What it means
Autocorrelation asks a series a personal question: does what happened yesterday tell you anything about today? If high values tend to follow high values, the series is positively autocorrelated, and if highs tend to follow lows, it is negatively autocorrelated.
If the past carries no information, autocorrelation sits near zero. The measurement works by shifting the series against itself.
Analysts compare the data with lagged copies of itself, one period back, two periods back, and so on, computing a correlation at each lag, and the collection of these values, the autocorrelation function, maps how memory fades with distance. Cross-correlation extends the idea to pairs of series, asking whether one leads the other.
Financial data shows why the concept matters. Daily returns of a broad stock index have historically shown little autocorrelation, which is one statement of market efficiency.
Volatility, by contrast, clusters strongly, with turbulent days following turbulent days and calm following calm, and both facts are autocorrelation statements. The statistic is a diagnostic, not just a description.
In regression analysis, autocorrelated residuals mean the model missed structure, and standard errors computed naively will be too small. The NIST e-Handbook of Statistical Methods treats autocorrelation as a core tool for checking randomness and for identifying the order of time-series models, and its standard plots mark with bars beyond the significance band the lags where the past still speaks.
Trading applications follow directly: momentum strategies bet on positive autocorrelation at some horizons, while mean-reversion strategies bet on negative autocorrelation at others. Pitfalls come with the territory, because trends, seasonality, or smoothing in the data construction can manufacture autocorrelation, and it can vanish out of sample when regimes change.
A monthly series can show seasonal autocorrelation at lag twelve that has nothing to do with momentum. For non-finance managers, the concept upgrades everyday chart reading, because sales that are autocorrelated are forecastable from their own history while sales that are not need external drivers to predict.
Computation is trivial in any spreadsheet or statistics package, and one lagged scatter plot often reveals the structure faster than a table of coefficients. The modern error is overconfidence: a significant lag mined from dozens of series often fails the simplest out-of-sample check, long histories make tiny autocorrelations significant without making them tradable, and restated data can show autocorrelation created by the revision process itself.
In practice
Real-world examples.
Example
An analyst studying daily index returns finds autocorrelation near zero at every lag. That is consistent with efficient-market expectations, because yesterday's return gives no usable edge for predicting today's.
Example
A risk team at a bank plots squared returns and finds high autocorrelation even though the returns themselves look random. The pattern is volatility clustering, so the team widens its risk limits after a turbulent week instead of assuming calm will return at once.
Example
An economist regresses retail sales on advertising spend and finds that the residuals are autocorrelated. She adds a lagged sales term and re-estimates, which restores trustworthy standard errors and a more honest view of the advertising effect.
Formula
Calculation
Autocorrelation at lag k = the sum of (value at time t minus the mean) x (value at time t+k minus the mean), divided by the sum of squared deviations from the mean.
Example: monthly sales, in thousands of dollars, are $10, $11, $13, $14 and $17 over five months. The mean is (10 + 11 + 13 + 14 + 17) / 5 = $13, so the deviations are -3, -2, 0, 1 and 4, and their squares sum to 9 + 4 + 0 + 1 + 16 = 30. The lag-1 products are (-3 x -2) + (-2 x 0) + (0 x 1) + (1 x 4) = 6 + 0 + 0 + 4 = 10, so lag-1 autocorrelation = 10 / 30 = 0.33. A value near 0.6 or above would mean this month explains much of next month, while a value near zero would mean history does not forecast.Case study
Seen in the real world.
This is a fictional, illustrative example. Kestrel Analytics models weekly call-centre volume for staffing. The lag-1 autocorrelation is 0.55 and lag-52 is 0.4, revealing both momentum and annual seasonality.
A simple time-series model using those lags cuts staffing forecast errors by a third versus a naive average. In this illustrative story, Kestrel also tested the model on a held-out year before trusting it, because patterns found in-sample often fade. The model kept most of its accuracy, so the call centre adopted it for the next planning cycle and rebuilt its shift roster around the forecast.
Watch out
Common mistakes.
- Reading autocorrelation as a trading signal without testing out of sample, since measured patterns decay and regimes shift. In-sample memory is easy to find and hard to bank.
- Ignoring trends and seasonality, which create spurious autocorrelation that disappears once the data is detrended. Diagnose the structure before modelling it.
- Running regressions on autocorrelated series with ordinary standard errors, which overstates significance. Time-series data needs time-series methods.
Questions
People also ask.
What does positive autocorrelation mean?
Values tend to persist: high follows high and low follows low at the measured lag, giving the series momentum-like memory.
How is it measured?
By correlating the series with lagged copies of itself at successive lags, summarised in an autocorrelation function plot with significance bands.
Why does it matter for modelling?
Autocorrelated residuals signal missing structure and invalidate ordinary standard errors; the pattern also guides time-series model selection.
From the founder's library
Take it further with the book.
Build your financial confidence beyond this definition. Shihan's full-length guide, Accounting Fundamentals, takes the same plain-English approach and turns it into a complete, practical playbook for non-finance managers, business owners and students - with chapter-end quiz answers and presentation slides included.
25% off with code MMHQ25, applied at checkout. Priced in USD - checkout may show the equivalent in your local currency.
View the book and save 25%Related