Understand the framework of ARIMA models (e.g., trends and seasonality).

Monthly auto-claim counts drift upward and spike every December. You need a forecasting model that handles the drift, the holiday spike, and the autocorrelation between consecutive months at once. That is the job ARIMA was built for.

Building blocks. Let $B$ be the backshift operator with $B{Y}_t=Y_{t-1}$ and $B^k{Y}_t=Y_{t-k}$. Three operators do all the work:

• AR(p): $\varphi (B)=1-{\varphi }_{1}B-{\varphi }_2{B}^2-\cdots -{\varphi }_p{B}^p$.
• MA(q): $\theta (B)=1+{\theta }_{1}B+{\theta }_2{B}^2+\cdots +{\theta }_q{B}^q$.
• Differencing: $(1-B{)}^d$ for regular drift, $(1-B^s{)}^D$ for seasonal drift.

The general ARIMA(p,d,q) model. Letting $\{{\epsilon }_t\}$ be white noise with mean zero and variance ${\sigma }^2$:

$$\varphi (B)(1-B{)}^d{Y}_t=\theta (B){\epsilon }_t$$

When $d=0$ you have a stationary ARMA(p,q). When $d\ge 1$ the model is integrated of order $d$, meaning $Y_t$ itself is not stationary but $(1-B{)}^d{Y}_t$ is.

A linear deterministic trend $Y_t=\alpha +\beta t+{\eta }_t$ becomes $(1-B)Y_t=\beta +({\eta }_t-{\eta }_{t-1})$, which has constant mean $\beta$. A stochastic trend (random walk with drift) is removed the same way.

Common mistakes

• Over-differencing. Applying $d=2$ when $d=1$ already stationarized the series introduces a unit MA root and inflates forecast variance. A common wrong answer is to second-difference whenever any ACF spike remains; the correct test is whether the lag-1 ACF of the once-differenced series is close to $-0.5$ (the unit-root signature).
• Confusing seasonal and regular orders. Writing ARIMA(1,1,1) when the data are monthly with a clear annual cycle ignores the seasonal block. The right specification carries a $(P,D,Q{)}_{12}$ tag with at least $D=1$ or $Q=1$.
• Reading the ACF and PACF backwards. AR(p) cuts off in the PACF, not the ACF. MA(q) cuts off in the ACF. Swapping them produces the wrong $p$ and $q$; the trap value is selecting AR(1) for a series whose PACF is the one that tails off.

Bottom line

• ARIMA(p,d,q) = AR order $p$, differencing order $d$, MA order $q$, applied to the series after $d$ regular differences. The "I" makes the series stationary; AR and MA model the stationary residual.
• Stationarity requires constant mean, constant variance, and an autocovariance that depends only on lag. Trends violate the mean condition; differencing fixes them.
• ACF and PACF signatures. Pure AR($p$): PACF cuts off at lag $p$, ACF tails off. Pure MA($q$): ACF cuts off at lag $q$, PACF tails off. Mixed ARMA: both tail off.
• Roots condition. Stationarity needs all AR roots outside the unit circle; invertibility needs all MA roots outside.

Exam shortcut

If the level ACF decays slowly and the differenced ACF has one negative spike then noise, default to ARIMA(0,1,1) with negative MA coefficient. If a monthly or quarterly series has a clear cycle, set seasonal $D=1$ first and re-identify; the airline model SARIMA(0,1,1)(0,1,1)${}_{12}$ is the right starting point unless evidence demands more terms.

The full lesson (about 2,309 words, 15 min read) adds 2 worked examples, all 7 common mistakes, a self-check, free in the app.

Learning objectives

• D2

Browse all free MAS-II lessons or jump into free MAS-II practice questions.