Autocorrelation and Serial Correlation

Serial correlation is heteroskedasticity's time-series sibling: instead of error variances differing, error terms are correlated across time. It is endemic in finance because returns are sampled sequentially, signals persist, and, most insidiously, researchers routinely build overlapping observations that manufacture correlation out of thin air. Like heteroskedasticity, it leaves OLS coefficients unbiased while corrupting the standard errors, but the corruption is usually worse and the fix (Newey–West) is more delicate.

The problem

In the model $y_t = x_t^\top\beta + \varepsilon_t$, serial correlation means the error covariance matrix has non-zero off-diagonals:

$\operatorname{Cov}(\varepsilon_t, \varepsilon_{t-k}) = \gamma_k \ne 0 \quad \text{for some } k \ge 1.$

The simplest model is a first-order autoregressive error, $\varepsilon_t = \rho\,\varepsilon_{t-1} + u_t$ with $|\rho| < 1$ and $u_t$ white noise, giving $\operatorname{Corr}(\varepsilon_t, \varepsilon_{t-k}) = \rho^k$. As with heteroskedasticity, strict exogeneity is intact, so OLS stays unbiased and consistent. What breaks is the variance: the true covariance is $(X^\top X)^{-1}(X^\top\Omega X)(X^\top X)^{-1}$ with a now-dense $\Omega$, and the naive $\sigma^2(X^\top X)^{-1}$ is wrong.

Why the direction matters

With positive autocorrelation ($\rho > 0$), by far the common case in finance, the naive standard errors are too small, sometimes dramatically. The intuition: correlated observations carry less independent information than their count suggests, so you have fewer "effective" data points than $n$. For an AR(1) error the effective sample size is roughly $n\cdot\frac{1-\rho}{1+\rho}$; at $\rho = 0.5$ that is one-third of $n$. Treating $n$ correlated points as $n$ independent ones overstates precision, inflates $t$-statistics, and produces false significance, the exact mechanism behind many "significant" backtests that evaporate.

The Newey–West HAC estimator

The robust fix generalizes White to correlated errors: estimate the sandwich meat $X^\top\Omega X$ = $S = \sum_{t,s}\mathbb{E}[\varepsilon_t\varepsilon_s x_t x_s^\top]$ directly from residuals, including cross-time products. Newey and West's heteroskedasticity-and-autocorrelation-consistent (HAC) estimator is

$\hat S = \hat\Gamma_0 + \sum_{k=1}^{L} w_k\big(\hat\Gamma_k + \hat\Gamma_k^\top\big), \qquad \hat\Gamma_k = \sum_{t=k+1}^n \hat\varepsilon_t\hat\varepsilon_{t-k}\,x_t x_{t-k}^\top,$

with Bartlett weights $w_k = 1 - \frac{k}{L+1}$ that taper to zero. The weights are not cosmetic: they guarantee $\hat S$ is positive semidefinite (a valid covariance), which a naive truncated sum would not be. The lag length $L$ is the key tuning choice, it must grow with the sample (a common rule is $L \approx \lfloor 4(n/100)^{2/9}\rfloor$) and, for overlapping returns, must be at least the overlap horizon. The resulting HAC standard errors are the standard defense for any time-series regression on a desk.

Detecting it: Durbin–Watson and beyond

The classic Durbin–Watson statistic tests for first-order autocorrelation:

$\text{DW} = \frac{\sum_{t=2}^n (\hat\varepsilon_t - \hat\varepsilon_{t-1})^2}{\sum_{t=1}^n \hat\varepsilon_t^2} \approx 2(1 - \hat\rho).$

So $\text{DW} \approx 2$ means no autocorrelation, $\text{DW} \to 0$ strong positive, $\text{DW} \to 4$ strong negative. Its limitations, it only checks lag 1, and is invalid when a lagged dependent variable is a regressor, make the Breusch–Godfrey LM test (regress residuals on their own lags and the regressors) and the Ljung–Box portmanteau test the more general modern tools. Plotting the residual ACF is the fastest diagnostic.

The overlapping-returns trap

This deserves its own spotlight because it fools so many practitioners. To study $k$-period returns you might regress the overlapping sum $r_{t\to t+k} = r_{t+1} + \dots + r_{t+k}$ on a signal, sampled every period. Even if one-period returns are i.i.d., the overlapping $k$-period returns share $k-1$ of their components with their neighbors, mechanically inducing autocorrelation up to lag $k-1$ in the dependent variable and hence the errors. Naive standard errors can understate the truth by a factor of $\sqrt{k}$ or more, turning noise into "predictability." Long-horizon return-predictability regressions (dividend yield predicting 5-year returns) are the notorious example. The mandatory fixes: Newey–West with $L \ge k-1$, Hansen–Hodrick standard errors, or simply using non-overlapping observations.

Worked example

You regress annual (12-month overlapping, sampled monthly) returns on the dividend–price ratio and get $\hat\beta$ with a naive $t = 3.0$. The 12-month overlap induces autocorrelation through lag 11. Applying Newey–West with $L = 12$ typically cuts the $t$-statistic to roughly $1.5$–$2.0$, the "strong" predictability was mostly manufactured standard-error understatement. Same coefficient, honest uncertainty, very different conclusion.

Failure modes in financial data

• Choosing $L$ badly. Too small under-corrects (residual autocorrelation leaks through); too large adds noise and can make the estimate erratic. For overlapping data, $L$ must cover the overlap.
• Autocorrelation as a signal, not a nuisance. In the errors it's a nuisance to robustify; in the series itself it's forecastable structure, that is the domain of ARMA Models. Don't robustify away predictability you could model.
• Combined with heteroskedasticity. Real financial errors have both; Newey–West handles both simultaneously (it is HAC, not just AC), which is why it, not White, is the time-series default.
• Illiquidity-induced smoothing. Stale prices in illiquid assets (credit, real estate, some hedge-fund NAVs) induce spurious positive autocorrelation that understates volatility and overstates Sharpe, a data artifact, not alpha.

In interviews

Lead with the same unbiased-but-wrong-SE structure as heteroskedasticity, then specialize: positive serial correlation makes naive standard errors too small, inflating $t$-stats. Be able to explain the effective-sample-size intuition ($n\frac{1-\rho}{1+\rho}$ for AR(1)) and to name Newey–West HAC as the fix, including why the Bartlett weights exist (to keep the covariance positive semidefinite). The overlapping-returns question is a favorite for research roles, explain how overlap manufactures autocorrelation and inflates significance, and that the fix is HAC with lag $\ge$ overlap or non-overlapping sampling. Knowing Durbin–Watson $\approx 2(1-\hat\rho)$ is a reliable quick win.