Flipping the Question: the KPSS Test and the Burden of Proof

Statistical tests have a defendant, and the defendant is presumed innocent. In a unit root test, the defendant is the claim "this series is a random walk that never comes home." Dickey-Fuller, Phillips-Perron, and their relatives all start by assuming that claim is true, and then ask whether the data can produce enough evidence to convict.

In 1992, four authors pointed out something uncomfortable about this arrangement. If the evidence is weak, the defendant walks free. And in unit root testing, the evidence is almost always weak. So the "random walk" verdict was being handed out constantly, not because the data supported it, but because the data could rarely disprove it.

Their response was to build the test that puts the other claim in the dock. It is now known by their initials: KPSS.

The problem: absence of evidence was being read as evidence of absence

Think about what a failed conviction actually means. If a jury cannot prove guilt, that does not mean the defendant is innocent. It means the case was not strong enough. Those are very different statements, and treating them as the same thing is one of the oldest mistakes in statistics.

By the early 1990s, the economics literature had reached a striking consensus: nearly every macroeconomic series, and most asset prices, appeared to contain a unit root. GDP, employment, prices, exchange rates: all random walks, all with no tendency to return to any trend. This had enormous consequences. It implied recessions were permanent, that shocks never washed out, that the long run had no anchor.

KPSS asked the awkward question that gives the paper its subtitle: how sure are we? Their suspicion was that the profession was not sure at all. The tests everyone was using had famously low power. They struggled to distinguish a true random walk from a series that reverts to its mean but does so slowly. Faced with a hundred series, such a test would return "random walk" for nearly all of them regardless of the truth. The consensus might have been an artifact of a weak instrument, not a discovery about the world.

The key idea via analogy: run the trial twice, with the roles swapped

Suppose you want to know whether a coin is fair. You could put "the coin is fair" on trial and see if the flips can convict it. Or you could put "the coin is rigged" on trial and see if the flips can convict that. A confident conclusion requires the two trials to agree.

KPSS built the second trial. Their test starts from the opposite assumption: presume the series is stable, hovering around a level or around a trend, and see if the data can disprove it.

The mechanics are intuitive. Imagine a series that really is stable. Take its deviations from its average and start adding them up, cumulatively, as you walk through time. Because the deviations are just as often positive as negative, this running total should stay small and keep crossing back over zero. It is like a drunk who wobbles around a lamppost: the total distance he has drifted from the lamppost never grows very large.

Now imagine a series with a unit root. Its deviations do not cancel out, because there is no fixed centre for them to cancel around. The running total wanders further and further away, like a drunk who has abandoned the lamppost entirely and is now walking down the street.

KPSS is essentially a formal measurement of how big that running total gets. If it stays modest, the stability story survives. If it balloons, the stability story is rejected. The paper works out exactly how big is too big, and (in the spirit of Phillips-Perron) corrects the calculation for the fact that real-world shocks are correlated and unevenly sized.

The really useful part: running both tests together

The practical genius of KPSS is not the test on its own. It is what happens when you pair it with a conventional unit root test. There are four possible outcomes, and each tells you something different:

• Unit root rejected, stationarity not rejected. Both tests point the same way. You have real evidence the series is stable. This is the result a pairs trader wants to see on a spread.
• Unit root not rejected, stationarity rejected. Both tests point the same way again, toward a random walk. Good evidence the series genuinely wanders.
• Both rejected. Something is wrong with your setup. Perhaps there is a structural break, or the series is neither cleanly stationary nor cleanly a random walk. Time to look at the chart.
• Neither rejected. The most honest and most common outcome: your data simply is not informative enough to tell. The sample is too short or the signal too weak. This is the case that a unit root test alone would have quietly reported as "random walk," and it is exactly the confusion KPSS exists to expose.

That fourth cell is the paper's real gift. It gives researchers a way to say "I do not know" out loud, with statistical backing.

Why it mattered

• It punctured a consensus. When KPSS applied their test to the same famous macroeconomic dataset that had produced the unit-root orthodoxy, they found that for many series the stationarity hypothesis could not be rejected either. The evidence was genuinely ambiguous. The confident "everything is a random walk" story was much shakier than advertised.
• It changed lab practice. Reporting both an augmented Dickey-Fuller result and a KPSS result became standard hygiene in applied time-series work. It is now expected in serious papers, and it should be expected in serious backtests.
• It is a directly useful tool in trading research. A quant testing whether a spread mean-reverts is exposed to precisely the failure KPSS identified: a low-power test happily reporting "random walk" for a spread that is actually tradeable, or worse, failing to warn you that your allegedly stationary spread has no real anchor at all.

The honest limitations

• KPSS has its own power problem, in the mirror. It is quite capable of failing to reject stationarity for a series that really does have a unit root, especially in short samples. Swapping the null hypothesis does not create information that the data does not contain.
• It is sensitive to the bandwidth choice. Like Phillips-Perron, KPSS needs an estimate of how correlated the shocks are, and that requires choosing how far back to look. Different choices can flip the verdict, which is a genuinely uncomfortable property for a test that people treat as decisive.
• It over-rejects when the series is highly persistent but still stationary. A slowly mean-reverting series can look, to KPSS, like a wanderer. So the two tests can both reject, leaving you in the confusing third cell above.
• Structural breaks fool it too. A single permanent shift in level will make a perfectly stationary series look like it wandered, and KPSS will reject stationarity. Every test in this family shares this blind spot.

The one-line takeaway

KPSS flipped the burden of proof in unit root testing, and in doing so revealed that a great deal of what the profession "knew" about random walks in economic data was not evidence for random walks at all, just a failure to prove otherwise, which is a completely different and much less interesting thing.