The Dataset That Ate the Literature: Gurkaynak, Sack and Wright

Not every important paper contains a theorem. Some contain a spreadsheet, and change the field more than the theorems do.

Gurkaynak, Sack and Wright, three economists at the Federal Reserve Board, took the Fed's internal estimate of the US Treasury yield curve, documented exactly how it was built, extended it back to 1961, and published the whole thing, updated daily, for free, forever.

If you have read almost any empirical bond paper written since, you have read a paper that used this data.

The problem: everyone was building their own curve, badly and differently

Here is a thing that is not obvious until you try it: there is no such thing as "the" ten-year Treasury yield.

What exists is a set of individual Treasury bonds, each with its own coupon, its own maturity date, its own quirks. None of them is a clean ten-year zero-coupon bond, which is what a researcher actually wants when they say "the ten-year yield." The bond closest to ten years might mature in 9 years and 7 months, pay a 6 percent coupon, and be trading rich because it is the newest issue and everyone wants it for repo.

So every researcher who wanted a yield curve had to build one: gather the bond prices, decide which bonds to exclude (bills? callable bonds? the on-the-run issue?), choose a fitting method, choose parameters, and produce a curve. Everyone made different choices. Nobody documented them fully.

The consequences were quietly corrosive. Two papers testing the same hypothesis on the same period could disagree, and no one could tell whether the disagreement was economics or data construction. Results were hard to replicate. And building a decent curve was a substantial engineering project, which meant a researcher with a good idea about bonds had to spend months on plumbing before testing it.

Meanwhile the Federal Reserve had a good curve, built carefully, used internally to inform policy. It just was not public.

The key idea via analogy: publish the ruler

Gurkaynak, Sack and Wright's contribution is not a new method. They use the Svensson extension of Nelson-Siegel: the parsimonious formula with a level, a slope and two humps that bends smoothly through the observed bond prices without chasing noise. That method was well established.

What they did was three things, none of them glamorous, all of them enormously valuable.

They documented it completely. Which bonds are included and which are thrown out, and why. (They exclude Treasury bills, which behave oddly relative to the rest of the curve, and they drop the newest issues, which trade at a liquidity premium that would otherwise contaminate the curve.) Every judgment call is written down. Anyone can reproduce the construction from the paper.

They extended it back to 1961. Decades of consistently constructed daily yield curve data, built the same way throughout, so a comparison between 1975 and 2005 is a comparison of economics rather than of methodology.

They gave it away and kept updating it. The fitted Svensson parameters and the resulting zero-coupon yields and forward rates are published on the Fed's website and refreshed. Not a snapshot. A living dataset.

The analogy is the metre bar. There was a time when the metre was a physical bar of platinum kept in Paris, and its existence mattered enormously: not because that particular bar was special, but because everyone measuring in metres was measuring against the same thing. A shared standard eliminates a whole category of disagreement. Gurkaynak, Sack and Wright published the ruler.

Why it mattered

• It became the default data for bond research. The GSW dataset is the empirical bedrock of a huge fraction of the modern term structure literature. When a paper says "we use zero-coupon yields from Gurkaynak, Sack and Wright," which they constantly do, this is what they mean.
• It made results comparable and replicable. Disagreements between papers are now about models and hypotheses rather than about whose curve-fitting code was used. That is a substantial upgrade in the scientific quality of the field.
• It democratised the subject. A graduate student with an idea can now test it against sixty years of clean daily yield curve data in an afternoon, with no funding and no data vendor contract. That lowers the barrier to entry enormously, and good ideas come from unexpected places.
• It gave the market and the public a window into the Fed's own measurement. The curve the Fed looks at is the curve you can look at. In an institution whose entire job is communication and credibility, that transparency has real value.
• The companion TIPS curve. The same team (with others) later did the same job for inflation-protected securities, giving researchers a consistent real yield curve and a breakeven inflation series. Same principle, same payoff.

The honest limitations

The dataset is a triumph, and using it uncritically is a mistake. Its authors say so themselves.

• It is a fit, not an observation. These are not prices. They are the output of a smooth formula bent through a scatter of bond prices. The 7-year zero-coupon yield in the dataset is an estimate, and if the fit is poor on a given day, it is a poor estimate. Treating it as raw data is a category error that a lot of users make.
• The smoothing erases exactly what some people want. Individual bonds deviate from the fitted curve for real reasons: on-the-run liquidity premia, repo specialness, index effects, flight-to-quality demand for particular issues. Those deviations are noise to a macroeconomist and signal to a relative-value trader. The Svensson fit deliberately averages them away. Anyone studying Treasury market microstructure or dislocations needs the individual bonds, not this.
• The short end is unreliable. Because Treasury bills are excluded and the parsimonious formula struggles with the sharp, angular features that policy expectations create at the front of the curve, the fitted yields at very short maturities are the least trustworthy part of the dataset. This is well documented and routinely ignored.
• The long end is extrapolated. For much of the sample there simply were no thirty-year Treasuries (the thirty-year was discontinued for several years in the early 2000s). Yields quoted at the very long end during those periods are the formula extrapolating, not the market speaking.
• The parameters are unstable. The Svensson parameters can jump around from day to day even when the curve barely moves, because different combinations produce near-identical curves. The fitted yields are far more reliable than the fitted parameters, and it is a mistake to interpret the parameters economically.
• The method carries its own assumptions. Svensson's formula can only draw a certain family of shapes. If the true curve on some day has structure the formula cannot represent, the dataset will not contain it, and you will never know from the dataset alone.

The one-line takeaway

Gurkaynak, Sack and Wright published the Federal Reserve's own daily US Treasury yield curve back to 1961, fully documented and free, and by giving everyone the same ruler they made a generation of bond research comparable, replicable and possible at all.