Statistics Glossary → ARIMA orders (p, d, q) and drift
ARIMA orders (p, d, q) and driftconcept
The three numbers that define the structure of an ARIMA model: the number of autoregressive terms (p), the number of differences (d) and the number of moving-average terms (q); drift is a constant term in the differenced series.
In an ARIMA(p, d, q) model the series is first differenced d times to make it stationary. Then p autoregressive (AR) terms explain each value by its own p previous values, and q moving-average (MA) terms explain it by the q previous forecast errors. ARIMA(1, 1, 1), for example, is a model in which the first-differenced series is related to the previous difference and to the previous shock. Seasonal ARIMA adds seasonal orders (P, D, Q) and the season length m: ARIMA(p, d, q)(P, D, Q)m.
The orders are chosen starting from the ACF and PACF plots or with information criteria such as AIC and BIC; the residuals of the chosen model should be white noise (Ljung-Box test). Drift is the constant term in the differenced series when d = 1 and corresponds to a linear trend in the original series: the series rises or falls by a constant amount per period on average. With d = 0 a constant represents only the mean, and with d = 2 it would imply a quadratic trend, so drift is usually not included in those cases. Orders that are too high can lead to overfitting and to AR and MA terms cancelling each other out (parameter redundancy).
An everyday example
A café's monthly sales keep rising, with on average 300 more cups of coffee sold each month than the month before. In the first-differenced series this steady growth shows up as a constant drift of 300 cups. If, in addition, an unexpectedly busy month carries over a little into the next one, an MA term is added; the result might be an ARIMA(0, 1, 1) with drift.
In MerQur
It is set in ARIMA Forecast with the 'AR (p), Difference (d), MA (q)', 'Include drift (d = 1)' and 'Seasonal (P, D, Q) m' parameters.
Analyses where this concept appears
Related concepts
DifferencingAutocorrelation (ACF) and partial autocorrelation (PACF) functionsInformation criteria (AIC, AICc, BIC)