AI Glossary / Time series and signal processing / Intermediate
ARIMA model
Definition
An ARIMA model is a time-series model combining autoregressive terms, differencing and moving-average error terms. ARIMA stands for autoregressive integrated moving average. A nonseasonal ARIMA(p, d, q) uses p autoregressive lags, d rounds of ordinary differencing and q moving-average lags to describe the series' temporal dependence.
What it means
ARIMA uses the history of a series to forecast its future behavior. The autoregressive part relates a modeled value to past values; the moving-average part relates it to current and past innovations. Differencing can make a changing level more suitable for a stationary dependence model.
The word integrated refers to reversing differencing when returning forecasts to the original scale. It does not refer to combining databases. The moving-average component is also different from taking a rolling average of observed values. ARIMA is a statistical forecasting family, and it can be useful as a benchmark when evaluating more complicated AI forecasting systems.
How it works
Start by inspecting the series, observation spacing, missing values and relevant transformations. Choose differencing appropriate to the data, then estimate autoregressive and moving-average parameters for the differenced series. Under a common sign convention, that series depends on a constant, weighted previous values, the current innovation and weighted previous innovations.
Select candidate orders using domain knowledge, diagnostic plots and appropriate model-selection criteria. Check residual behavior and evaluate forecasts on later observations. Generate forecasts for the differenced series and undo the differences using the latest observed levels. Seasonal ARIMA adds seasonal terms and seasonal differencing. Software packages can differ in how they parameterize constants and trends, so record the actual equation rather than relying only on a model label.
Three practical examples
1. Service volumes: A support operation models regularly recorded ticket totals and compares ARIMA forecasts with simple baselines. The evaluation uses the forecast horizon needed for staffing decisions.
2. Demand planning: An analyst models a product series whose first differences appear more stable than its levels. The analyst checks whether promotions or stockouts require additional explanation before trusting the model.
3. AI benchmarking: A forecasting team includes a well-specified ARIMA model alongside a neural model. Performance on held-out periods determines which method is useful for the particular series and horizon.
Think of it this way
Think of predicting how a moving walkway changes speed before reconstructing where it will carry you. Differencing focuses attention on changes, the dependence model forecasts those changes, and accumulation returns to the original level. The analogy does not imply that every series needs differencing or that future changes follow a fixed physical law.
Worked example
Consider an original ARIMA(1, 1, 0) illustration with no constant. Let the first difference be d_t = y_t - y_(t-1), and assume d_t = 0.5d_(t-1) + e_t, with future innovations having conditional mean zero. The latest level is 120 and the latest observed difference is 4.
The next expected difference is 0.5 × 4 = 2, so the one-step level forecast is 120 + 2 = 122. The following expected difference is 0.5 × 2 = 1, giving a two-step level forecast of 123. The third expected difference is 0.5, giving a three-step forecast of 123.5.
These calculations assume a specified coefficient rather than estimating it from three observations. They also show why forecasting differences of 2, 1 and 0.5 is not the same as reporting level forecasts of those amounts. Future observations include innovations, so they need not follow this smooth expected path. Prediction intervals require an innovation-variance model and appropriate treatment of uncertainty.
Illustrative case study
Fictional case: Birch Support introduces a new service package and sees its weekly ticket level move upward. Its analysts investigate whether modeling changes in volume produces more stable dynamics than modeling raw totals. They compare an ARIMA candidate with a seasonal naive forecast and an exponential-smoothing model.
The team evaluates several successive forecast origins, including weeks before and after the launch. A model that fits the full history neatly performs poorly around the launch, revealing the limits of purely historical extrapolation. The final workflow documents the intervention and monitors forecast errors after deployment. ARIMA remains one candidate rather than an automatic default for every department.
Common misconceptions and mistakes
ARIMA is not a guarantee that differencing will solve every nonstationarity problem. Excessive differencing can introduce unwanted dynamics, and structural breaks may require different treatment. The letters p and q count model lags, not the number of periods forecast ahead.
The moving-average terms concern innovations, not a simple window average of the observations. A low training error does not prove good forecasts. Randomly shuffling a time series into training and test rows can leak future information. Finally, a good residual diagnostic does not establish causality or guarantee that unexpected events will be captured.
Business use
ARIMA can provide interpretable, relatively compact forecasts for regularly spaced operational series. Its value depends on whether the historical dependence is informative for the decisions being made. It is particularly useful as a disciplined baseline against which added AI complexity can be judged.
Specify the forecast horizon, update frequency and cost of errors before choosing a model. Retain prediction intervals where they are justified, and document transformations, differencing, missing-data handling and external events. Reassess the model when the data-generating process changes; a successful historical evaluation does not permanently establish suitability.
People also ask
Can d be zero?
Yes. ARIMA(p, 0, q) is an ARMA model under the corresponding specification.
Does ARIMA automatically include seasonality?
The basic nonseasonal notation does not; seasonal specifications add separate terms.
Can additional predictors be included?
Extensions allow regression terms with ARIMA errors, but their future values and interpretation require attention.
Does every trend require differencing?
No. Model choice depends on the nature of the trend and the intended analysis.
Check your understanding
What does d = 1 mean?
Answer: One ordinary difference is modeled.
What is the one-step level forecast in the example?
Answer: It is 122.
Why is the third level forecast 123.5?
Answer: It accumulates the expected changes 2, 1 and 0.5 from the latest level of 120.
Related terms
- ARMA model
- Differencing
- Stationarity
- Prediction interval
- Box-Jenkins methodology
- autoregression
- innovation
- seasonal ARIMA
Learn next
Learn next how to distinguish stationary dynamics from a changing level, interpret ACF and PACF plots, and validate forecasts through rolling origins. Then examine seasonal terms and regression with ARIMA errors.
More in Time series and signal processing
- Adaptive response rate - Adaptive response rate describes a forecasting approach that changes how strongly it reacts to new observations according to evidence of changing conditions.
- Analog Filter Design - Analog filter design is the process of specifying a continuous-time system that changes the amplitude or phase of different frequency components of a signal.
- Assumed-density filter - An assumed-density filter is a sequential Bayesian approximation method that keeps the evolving probability distribution within a chosen manageable family.
- Augmented Lagrangian - An augmented Lagrangian is an optimization function that combines the original objective, multiplier terms for constraints and an additional penalty for...
- Autocorrelated errors - Autocorrelated errors are model errors that are correlated with errors at other times or ordered positions.
- Autocorrelation - Autocorrelation is the correlation between a variable and a time-shifted version of itself.
Sources
Hyndman and Athanasopoulos, Forecasting: Principles and Practice, ARIMA models: https://otexts.com/fpp3/arima.html . statsmodels developers, ARIMA: https://www.statsmodels.org/stable/generated/statsmodels.tsa.arima.model.ARIMA.html . These primary technical references define model components and parameterization. The example uses a specified illustrative model and independently checked recursive forecasts.
Previous: Argumentative zoning Next: Arithmetic coding