This feature is exclusively available to Starter, Research, and Pro. Upgrade when you need this workflow, review pricing, or send a question before changing plans.
What you keep on Free
How do you build a defensible univariate baseline forecast when the only information you trust is the series itself?
The ARIMA framework was formalized by George Box and Gwilym Jenkins in their 1970 monograph Time Series Analysis: Forecasting and Control, published by Holden-Day. The core problem it solved was principled: how to take a single observed time series, remove any deterministic non-stationarity through differencing, and then fit a parsimonious linear filter that captures the remaining autocorrelation structure. The Box-Jenkins methodology introduced a disciplined three-stage cycle of identification, estimation, and diagnostic checking that became the standard workflow for applied time-series analysis for the next three decades.
The statistical mechanism is a marriage of two linear filters. The autoregressive component regresses the differenced series on its own lagged values, capturing persistence and mean reversion. The moving-average component models the influence of past forecast errors, capturing the rate at which shocks dissipate. Differencing handles stochastic trends by reducing an integrated process to a stationary one. Together, the three pieces -- autoregressive order p, integration order d, and moving-average order q -- define the ARIMA(p, d, q) model. The Wold decomposition theorem guarantees that any covariance-stationary process admits an infinite-order MA representation, so ARIMA is not an arbitrary parametric choice but a finite-order approximation to the most general linear structure a stationary series can have.
ARIMA remains the workhorse univariate baseline in central bank forecasting divisions, statistical agencies, and applied macro research. The Federal Reserve Bank of St. Louis publishes automated ARIMA baselines for hundreds of FRED series. The European Central Bank and Bank of England use ARIMA benchmarks as the reference forecast against which richer multivariate and structural models are evaluated. The IMF's World Economic Outlook process includes ARIMA cross-checks for key aggregates. In academic econometrics, ARIMA is the natural null model in forecast comparison tests -- Diebold-Mariano tests and Model Confidence Set procedures almost always include an ARIMA benchmark in the comparison pool.
The Box-Jenkins framework predates modern information criteria. Originally, identification relied on visual inspection of sample autocorrelation and partial autocorrelation functions. Akaike (1974) and Schwarz (1978) introduced AIC and BIC, which enabled automated order selection. Hyndman and Khandakar (2008) operationalized fully automatic ARIMA fitting in the forecast package for R, which remains the most widely deployed automated ARIMA implementation. Python's statsmodels and pmdarima libraries provide equivalent functionality. The auto.arima algorithm combines unit-root pretesting (KPSS, ADF), stepwise order selection via AICc, and invertibility/stationarity constraint checking into a single pipeline.
The input is a single time series observed at regular intervals. The series enters the model through the differencing operator, which applies the lag operator L (defined by Ly_t = y_{t-1}) d times to produce a stationary working series. If the raw series has a unit root, first differencing (d=1) removes the stochastic trend. If the series has a second unit root -- rare in macro practice but possible for price levels -- second differencing (d=2) is applied. The differenced series is the object that the ARMA filter acts on.
The ARMA filter operates on the differenced series through two polynomial functions of the lag operator. The autoregressive polynomial maps lagged values of the series into the current-period conditional mean. The moving-average polynomial maps lagged innovations into the current-period conditional mean. The innovations are the one-step-ahead forecast errors under the model -- they are the new information arriving each period that the model could not have predicted from its own past. The conditional variance of the innovations is assumed constant (homoskedastic), which is the main structural constraint the basic ARIMA imposes.
Estimation proceeds by conditional or exact maximum likelihood. The likelihood function is constructed from the one-step-ahead prediction error decomposition: each observation contributes a Gaussian log-density evaluated at the prediction error, scaled by the conditional variance. The MLE jointly estimates the AR coefficients, the MA coefficients, and the innovation variance. Stationarity requires all roots of the AR polynomial to lie outside the unit circle; invertibility requires the same for the MA polynomial. These constraints are enforced either by reparameterization or by constrained optimization.
The primary use case is constructing a univariate forecast baseline for a single macro indicator. At the Federal Reserve Bank of Cleveland, ARIMA benchmarks for CPI components are generated monthly to detect whether multivariate models add genuine signal above the univariate floor. The ARIMA forecast serves as the null hypothesis: if a richer model cannot beat the ARIMA baseline in a pseudo-out-of-sample exercise, the added complexity is not justified. This role as a forecast benchmark makes ARIMA the most commonly estimated model in applied macroeconomic forecasting, even when it is not the final production forecast.
