ARIMA
Autoregressive integrated moving average for stationary time series forecasting
ARIMA models a time series as a combination of autoregressive (AR) lags, differencing (I) for stationarity, and moving average (MA) error terms. It is the foundational statistical approach for univariate time series forecasting.
When to use:
- Univariate time series with trend but no strong seasonality
- Data that becomes stationary after differencing
- Short-to-medium forecast horizons where statistical interpretability matters
Input:
- Trained model checkpoint — exported ARIMA fit from training
- Preprocessing config — scaling/normalization settings
- Training tail — last N observations for warm-starting lag features
- Steps — number of future time steps to forecast
Output: Forecasted values for the specified number of steps
Model Settings (set during training, used at inference)
AR Order (p) (default: 1, range: 0–5) Number of autoregressive lag terms. Determined by PACF analysis during training.
Differencing Order (d) (default: 0, range: 0–2) Number of differencing operations applied to achieve stationarity.
MA Order (q) (default: 0, range: 0–5) Number of moving average error terms. Determined by ACF analysis during training.
Inference Settings
No dedicated inference-time settings. The model generates forecasts by extending the fitted ARIMA process forward.