quantitative easing and yields pinned near a floor, a linear, Gaussian autoregression on the level factor is a questionable description of what is going on.

The one-line takeaway

Diebold and Li compressed the yield curve into three moving numbers, level, slope and curvature, forecast those three with simple time series models, and reassembled the curve, producing a forecasting method simple enough to explain in a paragraph and good enough to beat the random walk.