TLDR: A research paper introduces a deep learning framework that predicts short-term stock price movements by combining historical stock data with emotion features extracted from tweets. It uses Meta’s Llama 3.1-8B-Instruct model to preprocess tweets, enhancing the quality of emotion analysis from methods like DistilRoBERTa and NRC lexicons. An LSTM model is then trained on this integrated data. Experimental results on TSLA, AAPL, and AMZN stocks show that incorporating Llama-enhanced emotion features, especially with DistilRoBERTa, significantly improves the accuracy of predicting significant price changes, raising it from a baseline of 13.5% to 38.5%.
Predicting the short-term movements of stock prices has always been a complex challenge for investors and researchers alike. The stock market is highly volatile and sensitive to various factors, including investor sentiment. A recent research paper explores a novel approach to tackle this challenge by integrating emotion features extracted from social media tweets with historical stock price data to forecast significant price changes for the following day.
The study introduces a deep learning framework that leverages the power of large language models (LLMs) to enhance the quality of emotion analysis from tweet data. Specifically, Meta’s Llama 3.1-8B-Instruct model is utilized to preprocess tweets, filtering out non-emotional content and refining the emotional signals. This LLM-enhanced preprocessing step is crucial for improving the accuracy of subsequent emotion analysis.
Three distinct emotion analysis methods were employed in the framework: a transformer-based DistilRoBERTa classifier from the Hugging Face library, and two lexicon-based methods using National Research Council Canada (NRC) resources (NRC Emotion Intensity Lexicon and NRC Emotion Label Lexicon). These methods generate different representations of emotions expressed in tweets, such as anger, joy, fear, and anticipation.
Once the emotion features are extracted and aggregated daily, they are combined with historical stock price information, including open, close, high, low prices, and trading volume. The research focuses on predicting significant price movements, classifying daily changes into three categories: ‘Significant Increase,’ ‘Significant Decrease,’ or ‘Stable,’ based on the standard deviation of price changes. A Long Short-Term Memory (LSTM) model, known for its effectiveness in handling sequential data like financial time series, is then trained using this combined dataset.
The experiments were conducted on prominent U.S. technology stocks: Tesla (TSLA), Apple (AAPL), and Amazon (AMZN), using tweet data and historical prices from September 2021 to September 2022. The results demonstrate a significant improvement in predicting high-volatility stock movements when emotion features are included, compared to a baseline model that only uses historical stock prices. The baseline model achieved an average accuracy of 13.5% for predicting significant price movements.
The DistilRoBERTa-based stock prediction model, particularly when combined with Llama-enhanced emotion analysis, showed the most promising performance. Its average accuracy for predicting significant price increases and decreases rose from 23.6% (without Llama enhancement) to an impressive 38.5% (with Llama enhancement). This highlights that using large language models to refine tweet content significantly boosts the effectiveness of emotion analysis, which in turn leads to more accurate stock price movement predictions.
Also Read:
- Signature-Informed Transformer: A New Approach to Risk-Aware Asset Allocation
- Enhancing Financial Question Answering with Knowledge Graphs: A New Benchmark
This research underscores the growing importance of social media data and advanced AI techniques in financial forecasting. While the study primarily focuses on emotion analysis, future work could explore integrating these insights with technical indicators and financial news to further enhance prediction accuracy. For more details, you can read the full research paper here.
