Three Numbers Describe a Yield Curve: Nelson and Siegel

The yield curve is not something you can look up. It is something you have to construct.

What actually exists is a scattering of bond prices: a few Treasury bills here, some notes, some bonds, all with awkward maturities, some liquid and some not, some trading rich because they are the newest issue. From that scatter of dots you have to produce a smooth, continuous curve giving the yield at every maturity, because that is what you need to discount a cash flow arriving in 4.3 years, and that is what a central bank needs to read the market's expectations.

Drawing a smooth line through dots sounds trivial. It is not, and Nelson and Siegel's paper is the reason we are not still arguing about it.

The problem: too much flexibility is a disease

The naive approach is to fit a very flexible curve, a high-order polynomial, a spline with many knots, and let it pass through or near every observed point.

This goes wrong immediately, and it goes wrong in a specific and instructive way.

A flexible curve will fit your data beautifully and then do something insane between the points. It will develop wild oscillations. It will produce a forward rate curve, which is what you get when you look at the slope of the fitted curve, that snakes up and down implausibly, sometimes going negative, sometimes spiking, all in response to a single slightly odd bond price. Extrapolate it past your longest bond and it flies off to absurdity.

The problem is that the data is noisy, and a flexible curve dutifully fits the noise. Every quirk of liquidity, every stale quote, every bond that trades special in the repo market becomes a permanent feature of your yield curve. And because the forward curve is a derivative, small wiggles in the fitted yield curve become large wiggles in the forwards. Central bank economists trying to read policy expectations out of forward rates were reading their own fitting errors.

The key idea via analogy: three sliders on a graphic equaliser

Nelson and Siegel do the opposite of adding flexibility. They ask: what is the simplest formula that can still produce every shape a real yield curve actually takes?

Look at real yield curves and you find they come in a small number of shapes. Upward sloping (the usual). Downward sloping (inverted, a recession warning). Flat. And humped: rising for a while, peaking somewhere in the middle, then falling.

That is it. Real yield curves do not oscillate five times. They do not have three humps. They are smooth, and they belong to a small family.

So Nelson and Siegel write down a formula with just three building blocks, which combine to give you exactly that family. Think of a graphic equaliser with three sliders:

The level slider. A constant that shifts the entire curve up or down. Turn this up and every yield, from overnight to thirty years, rises together. This is the part of the curve that never fades: its influence is the same at every maturity, so it also determines where the curve settles in the very long run.
The slope slider. A component that is strongest at the short end and decays away as maturity grows. Turn this up and the short end moves while the long end stays put, so the curve tilts. This is what makes a curve upward-sloping or inverted.
The curvature slider. A component that is near zero at the very short end, rises to a peak somewhere in the middle, then decays back toward zero at the long end. This is the hump slider. It is what lets the curve bulge in the belly without disturbing either end.

Mix the three sliders in different proportions and you can draw every shape a real yield curve takes. There is one more parameter, a decay constant, that controls where the hump sits and how fast the slope component fades. That is the whole model: three sliders and a knob.

The magic is in what the formula cannot do. It cannot oscillate. It cannot spike. It cannot produce a forward curve with five wiggles in it, because there simply are not enough moving parts. The rigidity is the point. The curve is forced to be smooth and well behaved, so noise in the data gets averaged away instead of being carved into the answer. Extrapolate past your longest bond and the curve flattens out sensibly toward the level parameter, rather than exploding.

Nelson and Siegel build the formula on the forward rate curve rather than the yield curve, which is the right choice: forwards are where the pathologies show up, so that is where you impose the discipline.

Why it mattered

It is what central banks actually use. The Bank of England, the ECB, the Federal Reserve and many others fit Nelson-Siegel or its Svensson extension to their government bond markets, and publish the results. It became infrastructure.
The three sliders turned out to be the three factors. This is the beautiful part. A few years later, Litterman and Scheinkman ran a purely statistical analysis of bond returns and found that three factors explain almost all the variation, and those factors look like a level shift, a tilt, and a hump. Nelson and Siegel, working from a completely different direction (what shapes do curves take?), had written down the same three objects. Two independent roads to the same place is powerful evidence that the objects are real.
It made the curve a low-dimensional object. Instead of a thousand yields, a curve is three or four numbers. That compression is what makes it possible to model how curves evolve, which is exactly what Diebold and Li did in 2006 by treating the three sliders as time series and forecasting them.
Parsimony as a principle. The paper is a lasting argument that in a noisy world, a model that cannot fit the noise is often better than one that can.

The honest limitations

It is a description, not a theory. Nelson-Siegel is a curve-fitting formula. It contains no economics, no no-arbitrage condition, nothing about how rates evolve. In fact, in its raw form it is not arbitrage-free: nothing stops the fitted curve from implying prices that a clever trader could pick apart. For fitting today's curve this does not matter much. For pricing derivatives off it, it very much does, which is why the arbitrage-free Nelson-Siegel models exist as a separate, later strand.
It can be too rigid. With only one hump available, the model struggles with curves that have genuine complexity, a kink at the short end from a policy expectation plus a bulge further out. This is exactly the gap Svensson filled by adding a second hump.
The short end fits badly. The very front of the curve, where central bank policy expectations produce sharp, angular features, is often where Nelson-Siegel misses by the most. It smooths away features that are real.
The decay parameter is awkward. It enters the formula nonlinearly, so estimating it properly turns a simple linear fit into a fiddly nonlinear optimisation with multiple local optima. Most practitioners cheat by fixing it at a plausible value, which biases everything else and is one of the model's quiet embarrassments.
Fitted parameters are unstable. Refit the model day after day and the three sliders can jump around more than the curve itself does, because different combinations of sliders can produce almost identical curves. The parameters are not as economically meaningful as they look.
It hides the specials. By smoothing, the model deliberately ignores the fact that particular bonds trade rich or cheap for real reasons (on-the-run premium, repo specialness, index inclusion). A relative-value trader wants to see those deviations, and Nelson-Siegel's whole purpose is to erase them.

The one-line takeaway

Nelson and Siegel found that almost every real yield curve is just a mixture of a level, a tilt and a hump, and that a formula rigid enough to produce only those shapes fits the market better than a flexible one that also fits the noise.