Squeezing More Power from a Weak Test: Elliott, Rothenberg and Stock

The dirty secret of unit root testing is that the tests barely work.

Not in the sense that they are wrong. In the sense that they are weak. Hand an augmented Dickey-Fuller test a series that genuinely does revert to its mean, but slowly, and it will very often shrug and tell you it cannot rule out a random walk. Give it a few hundred observations of a series with a true autoregressive coefficient of 0.95, which is a real, tradeable, mean-reverting process, and the test will fail to detect the mean reversion a large fraction of the time.

This matters enormously in practice, because a slowly mean-reverting spread is exactly what a pairs trader is looking for, and the standard test will frequently tell you it is not there.

In 1996, Graham Elliott, Thomas Rothenberg and James Stock asked a sharper question than "can we build a better test?" They asked: how good could any test possibly be? And then they built one that comes close to that ceiling.

The problem: an accidental waste of information

Every unit root test has to deal with a nuisance. Before you can ask whether a series wanders, you have to deal with the fact that the series probably has a mean, and possibly a trend. Real data does not hover around zero. It hovers around something, and it may drift.

The standard approach is to remove that first. Fit a mean, or fit a trend, subtract it, and test what is left. This is called detrending, and it seems entirely innocent.

It is not innocent. The way you estimate the trend affects how much power your test has left. Ordinary least squares fits the trend to the data as if the data were nicely behaved and stationary. But if the series is close to a random walk, OLS is not the right way to estimate the trend, and the trend you subtract is subtly wrong in a way that eats your ability to detect mean reversion.

You have spent some of your statistical budget on the nuisance parameter, and there was not much budget to begin with.

The key idea via analogy: knowing what to listen for

Imagine you are trying to detect a very faint signal in a noisy radio band. A general-purpose detector, one designed to work reasonably well for any kind of signal, will have to spread its sensitivity across many possibilities and will be mediocre at all of them.

But suppose you have a strong prior about what you are listening for: you expect the signal to look approximately like this. Now you can build a detector tuned specifically to that shape. It will be dramatically more sensitive to the signal you expect, at the cost of being worse at detecting signals you did not expect.

Elliott, Rothenberg and Stock did exactly this. They said: in practice, the interesting alternative to a random walk is not a wildly mean-reverting series. It is a series that is very close to a random walk but not quite. That is where the real difficulty lies, and that is where you want your test to be sensitive.

So they tuned the detrending step. Instead of removing the trend with ordinary least squares (which implicitly assumes stationary, well-behaved data), they remove it using a procedure calibrated to a series that is nearly a random walk. This is called GLS detrending, and it is why the resulting test is commonly known as DF-GLS.

The change is small in appearance and large in effect. By spending your nuisance-parameter budget more efficiently, you have more statistical power left over for the question you actually care about.

The deeper contribution: knowing the ceiling

The part of the paper that impresses econometricians most is not the test itself. It is the power envelope.

Elliott, Rothenberg and Stock derived, from first principles, an upper bound on how much power any unit root test could achieve. Not any test they could think of. Any test, ever. It is a theoretical ceiling, a statement about the fundamental information content of the problem.

This is enormously clarifying, in two directions.

First, it tells you how much room for improvement exists. They showed that their test's power curve sits close to the envelope, and that the standard Dickey-Fuller test sits noticeably below it. There was real, quantifiable performance being left on the table, and now you know how much.

Second, and more soberingly, it tells you when to stop hoping. The envelope is not high. Even the theoretically optimal test cannot reliably distinguish a random walk from a slowly-reverting series in a modest sample. The weakness of unit root tests is not a failure of ingenuity. It is a property of the problem. No future paper will fix it, because the information is not in the data.

That second lesson is arguably worth more to a practitioner than the test itself.

Why it mattered

• It gave a meaningful power improvement on the exact cases that matter. DF-GLS detects mean reversion in near-unit-root series noticeably more often than the standard ADF test. For anyone testing whether a spread is tradeable, that is not academic. It is the difference between finding a strategy and discarding it.
• It became a standard part of the toolkit. DF-GLS (sometimes called the ERS test) is a standard option in econometrics packages and is routinely reported alongside ADF and Phillips-Perron.
• It established a benchmark. The power envelope gave the whole subfield a yardstick. Any new unit root test now has to be measured against it, which mostly means demonstrating that it is not worth publishing.
• It taught a general lesson about nuisance parameters. The idea that how you handle the boring part of your model determines how much power you have left for the interesting part generalises far beyond unit roots. It is a good thing to have internalised.

The honest limitations

• A better test is still a weak test. This must be said clearly. DF-GLS is more powerful than ADF, and it is still not powerful. In a sample of a few hundred observations, it will still frequently fail to detect genuine slow mean reversion. Upgrading your test is not a substitute for having enough data.
• The lag-length problem does not go away. Like ADF, DF-GLS requires you to choose how many lagged terms to include, and the answer still matters, still affects the conclusion, and still has no clean solution.
• The tuning is a bet, and you can lose it. The test is optimised for the near-random-walk region. If your series is strongly mean-reverting, the tuning buys you nothing, and the test can perform slightly worse than the standard one. In exchange for extra power where you need it, you gave up a little where you did not. That is usually a good trade, but it is a trade.
• It has size distortions in finite samples. Like most of its relatives, DF-GLS can reject too often when the underlying process has certain moving-average structures, which financial data frequently does.
• Structural breaks still ruin it. The entire family shares this blind spot. A single permanent level shift will make a stationary series look like a wanderer to DF-GLS just as reliably as to any other test in the family.
• It does not answer the question you actually have. A trader does not want to know "can I reject the unit root hypothesis at the 5% level." A trader wants to know "how fast does this spread revert, and is that fast enough to beat my costs." A more powerful hypothesis test does not tell you that. Estimating the speed of reversion, with an honest confidence interval, is usually the more useful exercise than testing whether the speed is exactly zero.

The one-line takeaway

Elliott, Rothenberg and Stock showed that the standard unit root test wastes statistical power on the boring business of removing a trend, built a version that wastes less, and, more valuably, derived the hard ceiling on how good any such test could ever be, thereby proving that the notorious weakness of unit root testing is a fact about the world rather than a gap in the literature.