This repository implements a dual-network framework based on the Mamba architecture for probabilistic time series forecasting. One network produces point forecasts, while the other estimates point-wise uncertainty for the forecasted region.

We discuss the motivation and impact of this approach in our manuscript:
"Mamba Time Series Forecasting with Uncertainty Quantification", published on IOP's Machine learning: science and technology and as preprint on arXiv

To reproduce the analysis/figures from our manuscript, follow these steps:

1. Generate synthetic data by running:

   bash synthetic_data.sh
   

2. Obtaining the real datasets from here

3. Run forecasting scripts for different datasets
   Use the following scripts to run experiments on different datasets:

• Sines dataset:

  bash sines_script.sh

• Van der pol dataset:

  bash VDP_script.sh

• Electricity consumption dataset dataset:

  bash ECL_script.sh

• Traffic occupancy dataset:

  bash traffic_script.sh

• Brownian motion dataset:

  brownian_script.sh
  

4. Generate the figures using the iPython notebook make_figures.ipynb

@article{pessoa2025mamba,
  title   = {Mamba time series forecasting with uncertainty quantification},
  author  = {Pedro Pessoa and Paul Campitelli and Douglas P. Shepherd and S. Banu Ozkan and Steve Pressé},
  journal = {Machine Learning: Science and Technology},
  volume  = {6},
  pages   = {035012},
  year    = {2025},
  doi     = {10.1088/2632-2153/adec3b}
}

Initial versions of this code were built on top of S-Mamba (GitHub, Neurocomputing article). With the express authorization of the authors, we have adapted and modified the original implementation.