Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time
September 6, 2022 10:09
October 1, 2022 17:07
December 1, 2021 12:06
October 1, 2022 17:18
January 3, 2021 16:30
September 20, 2021 17:06
December 3, 2020 10:07
September 6, 2022 10:09
September 14, 2020 09:47
September 6, 2022 11:41
December 1, 2021 12:06
December 3, 2020 10:07

PyTorch Forecasting

PyPI Version Conda Version Docs Status Linter Status Build Status Code Coverage

Documentation | Tutorials | Release Notes

PyTorch Forecasting is a PyTorch-based package for forecasting time series with state-of-the-art network architectures. It provides a high-level API for training networks on pandas data frames and leverages PyTorch Lightning for scalable training on (multiple) GPUs, CPUs and for automatic logging.

Our article on Towards Data Science introduces the package and provides background information.

PyTorch Forecasting aims to ease state-of-the-art timeseries forecasting with neural networks for real-world cases and research alike. The goal is to provide a high-level API with maximum flexibility for professionals and reasonable defaults for beginners. Specifically, the package provides

  • A timeseries dataset class which abstracts handling variable transformations, missing values, randomized subsampling, multiple history lengths, etc.
  • A base model class which provides basic training of timeseries models along with logging in tensorboard and generic visualizations such actual vs predictions and dependency plots
  • Multiple neural network architectures for timeseries forecasting that have been enhanced for real-world deployment and come with in-built interpretation capabilities
  • Multi-horizon timeseries metrics
  • Ranger optimizer for faster model training
  • Hyperparameter tuning with optuna

The package is built on pytorch-lightning to allow training on CPUs, single and multiple GPUs out-of-the-box.


If you are working on windows, you need to first install PyTorch with

pip install torch -f

Otherwise, you can proceed with

pip install pytorch-forecasting

Alternatively, you can install the package via conda

conda install pytorch-forecasting pytorch -c pytorch>=1.7 -c conda-forge

PyTorch Forecasting is now installed from the conda-forge channel while PyTorch is install from the pytorch channel.

To use the MQF2 loss (multivariate quantile loss), also install pip install pytorch-forecasting[mqf2]


Visit to read the documentation with detailed tutorials.

Available models

The documentation provides a comparison of available models.

To implement new models or other custom components, see the How to implement new models tutorial. It covers basic as well as advanced architectures.

Usage example

Networks can be trained with the PyTorch Lighning Trainer on pandas Dataframes which are first converted to a TimeSeriesDataSet.

# imports for training
import pytorch_lightning as pl
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning.callbacks import EarlyStopping, LearningRateMonitor
# import dataset, network to train and metric to optimize
from pytorch_forecasting import TimeSeriesDataSet, TemporalFusionTransformer, QuantileLoss

# load data: this is pandas dataframe with at least a column for
# * the target (what you want to predict)
# * the timeseries ID (which should be a unique string to identify each timeseries)
# * the time of the observation (which should be a monotonically increasing integer)
data = ...

# define the dataset, i.e. add metadata to pandas dataframe for the model to understand it
max_encoder_length = 36
max_prediction_length = 6
training_cutoff = "YYYY-MM-DD"  # day for cutoff

training = TimeSeriesDataSet(
    data[lambda x: <= training_cutoff],
    time_idx= ...,  # column name of time of observation
    target= ...,  # column name of target to predict
    group_ids=[ ... ],  # column name(s) for timeseries IDs
    max_encoder_length=max_encoder_length,  # how much history to use
    max_prediction_length=max_prediction_length,  # how far to predict into future
    # covariates static for a timeseries ID
    static_categoricals=[ ... ],
    static_reals=[ ... ],
    # covariates known and unknown in the future to inform prediction
    time_varying_known_categoricals=[ ... ],
    time_varying_known_reals=[ ... ],
    time_varying_unknown_categoricals=[ ... ],
    time_varying_unknown_reals=[ ... ],

# create validation dataset using the same normalization techniques as for the training dataset
validation = TimeSeriesDataSet.from_dataset(training, data, min_prediction_idx=training.index.time.max() + 1, stop_randomization=True)

# convert datasets to dataloaders for training
batch_size = 128
train_dataloader = training.to_dataloader(train=True, batch_size=batch_size, num_workers=2)
val_dataloader = validation.to_dataloader(train=False, batch_size=batch_size, num_workers=2)

# create PyTorch Lighning Trainer with early stopping
early_stop_callback = EarlyStopping(monitor="val_loss", min_delta=1e-4, patience=1, verbose=False, mode="min")
lr_logger = LearningRateMonitor()
trainer = pl.Trainer(
    gpus=0,  # run on CPU, if on multiple GPUs, use accelerator="ddp"
    limit_train_batches=30,  # 30 batches per epoch
    callbacks=[lr_logger, early_stop_callback],

# define network to train - the architecture is mostly inferred from the dataset, so that only a few hyperparameters have to be set by the user
tft = TemporalFusionTransformer.from_dataset(
    # dataset
    # architecture hyperparameters
    # loss metric to optimize
    # logging frequency
    # optimizer parameters
print(f"Number of parameters in network: {tft.size()/1e3:.1f}k")

# find the optimal learning rate
res = trainer.lr_find(
    tft, train_dataloaders=train_dataloader, val_dataloaders=val_dataloader, early_stop_threshold=1000.0, max_lr=0.3,
# and plot the result - always visually confirm that the suggested learning rate makes sense
print(f"suggested learning rate: {res.suggestion()}")
fig = res.plot(show=True, suggest=True)

# fit the model on the data - redefine the model with the correct learning rate if necessary
    tft, train_dataloaders=train_dataloader, val_dataloaders=val_dataloader,