Autoregressive Integrated Moving Average (ARIMA) Model

An autoregressive integrated moving average (ARIMA) model is a classic statistical method for forecasting a time series — a sequence of values measured over time, such as weekly sales, daily traffic, or monthly demand. The model is defined by three orders written as (p,d,q). The autoregressive part (the AR, or order p) says the next value depends on a number of past values of the series itself. The integrated part (the I, or order d) is differencing — subtracting each value from a previous one — repeated d times to make the series stationary, meaning its statistical behavior stops drifting over time. The moving-average part (the MA, or order q) says the next value depends on the recent forecast errors. Put together, ARIMA learns the pattern in a series' own history and projects it forward.

ARIMA grew out of the Box-Jenkins methodology, named for statisticians George Box and Gwilym Jenkins, who set out the now-standard procedure of identifying, estimating, and checking a model in the 1970s. Its appeal is that it forecasts a series using only the series itself — no outside drivers required — which makes it a sensible default whenever you have a reasonably long, regularly spaced history and want a baseline forecast. In marketing and growth work, ARIMA is used to forecast demand, web traffic, lead volume, or revenue, and to set the expectation a campaign or test is measured against. A seasonal extension, often written SARIMA, adds terms for repeating cycles such as weekly or yearly seasonality, which most real demand series have.

ARIMA is a univariate model — it forecasts a series from its own past, not from external predictors. That sets it apart from regression-style or machine-learning forecasters that fold in drivers such as price, weather, or ad spend. If your goal is to understand what moves the series, a model with explanatory variables (or a marketing-mix model) is the better fit; if your goal is a clean baseline projection of where the series is heading on its own, ARIMA is hard to beat for the effort involved. It also differs from simple exponential smoothing, which is closely related but frames the same problem differently. The honest framing is that ARIMA is a strong, interpretable baseline, not a universal best — newer methods can win on messy, high-dimensional, or highly nonlinear data.

The (p,d,q) orders are where judgment enters. Too few terms and the model misses real structure; too many and it overfits the past and forecasts poorly. The differencing order d matters most: ARIMA assumes a stationary series, so you difference until the series stops trending, then fit AR and MA terms to what remains. Get the differencing wrong and the whole forecast tilts. ARIMA also assumes the future resembles the past, so it handles steady patterns and seasonality well but cannot anticipate a structural break — a new competitor, a price change, a pandemic. That is the core limit: it extrapolates history, so it is most trustworthy for stable series over horizons where the underlying dynamics are not about to change.

Using an ARIMA model well starts with plotting the series and asking whether it is stationary. If it trends or drifts, difference it (raise d) until it settles; if it has seasonality, use a seasonal ARIMA with seasonal terms rather than forcing a plain model. Choose the AR order p and MA order q by inspecting the autocorrelation patterns or by letting an automatic routine search candidate orders and pick the best by a fit criterion that penalizes complexity. Then check the residuals — the leftover errors should look like random noise with no remaining pattern; if they do not, the model is missing structure. Always hold out recent data to test the forecast honestly rather than trusting the in-sample fit.

Treat ARIMA as a baseline you can defend, not an oracle. Pair the point forecast with its prediction interval so stakeholders see the uncertainty, and widen your skepticism over longer horizons, where small errors compound. Refit as new data arrives, because a model that fit last quarter can drift out of date. And keep its blind spot in view: ARIMA cannot foresee shocks or shifts that have no precedent in the history it learned from, so combine it with judgment, scenario thinking, or driver-based models when a known change — a launch, a price move, a market shift — is coming. Used this way, ARIMA gives a clear, reproducible, well-understood forecast that earns its long-standing place in the toolkit.

Autoregressive integrated moving average (ARIMA) — a Box-Jenkins time-series model with orders (p,d,q) — forecasts a series from its own history and remains a standard baseline for demand and marketing forecasting.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where autoregressive integrated moving average (arima) model is a core concern: