Newey-West HAC Standard Errors

A researcher regresses 12-month forward stock returns on a valuation ratio, using monthly data where each observation's 12-month return window overlaps the next observation's window by 11 months. The ordinary regression's standard errors assume each residual is unrelated to the next, but here consecutive residuals share almost all of the same underlying return data — they're heavily correlated with each other by construction, not by coincidence. Treating them as independent makes the model wildly overconfident: the reported standard errors come out too small, t-statistics too large, and a relationship that's actually statistical noise can look highly significant. Newey-West standard errors correct for exactly this, adjusting for both autocorrelation across time and unequal error variance without requiring the researcher to specify the exact nature of either.

An analogy: counting opinions that aren't actually independent

Imagine polling 100 people about a policy, but 90 of them are members of the same family who all discussed it over dinner and now hold nearly identical views, while the other 10 are strangers with independent opinions. Treating all 100 as independent data points wildly overstates how much genuine, independent evidence you have — the 90 family members are really contributing something closer to one independent opinion, heavily duplicated, not 90 separate pieces of evidence. Ordinary standard errors make exactly this mistake with time-series data that has overlapping or autocorrelated errors: consecutive months' residuals aren't independent voices, they're more like family members echoing each other, and the standard error formula needs to be told to discount the echo.

The mechanics, one symbol at a time

The ordinary OLS standard error formula assumes the regression's error covariance matrix is a simple diagonal, $\sigma^2 I$ — constant variance, zero correlation between any two residuals. The Newey-West correction instead estimates a more general covariance matrix that allows both unequal variances (heteroskedasticity) and correlation between nearby-in-time residuals (autocorrelation) up to some number of lags $L$:

$$\hat{V}_{NW} = \hat{V}_0 + \sum_{\ell=1}^{L} \left(1 - \frac{\ell}{L+1}\right) \left( \hat{\Gamma}_\ell + \hat{\Gamma}_\ell' \right),$$

where $\hat{V}_0$ is the usual heteroskedasticity-robust term and $\hat{\Gamma}_\ell$ captures the covariance between residuals $\ell$ periods apart, with a weight that fades linearly to zero as $\ell$ approaches the chosen lag cutoff $L$. In plain English: the correction adds in the extra correlation between an observation and its recent neighbors (rather than assuming it's zero), while trusting more distant lags less and less, until they're excluded entirely past $L$ — the "HAC" in the name stands for heteroskedasticity and autocorrelation consistent, meaning the resulting standard errors stay valid under both problems at once.

Worked example 1: overlapping 12-month return regressions

A regression of overlapping 12-month forward returns on a valuation ratio, using 240 months of data, produces an ordinary OLS standard error on the valuation coefficient of 0.015, giving a t-statistic of $0.06 / 0.015 = 4.0$ — seemingly very significant. Because consecutive observations overlap by 11 out of 12 months, a natural rule of thumb sets the lag length $L$ around the overlap window (roughly 11 to 18 lags). Applying Newey-West with $L=12$ inflates the standard error to roughly 0.045 — three times larger — giving a corrected t-statistic of $0.06 / 0.045 \approx 1.33$, which is no longer significant at conventional levels. The apparent strong relationship was substantially an artifact of treating 240 heavily overlapping observations as if they carried the independent information of 240 separate data points.

Worked example 2: choosing the lag length

A different regression on non-overlapping monthly data still shows modest autocorrelation in its residuals — trading strategies often have return persistence over a few days from slow information diffusion or position unwinding. A common rule of thumb sets $L \approx 4(n/100)^{2/9}$; for $n=200$ months, this suggests $L \approx 4(2)^{0.222} \approx 4 \times 1.167 \approx 4.7$, rounded to 5 lags. Using too few lags (say $L=1$) under-corrects and leaves some inflation in the t-statistics; using too many (say $L=40$) over-corrects, inflating standard errors unnecessarily and reducing the test's power to detect a real effect. The rule of thumb balances these two failure modes, though in practice it's worth checking sensitivity to a few nearby choices of $L$ rather than trusting a single formula blindly.

What this means in practice

Newey-West standard errors are close to mandatory whenever a regression uses time-series data with any plausible source of autocorrelation — overlapping return windows, slow-moving macro variables, trading strategies with return persistence from position building or unwinding — situations extremely common in finance research. They're a safer default than ordinary OLS standard errors in almost any time-series regression, since applying them when autocorrelation is genuinely absent costs little (the correction reduces close to the ordinary standard error), while skipping them when autocorrelation is present can produce dramatically overstated significance.