Stock Market Sentiment Analysis

Investor sentiment plays a crucial role in influencing stock market movements, but capturing and quantifying real-time sentiment remains challenging. In this project, we developed a sentiment analysis pipeline by crawling 71,888 posts from Eastmoney’s Shanghai Composite Index forum. We fine-tuned a Chinese ELECTRA model to classify post sentiment into three categories and constructed a rolling sentiment index to track market emotions dynamically.

Project Structure

Key files based on different Usage have sturcture like:

.
├── data/ # Data files for usage: "Train model to predict whole test data" only
│   └── electra_sentiment_chinese/
│       ├── train_data/           # Training datasets from various sources
│       │   ├── train_data_1.xlsx # Dataset from Alibaba Tianchi
│       │   ├── train_data_2.csv  # Dataset from Alibaba Tianchi
│       │   ├── train_data_3.txt  # Positive sentiment dataset from GitHub
│       │   ├── train_data_4.txt  # Negative sentiment dataset from GitHub
│       │   └── train_data_clean.txt # Processed training data
│       ├── test_data/            # Test datasets for model evaluation
│       │   ├── test_data.csv     # Raw scraped test data
│       │   └── test_data_clean.csv # Processed test data
│       └── pred_data/            # Output directory for prediction results
├── model/ # Fine-tuned Chinese ELECTRA model files for usage: "Use trained model to predict several input comments" only
│   └── electra_sentiment_chinese/
│       ├── config.json           # Model configuration
│       ├── model.safetensors     # Model weights
│       ├── special_tokens_map.json # Special tokens configuration
│       ├── tokenizer_config.json # Tokenizer configuration
│       └── vocab.txt             # Model vocabulary
├── script/
│   ├── data_test_loader.py       # Formatter to the test data: `./data/electra_sentiment_chinese/test_data`
│   ├── data_test_crawler.py       # Crawler to the test data: `./data/electra_sentiment_chinese/test_data`
│   ├── data_train_loader.py       # Formatter to the train data: `./data/electra_sentiment_chinese/train_data`
│   ├── main.py                   # Main script for training and evaluating the model
│   ├── model_finetuner.py        # Script for fine-tuning the ELECTRA model
│   └── model_predictor.py        # Script for predicting sentiment of input │
└── README.md                     # Project documentation
└── requirements.txt              # Dependencies for the project

Data Source

After finishing the Unzip step in Train model to predict whole test data part, you will be able to find several datasets in ./data/train_data or ./data/test_data. Here are their sources:

• train_data_1: https://tianchi.aliyun.com/dataset/158814
• train_data_2: https://tianchi.aliyun.com/dataset/179229
• train_data_3: https://github.com/algosenses/Stock_Market_Sentiment_Analysis/blob/master/data/positive.txt
• train_data_4: https://github.com/algosenses/Stock_Market_Sentiment_Analysis/blob/master/data/negative.txt
• test_data: Scraped by ../script/data_test_scraper.py.

Enviroment

See requirements.txt for a list of necessary Python packages.

Setup and Installation

1. Clone the repository

git clone https://github.com/yourusername/Stock-Market-Sentiment-Analysis1.git
cd Stock-Market-Sentiment-Analysis1

2. Set up the environment and activate

Choose either venv or conda:

• venv:

  python -m venv venv
  # On Windows
  .\venv\Scripts\activate
  # On macOS/Linux
  source venv/bin/activate
  

• conda:

  conda create --name risk_env python=3.9 # Or desired Python version
  conda activate risk_env
  

3. Install dependencies

pip install -r ./requirements.txt

Usage

You can either repeat the training pipeline to predict the sentiment using the comments from ./data/test_data or use a trained model to predict the sentiment of several input Chinese comments.

Train model to predict whole test data

To train model to predict whole test data by yourself:

1. Download: Download the data.rar in Google Drive.

2. Unzip: Unzip the data.rar in root directory ./.

3. Execute: Execute main.py and you will be given a model trained by yourself in ./model/electra_sentiment_chinese and a predicted dataset in ./data/electra_sentiment_chinese/pred_data based on a testing dataset in ./data/electra_sentiment_chinese/test_data/test_data_clean.csv.

Use trained model to predict several input comments

To use trained model to predict several input comments:

1. Download: Download the model.rar in Google Drive.

2. Unzip: Unzip the model.rar in root directory ./.

3. Execute: Execute model_predictor and you will be required to input Chinese comments about stock topics in your desired quantity. After inputting, you will be given a sentiment analysis for these comments.

Results

The sentiment analysis model classifies comment into three categories:

• Positive: Indicating optimistic market sentiment, with sentiment value of 2
• Neutral: Indicating balanced or uncertain market views, with sentiment value of 1
• Negative: Indicating pessimistic market sentiment with, sentiment value of 0

Here are different results example for different usages:

• Example sentiment analysis of market comments in Train model to predict whole test data part:

  User Age	Comment #	Influecen Power	Read #	Comment	Data	Sentiment Value	Confidence	Weighted Sentiment Value
  8.59	0	1.93	0	国内科创板	2019-03-28	2	0.56	1.50
  0.04	0	0.5	200	科创板融资问题	2019-03-28	2	0.65	1.58
  0.25	0	1.0	574	缩量都能把股指拉红神勇的大A	2019-03-27	2	0.73	1.59

• Example sentiment analysis of market comments in Use trained model to predict several input comments part:

  Comment	Sentiment	Confidence	Weighted Sentiment
  经济数据好于预期，市场上涨	Positive	0.89	1.88
  投资者对新政策持观望态度	Neutral	0.75	1.31
  通胀数据令人担忧，可能引发抛售	Negative	0.82	0.03

License

MIT License

Copyright (c) 2025 Stock Market Sentiment Analysis Project