What it means
Flip a fair coin and each toss forgets the last. Serial correlation is the failure of that forgetting: today's value carries information about yesterday's.
In prices, the question is whether returns echo: positive serial correlation means trends tend to continue, negative means moves tend to reverse, and zero is the random walk's claim. The NIST handbook's treatment is the practitioner's entry point: the lag plot, scatter of each value against its predecessor, answers directly whether the data are random or serially correlated.
In regression work, serial correlation is a diagnostic emergency: correlated residuals mean the model's standard errors are wrong, so coefficients look more certain than the data justifies. Finance cares for strategy reasons: measured momentum is serial correlation monetised, and its long-run profits are evidence that markets forget imperfectly, at least in places.
The tests have names: Durbin-Watson for regression residuals, Ljung-Box for return series, and the humble lag plot for anyone with a scatter and a minute. Structural breaks masquerade as correlation: a series that drifts between regimes will show echo in the aggregate, so persistence must be distinguished from mere nonstationarity.
For a non-finance reader, serial correlation is the question of whether yesterday tells you anything about tomorrow, asked with enough arithmetic that the answer is not just a feeling. Volatility shows its own memory: large moves cluster after large moves regardless of direction, a serial dependence in squared returns that volatility models like GARCH were built to capture.
Random-walk debates live on this measurement: early researchers found daily stock returns nearly uncorrelated, and that thin residue of echo was itself a finding, bounding how predictable prices could be. High-frequency data muddies the picture: bid-ask bounce creates artificial negative correlation tick to tick, so the honest analyst cleans microstructure noise before reading any memory into prices.
Portfolio construction quietly depends on the assumption: risk models that treat returns as independent day to day understate the chance of long losing streaks whenever serial correlation is positive.
In practice
Real-world examples.
Example
A lag plot of residuals reveals runs of same-direction misses, exposing inflated t-statistics in a returns model. The mistakes remembered themselves. After the analyst corrects for the correlation, several coefficients that looked significant no longer pass the test.
Example
Modelling changes instead of levels removes the inherited echo, leaving a shapeless residual cloud. A series of monthly sales that trends upward looks highly correlated with itself, but its month-to-month changes do not. Switching to changes gives honest standard errors.
Example
A momentum backtest's edge shrinks by a third once overlapping, serially dependent returns are handled honestly. Returns measured over rolling twelve-month windows share eleven months with their neighbours, which creates correlation by construction. The corrected estimate is smaller but far more believable.
Formula
Calculation
The lag-k autocorrelation is the correlation of the series with itself shifted k periods; the Durbin-Watson statistic tests regression residuals, near 2 meaning no first-order correlation, and Ljung-Box tests groups of lags jointly.
Worked example with invented monthly returns of 1%, 3%, 2%, 4%, 5% and 6%. The mean is 21 / 6 = 3.5%, and the deviations from the mean are -2.5, -0.5, -1.5, 0.5, 1.5 and 2.5.
- Sum of squared deviations = 6.25 + 0.25 + 2.25 + 0.25 + 2.25 + 6.25 = 17.5.
- Sum of lag-one products = (-2.5 x -0.5) + (-0.5 x -1.5) + (-1.5 x 0.5) + (0.5 x 1.5) + (1.5 x 2.5) = 1.25 + 0.75 - 0.75 + 0.75 + 3.75 = 5.75.
- Lag-one autocorrelation = 5.75 / 17.5 = 0.33, a positive echo in which a good month tends to be followed by another good one.
- The Durbin-Watson statistic is roughly 2 x (1 - 0.33) = 1.34, below the no-correlation value of 2.
Six data points are far too few to conclude anything. A common rule of thumb treats autocorrelations beyond about 2 divided by the square root of the number of observations as notable, which for 100 observations is 2 / 10 = 0.2.Case study
Seen in the real world.
This case study is fictional and illustrative. A made-up quant analyst at a regional asset manager is handed a colleague's model that predicts next month's returns with an impressive fit. Her first act is not to read the model but to plot the residuals against their own lags, the NIST lag-plot reflex. The picture kills the celebration: the errors cluster in runs, weeks of misses in one direction followed by weeks in the other, serial correlation strong enough that the model's standard errors are fiction and its impressive t-statistics are inflated by half.
The deeper dig finds the cause the plot implied: the target series itself trends for months at a time, so any model chasing it inherits the echo, and the fix is to model changes rather than levels, after which the residual cloud goes properly shapeless. The same habit pays on the strategy side the following quarter: a momentum sleeve's backtest is checked for serial dependence in its trades, and the honest estimate of its edge shrinks by a third once overlapping returns are handled correctly. Her training note for new quants distils the doctrine: before you believe a model, ask whether its mistakes remember themselves, because a model whose errors have memory is a model whose confidence is borrowed, and the lag plot is the cheapest lie detector in the building.
Watch out
Common mistakes.
- Trusting significance with correlated residuals; serial correlation corrupts standard errors, so the model's confidence intervals are wrong even when its line fits.
- Reading persistence as skill; regime drift and nonstationarity masquerade as exploitable correlation, and the remedies differ completely.
- Testing only lag one; dependence hides at specific lags, which is why lag plots and multi-lag tests complement the single-number statistics.
Questions
People also ask.
What is serial correlation?
The correlation of a series with its own past values, also called autocorrelation, indicating whether observations carry memory.
Why does it matter in regression?
Serially correlated residuals invalidate standard errors, overstating statistical significance even when coefficients look precise.
How is it detected?
Lag plots for a visual check, Durbin-Watson for first-order residual correlation, and Ljung-Box across multiple lags.
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%