Abstract
Predicting stock market movements is a complex task due to the volatile nature of financial markets and the influence of external news. This paper introduces a deep learning framework that extracts structured "events" from news headlines and converts them into dense vector embeddings. By combining a Neural Tensor Network (NTN) to learn these embeddings and a Convolutional Neural Network (CNN) to model their impact over different time scales, the authors achieved a significant improvement in prediction accuracy over state-of-the-art baselines.
Introduction: The Challenge of News-Driven Prediction
Predicting the stock market is notoriously difficult because prices react to a chaotic stream of information. Traditional quantitative models often rely on historical price data, while qualitative models try to "read" the news. However, previous news-based models faced three major hurdles:
- Lack of Structure: Simple "bag-of-words" models cannot distinguish between "Microsoft sues Barnes & Noble" and "Barnes & Noble sues Microsoft."
- Data Sparsity: While structured representations (Actor-Action-Object) solve the meaning problem, they create a sparsity problem—specific combinations of actors and actions are rare, making it hard for models to generalize.
- Temporal Decay: The influence of a news event isn't just a one-day spike; it lingers and decays over time. Most models failed to capture this long-term vs. short-term balance.
Methodology: The EB-CNN Framework
The authors propose a two-stage deep learning pipeline to address these challenges.
1. Learning Event Embeddings (The NTN)
Instead of treating an event as a string of text, they represent it as a tuple: $(O_1, P, O_2)$ (Actor, Action, Object). To solve the sparsity problem, they use a Neural Tensor Network (NTN). The NTN learns a dense vector (an embedding) for the whole event. By using tensors to model the specific interactions between the actor, action, and object, the model can recognize that "Nvidia's profits missed estimates" is semantically similar to "Delta's earnings fell short," even if they share no common words.
2. Multi-scale Modeling (The CNN)
Once event vectors are generated, they are fed into a Deep Convolutional Neural Network (EB-CNN). This architecture is designed to capture the diminishing influence of news over time by looking at three different horizons: Short-term (past day), Mid-term (past week), and Long-term (past month).
The convolution layers act as a sliding window, automatically picking out the most influential event combinations across these timeframes, while max-pooling layers retain only the most significant signals for the final prediction.
Experimental Results
The model was evaluated using over 10 million events extracted from Reuters and Bloomberg news (2006–2013) and stock data for the S&P 500 index and 15 major individual companies.
Prediction Accuracy
The proposed EB-CNN model achieved 65.08% accuracy on S&P 500 index prediction, significantly higher than the 58.94% achieved by previous state-of-the-art event-based models. On individual stocks, the model showed even stronger performance, with an average accuracy of 65.48%.
Profitability and Significance
In a simulated trading environment, the EB-CNN model generated a net profit of **$16,785** (starting with $10,000 per trade), nearly doubling the profits of the best-performing baseline ($8,694). A randomization test confirmed that these earnings were statistically significant at the 1% level, indicating they were not due to luck.
Discussion and Key Insights
- Structure and Embeddings: Moving from words to structured tuples was a known improvement, but this paper proves that embedding those tuples into a continuous vector space is the real "secret sauce" for generalization.
- Handling "No News" Days: A significant insight was that models relying only on daily news fail when there is no news for a specific stock. By incorporating mid-term and long-term event windows, the EB-CNN remains predictive even on quiet days.
- Limitations: The model currently ignores other vital signals like technical indicators (moving averages, RSI) or social media sentiment. Additionally, the event extraction relies on "Open Information Extraction" (Open IE), which can sometimes produce noisy tuples from complex sentences. Improving input cleanliness remains a primary area for future work.
Summary updated