A secondary application is signal extraction and seasonal adjustment. The X-13ARIMA-SEATS program, maintained by the U.S. Census Bureau, fits a seasonal ARIMA model to decompose a series into trend, seasonal, and irregular components. This decomposition underpins the official seasonally adjusted figures published for GDP, employment, industrial production, and retail sales across most OECD countries. The ARIMA model inside X-13 is not a forecast tool in this context -- it is a signal-extraction filter that enables other forecasters to work with clean, deseasonalized data.
ARIMA hits a hard wall when the forecasting question involves multiple interacting variables. If the analyst cares about the joint dynamics of output, inflation, and interest rates, a univariate model for each series misses the cross-variable predictive information that a VAR would capture. ARIMA also cannot incorporate exogenous regressors in its pure form (ARIMAX extends it, but the extension is fragile and rarely preferred over a transfer function model or a small VAR). When the forecast horizon extends beyond 4-8 quarters, ARIMA forecasts converge rapidly to the unconditional mean of the differenced series, offering little value over a random walk with drift.
In real-time forecasting competitions -- the M3 and M4 competitions organized by Makridakis and colleagues -- simple ARIMA specifications routinely perform within a few percentage points of the best submissions across thousands of series. This resilience comes from the parsimony of the model: with typical macro series lengths of 100-300 observations, an ARIMA(1,1,1) has only 4 free parameters (two coefficients, a constant, and the innovation variance), leaving very little room for overfitting.
The raw time series at date t, measured at a fixed frequency (monthly, quarterly, annual).
The polynomial 1 - \phi_1 L - \cdots - \phi_p L^p whose roots must lie outside the unit circle for stationarity.
Applied d times to remove stochastic trends; maps an I(d) process to a stationary I(0) residual.
The polynomial 1 + \theta_1 L + \cdots + \theta_q L^q whose roots must lie outside the unit circle for invertibility.
White-noise disturbance with E[\varepsilon_t] = 0 and Var(\varepsilon_t) = \sigma^2, serially uncorrelated by construction.
The constant variance of the one-step-ahead prediction error; scales forecast uncertainty bands.
Defined by L^k y_t = y_{t-k}; the algebraic device that makes polynomial manipulation of lagged series possible.
After d applications of the difference operator, the resulting series w_t = (1-L)^d y_t has a time-invariant mean, variance, and autocovariance function: E[w_t] = \mu, Var(w_t) = \gamma_0, Cov(w_t, w_{t-k}) = \gamma_k for all t.
If violated: If the differenced series is still non-stationary, AR coefficient estimates are inconsistent and forecast intervals have incorrect coverage. Apply an additional difference or consider a fractional integration model.
The conditional mean of w_t is a linear function of its own past values and past innovations. No threshold, regime-switching, or nonlinear dependence enters the conditional mean.
If violated: If the true DGP is nonlinear (e.g., SETAR, Markov-switching), ARIMA captures only the linear projection and may underperform nonlinear alternatives at medium horizons.
Var(\varepsilon_t | \mathcal{F}_{t-1}) = \sigma^2 for all t. The conditional variance does not depend on past values or past shocks.
If violated: If innovations exhibit ARCH/GARCH effects, the point forecast remains consistent but forecast intervals are miscalibrated -- too narrow during volatile periods, too wide during calm periods.
The AR and MA coefficients and the innovation variance are constant across the estimation window. The DGP does not change regime.
If violated: A structural break inside the sample biases coefficient estimates toward a weighted average of the pre- and post-break regimes, degrading both in-sample fit and out-of-sample forecast accuracy.
The order d is correctly specified: the raw series is I(d) and the differenced series is I(0). Over-differencing introduces an MA unit root; under-differencing leaves a unit root in the ARMA representation.
If violated: Over-differencing (d too large) inflates forecast variance and introduces artificial negative autocorrelation. Under-differencing (d too small) leaves a spurious trend in the forecast path.
\varepsilon_t \sim N(0, \sigma^2). Required for exact MLE inference; relaxed under quasi-MLE where only the first two moments matter.
If violated: Non-Gaussian innovations do not affect consistency of the quasi-MLE but invalidate small-sample likelihood ratio tests and exact confidence intervals. Robust standard errors or bootstrap inference may be needed.