CSMD: Curated Multimodal Dataset for Chinese Stock Analysis

📝 CSMD version 1.0: — Chinese Stock Multimodal Dataset

CSMD, a novel multimodal dataset for the Chinese stock market that encompasses stock prices and financial news texts, which is publicly available for research use. It includes mainly five advantages:

• ✅ CSMD integrates stock prices and financial news texts, with data rigorously curated and processed to ensure high quality and relevance.

• ✅ By leveraging large language models for factor extraction from news texts, CSMD enhances interpretability and facilitates understanding of market dynamics.

• ✅ News is sourced from authoritative outlets like Securities Times, ensuring authenticity and credibility of the information.

• ✅ The data undergoes meticulous preprocessing and denoising, effectively improving data quality and training efficiency.

In summary, CSMD is dedicated to providing high-quality multimodal data and flexible tools to advance research and applications in financial prediction.

Currently, the CSMD V1.0 dataset including CSMD50 and CSMD300 is available to the public, with plans to continuously update and refine datasets in the future.

💡 LightQuant version 1.0: — Lightweight Quantitative Framework for Stock Market Prediction

LightQuant is a lightweight quantitative framework for stock market prediction. It is designed to provide a flexible, modular framework supports multimodal data analysis, model training, and strategy backtesting, meeting diverse research requirements. With its modular design and flexible data processing capabilities, LightQuant aims to simplify the process of building and testing stock market prediction models.

🤗 We hope to provide a comprehensive, convenient, one-stop evaluation solution for the entire industry.

🚀 Changelog

• [2025.06.16] CSMD has now open-sourced part of the datasets and the complete data processing flow and code! We hope to open source the complete dataset and more detailed code after the paper is accepted. Download dataset here 🚀🚀🚀
• [2025.08.05] Congratulations!🎉🎉🎉 We are delighted to announce that our paper has been accepted by 34th ACM International Conference on Information and Knowledge Management(CIKM2025).
• [2025.09.08] We open source the complete curated multimodal dataset CSMD50 and CSMD300, hope our work can really help you. Download CSMD50 here Download CSMD300 here

✨ Table of Contents

• Data
• Quick Start
• Leaderboard
• TODO
• Disclaimer
• Licenses

Data

Our dataset collected data from trusted, authoritative sources, primarily including official Chinese financial media such as Securities Times, ensuring high authenticity and relevance. The dataset features rich multimodal information—combining stock prices and financial news texts—to support comprehensive research and modeling. We implement rigorous preprocessing steps, including data cleaning, denoising, and feature extraction, to enhance data quality and usability. Additionally, we leverage large language models (LLMs) like GPT-4 to extract impactful financial factors from news texts, significantly improving interpretability and aiding downstream tasks. This integrated approach ensures our dataset not only captures real-market dynamics but also supports robust, explainable financial analysis.

Here we show the process of how LLM extracts financial factors from news texts:

Quick Start

To get started with CSMD, follow these steps:

1. Installation: Install the necessary dependencies by running the following command:

pip install -r requirements.txt

2. Download: Download our dataset You can download the dataset from here.

Of cource, you can build your own dataset with our method. We open source our data collecting and processing flow and code, you can refer to ./data_construction to build your own dataset.

./dataset_construction/news_scraper.py  # this file is used to collect news data
./llm_factor/extract_factors.py  # this file is used to extract financial factors from news texts

3. Usage: Use the dataset in your own projects or use our framework to evaluate. Let me show you the structure of our framework:

./dataset  # this folder contains all the datasets
./evaluate # this folder contains the evaluation method of LLM's readability.
./model    # this folder contains the main models of our framework.
./price_experiment # this folder contains the code with single-modal(price) experiment.
./price_news_experiment # this folder contains the code with multi-modal(price and news) experiment.
./result # this folder contains the results of the experiment.
./utils # this folder contains the main utils.
./backtest # this folder contains the code with backtesting.
./llm_factor # this folder contains the code with extracting financial factors from news texts.

Then we can start to use our framework:

1. python ./utils/word2vec.py  --dataset CSMD50 --csv_news_path --embedding_path --local_model_path --trading_date_list #  generate word2vec embedding
2. python ./utils/util.py  --dataset CSMD50  # process and split the dataset
3. bash run.sh or python run.py # you can configurate the parameters in run.sh or run.py
4. bash run_backtest.sh # you can configurate the parameters in run_backtest.sh to backtest the model

Leaderboard

Below, we list the performance of the models we evaluated in the initial version. We have conducted a comprehensive assessment of various models on CSMD50 and CSMD300.

Performance comparison on the benchmark dataset.

The backtesting results on the CSMD 50 dataset.

TODO

• [√] Release the complete dataset;
• Add more prediction models;
• Optimize the backtesting strategy;
• Refine the details of our framework;

Disclaimer

⚠️ This project is for academic and non-commercial use only. Commercial use is strictly prohibited.

We take no legal responsibility for any use of the models or datasets. Users should independently assess potential risks and verify model outputs.

The original news content is copyrighted by Securities Times. We only provide derived data (e.g., timestamps, sentiment scores, embeddings), excluding any raw or verbatim text.

Users are encouraged to explore and expand upon this work to support research and AI innovation.

Licenses

This project adheres to MIT License.

The CSMD dataset follows Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.