The Six-Page Paper Everyone Cites: Newey-West Standard Errors

This is a six-page paper. It contains no dramatic economic finding, no new theory of markets, no counterintuitive result. It is a technical patch for a plumbing problem.

It is also one of the most cited papers in the history of economics, and if you have ever run a predictive regression on overlapping returns, you have almost certainly used it, whether you knew it or not.

The problem: the coefficient is fine, the t-statistic is fiction

Every regression gives you two things: an estimate, and a measure of how uncertain that estimate is. The estimate is the coefficient. The uncertainty is the standard error, and it is the standard error that generates your t-statistic, your p-value, and ultimately your decision about whether you have found something real.

Ordinary least squares computes that standard error using a formula that rests on two assumptions:

All the errors are the same size. (No heteroskedasticity.)
The errors are unrelated to each other over time. (No autocorrelation.)

In financial and economic data, both assumptions are routinely false, and the second one is catastrophically false in one extremely common setting: overlapping observations.

Here is the classic case. You want to know whether the dividend yield predicts the next twelve months of stock returns. You have monthly data. So in January you regress the next twelve months' return on January's yield. In February you regress the next twelve months' return on February's yield. And so on.

Look at what you just did. January's forecast period and February's forecast period share eleven months. They are almost the same observation. The regression treats them as two independent data points, but they are nearly the same data point wearing a different hat.

Do this across thirty years of monthly data and you appear to have 360 observations. In terms of genuinely independent information, you have something closer to thirty. Your regression believes it has twelve times more evidence than it actually has. So it computes a standard error that is far too small, a t-statistic that is far too large, and hands you a p-value that is, in a word, fake.

This is the machine that produced a large chunk of the return-predictability literature.

The key idea via analogy: count your independent witnesses

Suppose you are investigating a claim and you interview a hundred witnesses. If all hundred are independent people who each saw the event separately, that is powerful evidence.

But suppose ninety of them were standing in a group, all repeating what one person told them. You do not have a hundred pieces of evidence. You have roughly eleven. If you compute your confidence as though all hundred were independent, you will be dramatically overconfident.

Newey-West is a way of counting how much genuine independent information you actually have.

The mechanics are more intuitive than the name suggests. To measure how much the errors are related over time, you look at the correlation between an error and the error one period later, and two periods later, and three, and so on. Add up those relationships and you get a picture of how much redundancy is baked into your data. Then you inflate the standard error accordingly. More redundancy means a bigger inflation, which means a smaller t-statistic and a more honest verdict.

The coefficient itself does not change at all. Your estimate of the relationship stays exactly the same. What changes is your confidence in it, and it always changes in the direction of humility.

The technical trick that made it work

There was a problem with the obvious version of this idea. If you just add up the correlations at each lag with equal weight, the resulting variance estimate can come out negative. A negative variance is not merely wrong, it is nonsensical, and it makes the whole procedure unusable in practice.

Newey and West's contribution, and the reason their names are on the tool rather than someone else's, was to fix this with a weighting scheme. Instead of counting each lag equally, you count nearby lags heavily and progressively discount distant ones, with the weights declining linearly to zero. The weighting is chosen so that the resulting estimate is guaranteed to be non-negative, always. That is what the "positive semi-definite" in the title means, and it is the entire content of the paper's contribution.

An estimator that sometimes returns nonsense is a research topic. An estimator that always returns something usable is a button in every statistics package. That is the difference this weighting made.

Why it mattered

It made a huge class of financial studies possible to run honestly. Long-horizon return predictability, overlapping-window regressions, term-structure studies, macro predictability: every one of these has the overlapping-observations problem. Without a fix, none of their t-statistics mean anything.
It became a default. In empirical finance, reporting Newey-West standard errors is not a sophisticated flourish. It is table stakes. A regression on overlapping returns without them will not survive a referee, or a serious risk committee.
It quietly killed a lot of "findings." Numerous celebrated predictive relationships lost most of their statistical significance the moment their standard errors were corrected. That is exactly what a good diagnostic tool is supposed to do.
It generalises. The same machinery is used inside GMM estimation, inside the Diebold-Mariano forecast comparison test, and inside Fama-MacBeth regressions. It is a piece of infrastructure that many other tools quietly depend on.

The honest limitations

You have to choose the bandwidth, and the answer depends on your choice. How many lags should you include? Too few and you have not corrected enough, leaving you overconfident. Too many and you add noise, making the test erratic. There are rules of thumb, and Andrews later provided a principled automatic procedure, but in practice a great many researchers pick a round number and move on. The choice can matter enough to flip a conclusion, which is uncomfortable for something so widely used.
It is an asymptotic fix, and it under-corrects in small samples. With limited data and strong serial correlation, Newey-West standard errors are still too small. They make you more honest, but not honest enough. In the overlapping-return setting specifically, several studies have shown the correction leaves substantial over-rejection on the table.
It fixes your inference, not your data. If your genuinely independent sample size is thirty, no correction conjures more information. Newey-West tells you the truth about how little you know. It cannot make you know more.
It does not save you from a misspecified model. Serially correlated errors are frequently a symptom. They can mean you omitted a variable, or got the functional form wrong, or are running a spurious regression on non-stationary data. Slapping a robust standard error on the problem silences the alarm without addressing the fire. This is probably the most common misuse: treating autocorrelated residuals as a nuisance to be patched rather than as evidence that something is wrong with the model.

The one-line takeaway

Newey and West wrote the short, unglamorous paper that stops a regression from mistaking redundant, overlapping data for genuine independent evidence, and their weighting trick is what turned a good idea into a button that always works, which is why their names sit under nearly every table in empirical finance.