This repository contains the code for the paper EXPRTS: Exploring and Probing the Robustness of Time Series Forecasting Models. We present a framework for transforming time series data and a visual analytics tool called EXPRTS, that allows users to explore the robustness of time series forecasting models. The framework supports a variety of transformations, including simple noise addition, scaling, and shifting, and more advanced transformations that modifies specific time series characteristics.

To run the application start by installing and activating the environment:

conda env create -f environment.yaml
conda activate EXPRTS

We have uploadedthe necessary files to run an example with a dense network on the electricity dataset. Use the following command to run the example:

bokeh serve src/ --args experiments/electricity_nips/feedforward_prob/config.yaml

To run the application with a new model and/or dataset, create a config.yaml file in experiments/<dataset>/<model>. Then the model can be trained by running

python scripts/fit.py <config_path>

During training, the script saves batches of training data to ``data//training_data`. We can calculate features from this data like this:

python scripts/calculate_features.py <config_path>

To calculate the features for predefined test data from GluonTS use the optional --test-data argument:

python scripts/calculate_features.py <config_path> --test-data=1

Evaluating models can be done like this:

python scripts/evaluate.py <config_path>

Once all of the above steps have been completed for the new dataset and/or model, the application can be started by running

bokeh serve src/ --args experiments/<dataset>/<model>/config.yaml

You can cite our work using the following BibTeX entry:

@misc{kjaernli2025exprts,
      title={EXPRTS: Exploring and Probing the Robustness of Time Series Forecasting Models}, 
      author={Håkon Hanisch Kjærnli and Lluis Mas-Ribas and Hans Jakob Håland and Vegard Sjåvik and Aida Ashrafi and Helge Langseth and Odd Erik Gundersen},
      year={2025},
      journal={arXiv preprint arxiv:2403.03508}
}