Parametric Prior Mapping Framework for Non-stationary Probabilistic Time Series Forecasting

This repository contains the official implementation of Parametric Prior Mapping Framework for Non-stationary Probabilistic Time Series Forecasting (PPM).

News

📌 [2026] 🎉 PPM is accepted by ICML 2026.

Introduction

PPM is a probabilistic forecasting framework designed for non-stationary time series. Instead of relying on an approximate posterior or iterative denoising process, PPM directly learns an input-conditioned parametric prior and pushes it forward through a generative mapping to obtain the predictive distribution.

The framework aims to provide accurate and efficient probabilistic forecasting, especially under non-stationary dynamics.

Installation

Please install the required packages with:

pip install -r requirements.txt

Data

Please place the datasets in dataset/.

Usage

You can run the experiments using the scripts provided in scripts/MLP/.

For example:

bash scripts/MLP/ETTh1.sh

Please check the scripts for detailed configurations, including dataset name, prediction length, model settings, and training hyperparameters.

Contacts

For questions about the paper or code, please contact:

• Jinglin Li: ljinglin8336@gmail.com
• Jun Tan: juntan.csu@gmail.com
• Qi Fang: csqifang@csu.edu.cn
• Ning Gui: ninggui@gmail.com

License

This project is released under the Apache License 2.0.