Autocorrelation is correlation between a series and a shifted copy of itself.
At lag 1, pair today with yesterday.
At lag 7, pair today with the same weekday last week.
The autocorrelation function (ACF) repeats this for many lags.
Autocorrelation: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
Concept
What it means
Northstar example
↩️Lag means how far back.
For daily sales, lag 1 is one day and lag 7 is one week.
Positive correlation means high values tend to pair with high values; negative means high tends to pair with low.
🛒 For Northstar's daily data, lag 7 pairs this Monday's sales with last Monday's.
Lag 1 pairs today with yesterday.
🔎Correlation is a clue, not a model.
A peak at lag 7 may reveal weekly seasonality.
A slowly decaying ACF may reflect trend, persistence, or nonstationarity.
Remove trend and known seasonality before interpreting short-run memory.
🛒 Two adjacent high sales days may simply share a growing business level.
Remove trend and known seasonal effects before calling that useful short-run memory.
🔗ACF and PACF answer different questions.
The ACF includes indirect relationships through intermediate lags.
The partial ACF measures the linear relationship at a lag after accounting for intervening lags.
A stationary pure AR(1) has a decaying ACF and a theoretical PACF cut-off after lag 1.
🛒 After adjustment, compare direct lag relationships with indirect ones.
The pure AR(1) identity ACF(k) = φ^k is a teaching case, not a rule for Northstar's raw sales.
🎲A spike isn’t proof.
Approximate ±1.96/√n bounds are a white-noise reference.
Sampling variation and testing many lags produce some spikes by chance.
🛒 One ACF spike in a large collection of tested lags can arise by chance.
Check repeated structure and future forecast performance before adding a lag.
ACF(k) = correlation(yₜ, yₜ₋ₖ)
FIG 09.1
Meet a series and its past self
WHAT THIS SHOWS
Each dot pairs a day's sales index with its value a chosen number of days earlier. The bars repeat that correlation calculation for all lags. A common trend can make many nearby lags look useful.
The question it answers: How related is Northstar's series to its own past, and is that relationship useful for forecasting?
Series
Today vs 7 days earlier · r = 0.99
Correlation at every lag
Correlation at lag 70.99
Expanding-mean MAE49.91
Yesterday MAE7.33
Same weekday last week MAE3.86
Fine print & method
Constructed daily sales index. Bars and scatter both use Pearson correlation on the available lag pairs. Differencing changes the diagnostic plots; the three forecast scores always use the original level series and the last 64 days as one-step rolling tests, with every predictor strictly in the past. Correlation is descriptive, not causality or out-of-sample R². Differencing white noise can introduce negative lag-one dependence.
CHECK YOURSELF · CHOOSE AN ANSWER
Why can raw growing sales show strong lag-1 autocorrelation without useful short-run memory?