CSMD: Curated Multimodal Dataset for Chinese Stock Analysis
CSMD: Curated Multimodal Dataset for Chinese Stock Analysis
This paper presents CSMD, a curated multimodal dataset for Chinese stock market analysis, addressing the lack of high-quality, time-aligned price and news data for the Chinese market. The dataset covers two major indices (SSE 50 and CSI 300) from 2021 to 2024, with financial news sourced from Securities Times. To enhance data quality, the authors employ large language models (LLMs) guided by sequential knowledge prompts to extract interpretable and impactful factors from raw news texts, followed by rigorous quality validation through manual and automatic evaluations. Additionally, they introduce LightQuant, a lightweight and modular backtesting framework that simplifies data processing, model integration, and strategy evaluation.
Experimental results demonstrate that CSMD outperforms the existing Chinese dataset CMIN-CN across various single-modal and multimodal models in stock trend prediction tasks, with improvements in accuracy and Matthews correlation coefficient. Backtesting experiments on CSMD 50 further validate the practical trading effectiveness, with multimodal models like StockNet achieving strong risk-adjusted returns. The work provides a valuable resource for researchers and practitioners in Chinese quantitative finance, enabling more reliable and interpretable stock movement prediction.
Highlights
- 1Introduces CSMD, a multimodal dataset for Chinese stock analysis covering SSE 50 and CSI 300 indices with price and news data from 2021-2024.
- 2Uses LLMs guided by sequential knowledge prompts to extract interpretable and impactful news-based factors, enhancing data quality.
- 3Develops LightQuant, a lightweight and user-friendly backtesting framework with modular architecture for rapid prototyping.
- 4Demonstrates superior performance of CSMD over existing Chinese dataset CMIN-CN across multiple models in stock trend prediction and backtesting.
Methods
- MFinancial news collection from Securities Times with automated pipeline and LLM-based factor extraction using prompts.
- MData quality validation through manual evaluation by financial experts and automatic scoring using MiniLM-L6-v2 and GPT-4.
- MLightQuant framework with three layers (Data, Model, Evaluation) for data processing, model training, and backtesting.
- MEvaluation using single-modal (LSTM, BiLSTM, ALSTM, Adv-LSTM, SCINet, DTML) and multimodal models (StockNet, HAN, PEN).
Results
- RCSMD 300 and CSMD 50 consistently outperform CMIN-CN in accuracy and MCC across most models, e.g., StockNet achieves 55.47% ACC on CSMD 300 vs 54.53% on CMIN-CN.
- RMultimodal models (StockNet, HAN) achieve higher ACC and MCC than single-modal models on CSMD datasets.
- RBacktesting on CSMD 50 shows StockNet achieves highest annualized return (0.1301) and Calmar ratio (3.0182), while ALSTM achieves highest Sharpe ratio (0.8192).
- RCSMD datasets exhibit high quality in denoising, financial sentiment, and readability for both humans and LLMs.
Analyze Paper
Generate insights from "CSMD: Curated Multimodal Dataset for Chinese Stock Analysis".