API Reference / algorithms / timeseries /

deep/

Deep, torch-backed forecasters.

Three neural window forecasters that learn a mapping from a lookback window to a horizon, rather than from a hand-specified process as the classical models in timeseries do.

Algorithms

  • NBEATSForecaster: doubly-residual stacks of fully connected blocks, with
an optional interpretable trend/seasonality decomposition.
  • NHITSForecaster: N-BEATS plus multi-rate max pooling and hierarchical
interpolation, so each stack specialises in one frequency band.
  • PatchTSTForecaster: a Transformer over patch tokens, with channel
independence and reversible instance normalisation.

Notes

These models require PyTorch, which TuiML treats as an optional extra: 

pip install 'tuiml[torch]'

Importing this package does not import torch, and the classes construct, register and report their parameter schema without it, so list_algorithms() shows the same catalog on every install. The dependency is demanded by fit, which raises an ImportError naming the install command.

They all follow the forecasting convention used by the classical models: fit(y) takes the series itself rather than a design matrix, and predict(steps) returns the next steps values.

python 
>>> from tuiml.algorithms.timeseries.deep import NBEATSForecaster
>>> model = NBEATSForecaster(random_state=0)
>>> model.lookback, model.horizon
(24, 8)

Modules

nbeats

N-BEATS: neural basis expansion analysis for time series forecasting.

1 class · 2 functions
nhits

N-HiTS: neural hierarchical interpolation for time series forecasting.

1 class · 2 functions
patchtst

PatchTST: a patch-based Transformer forecaster with channel independence.

1 class · 3 functions