WorldTS

This repository provides the PyTorch implementation of WorldTS.

Introduction

WorldTS learns latent-state dynamics for time series forecasting with numerical, textual, and image covariates. It predicts future latent states and decodes them into forecasts through a two-stage training framework.

Quickstart

1. Requirements

Use separate environments for the three benchmarks. Run the following commands from the repository root:

# Numerical covariates
conda create -n worldts-dag python=3.8.20
conda run -n worldts-dag python -m pip install -r DAG-bench/requirements.txt

# Textual covariates
conda create -n worldts-vot python=3.10.20
conda run -n worldts-vot python -m pip install -r TimeMMD-bench/requirements.txt

# Image covariates
conda create -n worldts-unica python=3.10.20
conda run -n worldts-unica python -m pip install -r MMSP-bench/requirements.txt

2. Data preparation

Download the prepared datasets from the following sources and place them in the corresponding directories:

Benchmark	Data	Location
DAG-bench	Google Drive	DAG-bench/dataset/forecasting/
TimeMMD-bench	Prepared CSV files	TimeMMD-bench/data/
MMSP-bench	Google Drive	MMSP-bench/unica_datasets/mmsp/

For TimeMMD-bench, download the GPT-2 model and tokenizer to TimeMMD-bench/language_model/openai-community/gpt2/.

3. Train and evaluate

WorldTS implementations: Numerical · Textual · Image.

Run the commands below from the repository root in the corresponding environment.

# Numerical covariates: 12 datasets
conda activate worldts-dag
GPU=0 bash DAG-bench/scripts/w_future/worldts.sh

# Textual covariates: 8 datasets
conda activate worldts-vot
GPU=0 bash TimeMMD-bench/scripts/worldts.sh

# Image covariates: MMSP
conda activate worldts-unica
GPU=0 bash MMSP-bench/scripts/worldts.sh

Results

Forecasting results with numerical, textual, and image covariates are shown below.

Acknowledgements

We thank the authors of DAG, VoT, and UniCA for sharing their code and datasets.