Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction

The official PyTorch implementation of "Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction".

Overview

We propose Neural MJD, a neural forecasting model that combines time-varying diffusion and jump processes to handle non-stationary time series with abrupt changes.

Install Python Environment

conda env create -f setup/conda.yaml
conda activate neuralmjd

Prepare datasets

### SP500 data
curl -L -o ./data/sandp500.zip https://www.kaggle.com/api/v1/datasets/download/camnugent/sandp500
unzip ./data/sandp500.zip -d ./data/sandp500

Training and Evaluation

SP500 data

# Training
python train.py -c config/sp500/mjd/neural_mjd.yaml --dataset_name sp500 -m=$(hostname)

# Testing
python test.py -p=${MODEL_PATH} -m=$(hostname)

Demo use case

Please see demo_notebook.ipynb for a demo use case on synthetic data.

Citation

If you find this work useful, please consider citing:

@article{gao2025neural,
  title={Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction},
  author={Gao, Yuanpei and Yan, Qi and Leng, Yan and Liao, Renjie},
  journal={arXiv preprint arXiv:2506.04542},
  year={2025}
}