Leveraging Language Models for Stock Predictions

In recent years, large language models (LLMs) have become popular tools in various fields, including finance. These advanced models can analyze text and provide insights based on sentiment. This study looks at the potential use of LLMs to help predict stock price changes in China by analyzing news articles related to publicly traded companies.

#The Role of Sentiment in Trading

Sentiment analysis is the process of determining the attitude or feeling expressed in a piece of text. In finance, sentiment can significantly influence stock prices. Positive news may lead to higher stock prices, while negative news can result in declines. By assessing sentiment in news articles, traders can make better-informed decisions regarding their investments. This research aims to understand how well LLMs can extract sentiment from Chinese news summaries to create strategies for trading stocks.

#The Challenge of Language

Using LLMs for sentiment analysis in Chinese presents unique challenges. Most LLMs were primarily trained on English texts, which makes them less effective when applied to other languages. Chinese has distinct characteristics and rules that differ from English, making it essential to employ models specifically designed or fine-tuned for Chinese text. This research will explore different LLMs' effectiveness in extracting sentiment from Chinese financial news and establishing trading strategies based on these insights.

#Experimental Design

To assess the potential of LLMs in sentiment analysis, a systematic experimental design was put in place. The research involved several steps:

1. Data Collection: A large set of news articles related to Chinese publicly traded companies was gathered. A total of 394,429 articles were collected from credible financial news sources. These articles were filtered to ensure they were published before the stock market opened, allowing traders to use the sentiment analysis right before making investment decisions.

2. Sentiment Extraction: Three different LLMs were used in this study: a baseline model, a Chinese language-specific model, and a model fine-tuned for financial text. Each model was tasked with analyzing the sentiment in the news articles and classifying them as positive, negative, or neutral.

3. Trading Strategy Development: Based on the extracted sentiment, trading strategies were developed. The goal was to buy stocks with the highest positive sentiment and sell those with negative sentiment. The performance of these strategies was then tested using simulations that accounted for real-world trading conditions.

#Methods for Sentiment Analysis

#Baseline Model: ChatGPT

The first model used was ChatGPT, a well-known LLM with extensive training data. This model was employed without any additional modifications. It analyzed the sentiment of news articles by generating a response based on provided prompts. Each article was classified as having good, bad, or neutral sentiment, and numerical values were assigned to each classification for further analysis.

#Chinese Language-Specific Model: Erlangshen-RoBERTa

The second model was Erlangshen-RoBERTa, which was specifically trained on Chinese texts. This model takes into account the unique features of the Chinese language, aiming to provide more accurate sentiment analysis. It was used in a similar manner as ChatGPT, analyzing the same set of news articles and classifying them based on sentiment.

#Financial Domain-Specific Model: Chinese FinBERT

The final model used was Chinese FinBERT, designed for financial text analysis. It was further trained on financial documents, allowing it to focus on the specific nuances and vocabularies commonly found in financial news. Like the other models, it was applied to the same news articles to classify sentiment.

#Trading Simulation

Once sentiment factors were extracted using the three models, trading strategies were implemented. The simulation followed these key steps:

• Daily Portfolio Updates: Investment portfolios were adjusted daily based on the latest sentiment analysis, with stocks bought or sold at market open. Sentiment factors were used to rank stocks and determine which ones should be included in the portfolio for trading.

• Realistic Trading Conditions: To simulate real trading conditions, several factors were incorporated into the simulation. This included accounting for delays and transaction fees. A realistic approach was taken by using the Volume-Weighted Average Price (VWAP) for trades, which reflects more realistic market prices over a short period.

• Performance Metrics: The success of the trading strategies was measured using various performance indicators, including annual returns, win rates, and risk-adjusted returns. These metrics helped to evaluate how well each strategy performed based on the sentiment extracted from the news articles.

#Results

After running the trading simulations, the performance of the portfolios generated from each sentiment extraction model was analyzed. Key findings include:

1. Erlangshen-RoBERTa Model Outperforming Others: The Chinese language-specific model, Erlangshen-RoBERTa, showed the best performance in terms of returns compared to the other models. Despite having fewer parameters than some competitors, it proved highly effective at extracting sentiment from the Chinese financial news.

2. Continuous vs. Discrete Sentiment Rating: The difference in sentiment rating methodologies played a significant role in performance. Erlangshen-RoBERTa provided continuous sentiment scores, which offered more granularity in differentiating between varying Sentiments. This was contrasted with the discrete ratings from ChatGPT.

3. Importance of Broad Training Data: The research highlighted that having a broader training dataset on the Chinese language could be more beneficial than focusing solely on domain-specific training. This was evident as the Chinese FinBERT model underperformed relative to Erlangshen-RoBERTa, despite its specialization in finance.

#Discussion

The findings indicate that language models can significantly influence trading strategies through sentiment analysis, especially when tailored to the language and context of the information being analyzed. The study reveals that:

• Models specifically trained on the Chinese language can offer improved sentiment classification than those primarily trained on English.

• Utilizing models that incorporate a wider range of contexts, such as general Chinese texts, may provide stronger results in sentiment analysis tasks, even in specialized fields like finance.

• Continuous sentiment scoring can enhance the selection process for stocks, as it allows for better differentiation between news items.

#Conclusion

This research explored the potential of large language models in predicting stock movements by analyzing sentiment in Chinese financial news. The study demonstrated that different models vary in effectiveness, with Erlangshen-RoBERTa showing superior performance for this task. By establishing a clear process for sentiment analysis and trading strategy development, the research offers valuable insights for traders and researchers in the financial domain.

The findings advocate for the development and application of language models that consider the unique characteristics of the language in which they operate. For traders looking to leverage sentiment analysis in their strategies, this study emphasizes the importance of using appropriate models and methodologies that align with their specific requirements. Further research may continue to refine these approaches and explore their applications in other contexts.