Autocorrelation: Why Your Time Series Data Is Cheating on Your Model

Your regression model looks great. The R-squared is solid. The p-values are below 0.05. You write up the results and feel confident.

Photo by HRK Gallery on Pexels.

Then someone asks: "Did you check for autocorrelation?"

If you just felt a flicker of uncertainty, this post is for you.

What Autocorrelation Actually Means

Most statistical models assume your observations are independent. Flip a coin ten times and each flip has nothing to say about the next one. Pull ten random customers from a population and the same logic holds. Independence is the quiet assumption running underneath almost every standard inference procedure.

Time series data breaks that assumption constantly.

Autocorrelation means a variable is correlated with its own past values. Today's stock price predicts tomorrow's. This month's sales figures carry information about next month's. Yesterday's temperature tells you something real about today's. The data points are not independent witnesses; they are a chain of related events, each one leaning on the last.

The word itself is just "correlation with yourself, lagged in time." If you plot a variable against a version of itself shifted one period back, and that scatterplot shows a clear relationship, you have autocorrelation.

Why This Breaks Your Statistics

Here is where the damage happens. When you run an ordinary least squares regression on autocorrelated data, the coefficient estimates can still be unbiased. That part is fine. What breaks completely is your standard errors.

OLS assumes the residuals (the gaps between your model's predictions and reality) are random noise with no pattern. Autocorrelated data produces autocorrelated residuals. Those residuals carry structure. Your model cannot see that structure, so it interprets it as random variation and estimates the wrong amount of uncertainty.

The result: standard errors that are too small. Which means t-statistics that are too large. Which means p-values that are falsely significant. You think you have found a real effect. You have found an artifact of correlated noise.

This is not a minor adjustment. In some time series contexts, ignoring autocorrelation can shrink your standard errors by a factor of three or four. A result that should have a p-value of 0.15 gets reported as 0.02. Publications happen. Decisions get made.

graph TD
    A[Autocorrelated Data] --> B[OLS Regression]
    B --> C{Residuals Checked?}
    C -->|No| D[Underestimated Standard Errors]
    D --> E[False Significance]
    C -->|Yes| F(Corrected Model)
    F --> G[Valid Inference]

How to Spot It

The Durbin-Watson test is the classic diagnostic. It produces a statistic between 0 and 4; values near 2 suggest no autocorrelation, values near 0 indicate positive autocorrelation (high values following high values), and values near 4 indicate negative autocorrelation (high values followed by low ones).

The more informative tool is the ACF plot: the autocorrelation function plotted across multiple lags. If lag 1 shows a spike above the confidence bands, your data remembers yesterday. If lags 1 through 12 all show spikes, your data remembers an entire year. That plot tells you the shape and depth of the dependency, not just whether it exists.

Plot your residuals over time, too. If they drift up together for a stretch, then drift down together, that is positive autocorrelation showing up visually. Residuals should look like static, not like weather.

What to Do About It

Several approaches work, depending on your goal.

Newey-West standard errors (sometimes called HAC: heteroscedasticity and autocorrelation consistent) adjust your standard errors without changing the coefficient estimates. This is often the pragmatic fix when the autocorrelation is modest and you are primarily interested in inference rather than forecasting.

For modeling the time series itself, ARIMA models explicitly build the autocorrelation structure into the model. You estimate how much each period depends on previous periods and previous errors, and then the remaining residuals should be white noise. That is the actual goal: residuals with no memory.

First-differencing is a blunter instrument. Subtract each observation from the previous one and model the changes instead of the levels. Many economic time series that look highly autocorrelated in levels become much better behaved in differences.

The Deeper Issue

Autocorrelation is a symptom worth interrogating. When your data's past strongly predicts its future, that tells you something about the process generating the data. Prices have momentum. Diseases spread in waves. Habits persist. That structure deserves to be modeled explicitly, not swept under a rug of corrected standard errors.

Ignoring it is convenient. Your standard tools still run. The output looks normal. Nothing throws an error.

That is precisely what makes it dangerous. Silent violations are the ones that ruin analyses quietly, in the background, while you are busy trusting the numbers.