native

Base for deterministic local forecasts with explicit horizons.

Fit returns the model; predict(n, series=...) returns the next n time steps. A supplied series changes the prediction context without changing fitted parameters. Subclasses declare min_train_series_length and implement _fit_values and _predict_values.

Fit an autoregression with jointly modeled components and ridge regularization.

A positive integer lags uses the preceding lags observations. A sequence of distinct negative offsets selects specific history positions. ridge=0 uses ordinary least squares; the intercept is never penalized. Prediction recursively appends each one-step forecast to the context. Covariates and probabilistic forecasts are deliberately not implicit.

Extrapolate the line joining the first and last context observation.

Repeat the fitted per-component mean for every forecast step.

Repeat the last K observations; K=1 is persistence.

Train any PyTorch module on chronological time-series windows.

The module receives (batch, input_chunk_length, components) and must return (batch, output_chunk_length, components). If omitted, a small linear module is constructed on first fit. The model uses GenericModel's training adapter and recursively forecasts in output-sized chunks.