STOP 20 / 31Predict the next value from its own past.
Layout
STOP 20 · More models

Predict the next value from its own past.

EXPLORE~10 min
Back to route map
  • Autocorrelation describes how a series relates to shifted copies of itself.
  • Autoregression (AR) turns lagged values into predictors.
  • For a stable AR(1), today’s deviation from the mean is φ times yesterday’s deviation, plus a new unpredictable innovation.
ARIMA & autoregression: concepts and Northstar examples. Illustrative Northstar examples, not measured business results.
ConceptWhat it meansNorthstar example
AR: the series regresses on itself.
  • AR(1) uses one lag; AR(p) uses p lags with separate coefficients.
  • For AR(1), |φ| < 1 gives stationarity: positive φ creates persistence, negative φ creates alternation, and φ = 0 leaves white noise.

An AR term can use Northstar's recent adjusted-sales deviations as predictors instead of assuming every recent deviation immediately disappears.

Autocorrelation is not the AR coefficient in general.
  • For a stationary pure AR(1), theoretical ACF(k) = φ^k, so lag-1 autocorrelation equals φ.
  • That shortcut does not hold for arbitrary ARMA models or raw trending sales.
  • PACF helps distinguish direct lag relationships.
  • In the stationary pure AR(1) teaching process, lag-one correlation equals φ.
  • Do not transfer that shortcut to Northstar's trending, seasonal raw sales.
I: difference, then model.
  • ARIMA(p,d,q) applies AR and MA terms to a series differenced d times.
  • When d > 0, stationarity refers to that differenced working series—not necessarily the original sales levels.

If Northstar models changes in sales, it must undo the differencing to express forecasts back in revenue levels.

MA: past innovations, not a rolling average.
  • The moving-average part uses lagged forecast innovations.
  • In ARIMA, MA(q) is not a q-day mean of sales.
  • Innovations are inferred during fitting; future innovations have expected value zero.
  • ARIMA's MA term uses past unexpected forecast innovations.
  • It is not the seven-day rolling mean Northstar plots on a dashboard.
A concrete AR(1) forecast.
  • If the long-run mean is $30k, today is $35k, and φ = 0.7, tomorrow’s conditional mean is $33.5k.
  • Two days ahead it is $32.45k.
  • The shock fades toward the mean while uncertainty accumulates.

With mean $30k, current level $35k, and φ = 0.7, the supplied AR(1) conditional mean is $33.5k tomorrow—not a guarantee of tomorrow's sales.

yₜ − μ = φ(yₜ₋₁ − μ) + εₜ · E[yₜ₊ₕ | yₜ] = μ + φʰ(yₜ − μ)
FIG 20.1

Autocorrelation → autoregression

WHAT THIS SHOWS

This stationary AR(1) illustration shows lag memory and a conditional mean after a shock. It is not an automatic ARIMA/SARIMA fit.

The question it answers: How does positive or negative short-run memory affect a forecast?

ACF view
TOMORROW’S CONDITIONAL MEAN$30k + 0.70 × (+$5k) = $33.5K
Interactive time series chart$29k$31k$34k$36kToday+5 days+10 days+20 days
Conditional mean after today’s shockLong-run mean μ
THEORETICAL ACF(k) = φᵏPositive → persistent
Autocorrelation by lag-1-0.500.5112345678910111213141516Lag (days)
Lag 1 correlation0.70
Lag 7 correlation0.08
AR order p1 lag

Try φ = 0.9, 0, and −0.7. In a stationary pure AR(1), the theoretical ACF decays as φᵏ. Sample estimates wobble around it. The forecast line is a conditional mean with future innovations set to their expected value, zero—not a simulated future path. ARIMA can apply AR terms after differencing; its MA part uses past innovations, not rolling sales averages.

Fine print & assumptions

Constructed teaching data and illustrative calculations, not a production forecasting service. Read the diagram's method notes and distinguish supplied simulator components from fitted estimates.

CHECK YOURSELF · CHOOSE AN ANSWER

In a stationary AR(1), μ = $30k, today = $35k and φ = 0.7. What is tomorrow’s point forecast?

Your learning note

Private · just for you