Auto-ARIMA

Auto-ARIMA performs a stepwise search over ARIMA parameter combinations and selects the best model using AIC or BIC. This removes the need for manual ACF/PACF analysis and order selection.

When to use:

• When ARIMA order selection is unclear or time-consuming
• Automated pipelines where manual tuning is not feasible
• Multiple series forecasting where each series may have different optimal orders

Input:

• Trained model checkpoint - exported auto-selected ARIMA model
• Preprocessing config - scaling settings
• Training tail - last N observations
• Steps - forecast horizon

Output: Forecasted values using the auto-selected model

Model Settings (set during training, used at inference)

Auto-ARIMA searches within bounds set during training:

• Max P / Max Q - upper bounds for AR and MA order search
• Max D - maximum differencing order
• Seasonal - whether to include seasonal components
• Information Criterion - aic or bic for model selection

The selected (p, d, q) and seasonal orders are fixed at training time.

Inference Settings

No dedicated inference-time settings. The selected order is baked into the trained model.