Vector Autoregression
Multivariate time series model capturing interdependencies between multiple series
Vector Autoregression (VAR) models multiple time series simultaneously, capturing how each variable's past values influence the others. It is the multivariate generalization of ARIMA.
When to use:
- Multiple interrelated time series (e.g., economic indicators, stock prices, system metrics)
- When cross-series causal effects are important for forecasting accuracy
- Impulse response analysis and Granger causality testing
Input:
- Trained model checkpoint — exported VAR model
- Preprocessing config — scaling settings
- Training tail — last N observations for all series
- Steps — forecast horizon
Output: Forecasted values for all modeled series simultaneously
Model Settings (set during training, used at inference)
Lag Order (p) (default: auto) Number of lag periods included. Higher orders capture longer-range dependencies but require more data.
IC (default: aic)
Information criterion for automatic lag order selection. aic, bic, hqic, or fpe.
Trend (default: c)
Deterministic terms. n = none, c = constant, ct = constant + linear trend.
Inference Settings
No dedicated inference-time settings. All series are forecast jointly using the trained coefficient matrices.