Autocorrelation

Autocorrelation is a key idea in time-series analysis: it measures the degree to which a series of values is correlated with past values of itself. If today's figure tends to look like yesterday's, the series is autocorrelated. The concept is central to forecasting, econometrics and quantitative finance, and it's a common topic in risk qualifications like the FRM. This guide explains what autocorrelation is, why it matters, how it's detected, and what to do about it.

What is autocorrelation?

Autocorrelation — also called serial correlation — is the correlation of a time series with a lagged version of itself. "Lag" simply means how far back you look: a lag of one compares each value with the one immediately before it; a lag of twelve might compare each month with the same month a year earlier. Like any correlation, autocorrelation runs from +1 to −1:

• Positive autocorrelation means a high value tends to be followed by another high value (and low by low) — the series is persistent or trending.
• Negative autocorrelation means a high value tends to be followed by a low one — the series alternates.
• Zero autocorrelation means past values tell you nothing about the next one — the series behaves randomly with respect to that lag.

A simple example

Imagine a coffee shop's daily sales. If a busy day is usually followed by another busy day — because, say, good weather persists for a week — then sales show positive autocorrelation at a lag of one day. Now think about monthly sales: every December spikes for the holidays and every January dips. Compare each month with the same month a year earlier and you'll find strong positive autocorrelation at a lag of twelve. That seasonal pattern is autocorrelation, and recognising it is what lets the shop forecast next December from last December.

Why autocorrelation matters

Autocorrelation matters for two main reasons. First, it can be useful information: if a series is autocorrelated, past values help predict future ones, which is the basis of many forecasting models. Detecting a seasonal pattern — sales that spike every December, say — is really detecting autocorrelation at a seasonal lag.

Second, it can be a problem. Many statistical techniques, including standard regression analysis, assume that the errors (residuals) are independent of one another. When those errors are autocorrelated — common with time-series data — the model can still produce estimates, but the reported standard errors become unreliable, making the results look more precise and statistically significant than they really are. Spotting and correcting this is essential for trustworthy analysis.

How autocorrelation is detected

Several standard tools are used to identify autocorrelation:

• The autocorrelation function (ACF). A plot (often called a correlogram) showing the correlation of the series with itself at a range of lags. It's the quickest way to see at which lags a relationship exists.
• The Durbin–Watson test. A widely used statistic that tests specifically for first-order autocorrelation in regression residuals. Its value ranges from 0 to 4, with a figure near 2 suggesting little autocorrelation, below 2 indicating positive autocorrelation, and above 2 indicating negative.
• The Ljung–Box test. A test for whether autocorrelation is present across several lags at once, rather than just one.

What to do about it

If autocorrelation is the signal you're after — as in forecasting — you model it directly, using time-series methods such as autoregressive (AR) models that are built precisely to capture how a value depends on its own past. If, instead, autocorrelation is contaminating a regression, analysts typically respond by adding missing variables or lagged terms that explain the pattern, transforming the data, or using techniques that adjust the standard errors so inference is reliable again.

Why it matters for finance professionals

Financial data is overwhelmingly time-series data — prices, returns, interest rates and economic indicators all arrive in sequence over time — so autocorrelation is everywhere. Understanding it helps analysts build better forecasts, avoid being misled by spurious statistical significance, and interpret market data correctly. It's a foundational concept for anyone working in quantitative finance, risk or financial modelling.

Frequently asked questions

What is autocorrelation?

The correlation of a time series with a lagged version of itself — a measure of how much current values relate to past values. It's also known as serial correlation.

What does positive autocorrelation mean?

That high values tend to be followed by high values and low by low — the series is persistent or trending. Negative autocorrelation means values tend to alternate between high and low.

Why is autocorrelation a problem in regression?

Standard regression assumes the errors are independent. Autocorrelated errors make the reported standard errors unreliable, so results can appear more statistically significant than they actually are.

How is autocorrelation detected?

Through the autocorrelation function (ACF) plot, the Durbin–Watson test for first-order autocorrelation in regression residuals, and the Ljung–Box test for autocorrelation across multiple lags.

Build your quant skills with Learnsignal

Autocorrelation is part of the time-series toolkit at the heart of quantitative finance. Learnsignal's tutor-led courses, including the FRM, develop the statistics and modelling understanding that concepts like this build on — with clear teaching that makes the theory click.