The Oldest Alarm in Econometrics: Durbin-Watson

Run a regression in almost any statistics package and, somewhere in the output, you will find a number called the Durbin-Watson statistic. It sits there, quietly, in nearly every regression table ever printed.

Most people glance past it. That is a mistake, and it has been a costly one. The Durbin-Watson statistic is the oldest and most direct answer to a question that decides whether your regression means anything at all: do the errors have a memory?

James Durbin and Geoffrey Watson published this in two parts, in 1950 and 1951, and it is still doing its job seventy-five years later.

The problem: regression assumes each mistake is a fresh mistake

Ordinary least squares rests on an assumption that is easy to state and easy to forget: the errors are independent of each other. If your model over-predicted today, that tells you nothing about whether it will over-predict tomorrow. Each mistake is a fresh, independent draw.

In cross-sectional data, this is often reasonable. In time-series data, it is routinely and catastrophically false.

Here is what serial correlation looks like. Your model under-predicts in January, and February, and March, and April. It is not making independent mistakes. It is making the same mistake over and over, because it is missing something systematic. The residuals do not scatter randomly around zero. They wander, drift, and stay on one side for long stretches.

The consequences are two, and the second is much worse than the first.

First, your coefficient is inefficient. You could do better with a properly specified model.

Second, and fatally, your standard errors are wrong. Positively correlated errors mean your data contains far less independent information than your observation count suggests. So OLS computes a standard error that is far too small, a t-statistic that is far too big, and hands you a p-value that is fiction. You will find relationships that are not there, and you will be very confident about them.

The key idea via analogy: does the residual remember?

Durbin and Watson's test is beautifully direct. It asks a single question of the residuals:

Does each residual look like the one before it?

If your model is healthy, the residuals are random static. Knowing that yesterday's residual was positive tells you nothing about today's. Consecutive residuals will differ a lot, sign-flipping constantly, because they are independent draws.

If your model is sick with positive serial correlation, consecutive residuals will look similar. A positive residual is followed by another positive one. The residuals move slowly and smoothly, like a wandering line rather than a spray of dots.

So the statistic measures exactly this: how big are the differences between consecutive residuals, relative to how big the residuals themselves are?

• If consecutive residuals are very different from each other (healthy, no memory), the statistic comes out around 2.
• If consecutive residuals are nearly identical (strong positive memory, the disease), the differences are tiny and the statistic comes out near 0.
• If consecutive residuals alternate in sign (negative memory, rarer but real), the differences are huge and the statistic comes out near 4.

Around 2 is healthy. Near 0 is an alarm. Near 4 is a different alarm.

The genuinely hard part of the original papers, and the reason it took two of them, was working out the distribution of this statistic. It turns out that the critical values depend on the actual values of your regressors, which is awkward: it means there is no single universal table. Durbin and Watson's solution was to compute bounds: an upper and lower critical value, such that if your statistic falls below the lower bound you can definitely reject, above the upper bound you definitely cannot, and in between you are in an inconclusive zone. That inconclusive region is inelegant, and it is the price of a hard problem honestly solved.

Why it mattered, and the connection you must not miss

The Durbin-Watson statistic's greatest moment came twenty-four years after it was published, in the hands of Granger and Newbold.

They were investigating spurious regressions: the phenomenon where two completely unrelated wandering series produce a beautiful, highly significant regression that means absolutely nothing. And they noticed something. Spurious regressions have a signature.

They have a high R-squared and a very low Durbin-Watson statistic.

That combination is the fingerprint of a fake relationship. The high R-squared says "look how much I explain." The low Durbin-Watson says "my errors have enormous memory, which means I am not explaining anything, I am simply two trends drifting together."

Granger and Newbold's rule of thumb, still one of the fastest and best diagnostics in applied work, is brutally simple:

If R-squared is greater than the Durbin-Watson statistic, be very suspicious.

That single heuristic, which costs nothing to check, would have prevented a large fraction of the nonsense published in empirical economics in the 1960s and 1970s, and would prevent a large fraction of the nonsense in quant research today.

Why it matters in trading research

• Every price-on-price regression is at risk. Regressing one asset's price level on another's is exactly the Granger-Newbold trap. Check the Durbin-Watson. If it is near zero and your R-squared is 0.9, you have not found a relationship. You have found two things that both went up.
• Autocorrelated residuals mean leftover signal. If your model's errors are predictable, then by definition there is structure you failed to capture. For a forecaster, that is money on the table. For a risk model, that is a systematically underestimated risk.
• It is a warning that your standard errors are lying. A low Durbin-Watson is the signal that you need Newey-West standard errors, or that you need a better model. Preferably the latter.

The honest limitations

• It only detects first-order correlation. Durbin-Watson tests whether each residual resembles the one immediately before it. If your residuals have seasonal memory, correlated at lag 12 in monthly data but not at lag 1, the statistic will come out beautifully near 2 while your model is badly misspecified. This is a real and common blind spot. The Breusch-Godfrey test and the Ljung-Box test handle general lag structures and should be preferred when you suspect anything more complex.
• It is invalid when the regression contains a lagged dependent variable. This is a serious and frequently violated restriction. If your regression includes yesterday's value of the dependent variable as a predictor, which is extremely common in time-series work, the Durbin-Watson statistic is biased toward 2 and will systematically fail to detect serial correlation that is really there. It gives you a false all-clear precisely in the setting where you most need a warning. Durbin himself later developed an alternative (Durbin's h) for this case.
• The inconclusive zone is genuinely annoying. Landing between the bounds means the test has nothing to say, which happens more often than anyone would like.
• It diagnoses, it does not cure. A low statistic tells you something is wrong. It does not tell you whether the fix is an omitted variable, a wrong functional form, a needed lag, or the recognition that you are regressing two random walks on each other and should stop immediately.
• The reflex fix is often the wrong one. The historical response to a low Durbin-Watson was to apply a mechanical correction for autocorrelation and move on. Granger and Newbold's point was that this is often the worst possible response: the autocorrelation is not a nuisance to be patched, it is evidence that your entire model is wrong. Silencing the alarm is not the same as putting out the fire.

The one-line takeaway

Durbin and Watson gave every regression a smoke detector, a single number that tells you whether your model's errors have a memory, and the fact that a low value alongside a high R-squared is the unmistakable fingerprint of a spurious relationship makes it, seventy-five years on, one of the cheapest and most valuable checks in the entire quantitative toolkit.