Multivariate

polars_ts.dl.multivariate

Native multivariate deep forecasting models.

MultivariatePatchTST: Channel-mixing PatchTST that processes all variates jointly through a shared transformer encoder.

iTransformerForecaster: Inverted transformer that treats each variate as a token (Liu et al., 2024), enabling cross-variate attention.

References

Nie et al. (2023). PatchTST. ICLR. Liu et al. (2024). iTransformer. ICLR.

_MVPatchTSTNet

Bases: Module

Channel-mixing PatchTST: all variates processed jointly.

forward(x)

Forward pass.

Parameters

x : Tensor of shape (batch, input_size, n_vars)

Returns

Tensor of shape (batch, h, n_vars)

_iTransformerNet

Bases: Module

Inverted transformer: each variate is a token.

forward(x)

Forward pass.

Parameters

x : Tensor of shape (batch, input_size, n_vars)

Returns

Tensor of shape (batch, h, n_vars)

MultivariatePatchTST

Channel-mixing PatchTST for multivariate forecasting.

Parameters

h Forecast horizon. input_size Lookback window length. patch_len Length of each input patch. target_cols List of target column names to forecast jointly. d_model Transformer hidden dimension. n_heads Number of attention heads. n_layers Number of transformer encoder layers. dropout Dropout rate. lr Learning rate. max_epochs Maximum training epochs. id_col, time_col Column names.

fit(df)

Train on multivariate panel data.

predict(df)

Generate multivariate forecasts.

iTransformerForecaster

Inverted transformer forecaster for multivariate time series.

Treats each variate as a token, enabling cross-variate attention.

Parameters

h Forecast horizon. input_size Lookback window length. target_cols List of target column names to forecast jointly. d_model Transformer hidden dimension. n_heads Number of attention heads. n_layers Number of transformer encoder layers. dropout Dropout rate. lr Learning rate. max_epochs Maximum training epochs. id_col, time_col Column names.

fit(df)

Train on multivariate panel data.

predict(df)

Generate multivariate forecasts.

_extract_multivariate(df, target_cols, id_col, time_col)

Extract per-series multivariate arrays.

Returns

ids List of series identifiers. arrays List of arrays, each of shape (seq_len, n_vars).

_build_mv_windows(arrays, input_size, h)

Build multivariate sliding windows.

Returns

X Shape (n_windows, input_size, n_vars). Y Shape (n_windows, h, n_vars).

_build_mv_forecast_df(ids, forecasts, target_cols, df, h, id_col, time_col)

Build output DataFrame with future dates and multivariate forecasts.

Parameters

forecasts Shape (n_series, h, n_vars).