TL;DR
This paper proposes a novel deep learning framework that extracts financial events from news, represents them using BERT/ALBERT, and uses a hierarchical attention network to predict stock market movements. The approach significantly outperforms existing methods in accuracy and simulated trading returns, demonstrating the power of event-driven analysis for understanding market behavior.
What Problem Does This Paper Address?
The financial world is awash in news, and it's widely accepted that this information influences stock market movements. However, automatically extracting and leveraging this information for prediction is a complex challenge. Key difficulties include:
- Ambiguity of "Good" or "Bad" News: News can have varied impacts on different entities (e.g., a company's bad news might be good for its competitors).
- Temporal and Cumulative Effects: The impact of news isn't always immediate; it can accumulate over time and across different articles.
- Information Granularity: Raw news articles are long and unstructured. Extracting concise, meaningful financial events that capture the essence of the information is crucial.
- Representing Event Similarity: How can we effectively represent financial events so that their similarities and relationships can be learned for prediction?
This paper aims to address these challenges by developing a system that can systematically learn from financial event streams to predict stock market behavior.
Methods
The authors introduce a three-pronged approach:
- Combined Event Extraction: They use a pipeline that combines Open Information Extraction (OpenIE) with neural co-reference resolution. This aims to extract events that are not only factual but also "compact and complete with meaning," by resolving pronoun references and ensuring events are independent.
- BERT/ALBERT Enhanced Event Representation: Instead of traditional methods, they leverage pre-trained language models like BERT and ALBERT to create rich, contextualized representations of the extracted financial events. This allows for a deeper understanding of event semantics and their relationships.
- Extended Hierarchical Attention Network (HAN): They build upon existing HAN architectures by adding an "event-level" attention mechanism. This network then incorporates attention at the event, news, and temporal (daily) levels, allowing it to weigh the importance of different pieces of information at various granularities to predict stock movements.
Datasets
The study utilizes financial news data from Bloomberg and Reuters covering the period from 2006/10 to 2013/11. This dataset is significantly larger than those used in prior work, containing over 10 million events. For evaluation, they use this data to predict movements of the S&P 500, Dow Jones, and Nasdaq indices, as well as 10 individual stocks. They also create a new test set using data from November 2013 to April 2020 to assess performance on more recent data and to investigate the impact of pre-training data timing.
Key Insights and Ideas
- The Power of Large-Scale, Coreference-Free Events: The sheer volume of extracted events (millions) and the process of making them coreference-free significantly boosts prediction accuracy, highlighting the importance of comprehensive data.
- BERT/ALBERT for Financial Event Semantics: Using advanced pre-trained language models for event representation is crucial for capturing nuanced meanings and relationships that simpler methods miss.
- Hierarchical Attention is Key: The multi-level attention mechanism (event, news, temporal) effectively models how different types of information contribute to market movements over time.
- Market Behavior is Learnable from News: The success in predicting stock movements, especially with simulated trading, suggests that systematic patterns driven by financial news are indeed learnable.
Key Results
The proposed ALBERT+eventHAN model achieved state-of-the-art results:
- Accuracy: Significantly higher accuracies across 2, 3, and 5-category stock movement predictions for S&P 500, Dow Jones, and Nasdaq indices compared to four strong baselines (wordHAN, ECK+event, HATS, docBERT). For example, on the 2013 test set, ALBERT+eventHAN achieved 93.3% accuracy for S&P 500 2-category prediction, compared to the best baseline's 85.1%.
- Simulated Annualized Returns: The model demonstrated superior performance in market trading simulations, achieving up to 93.2% annualized returns, significantly outperforming baselines (up to 71.9%).
- Individual Stocks: The approach also showed strong performance on individual stock predictions, with accuracies up to 55.7%, outperforming baselines.
- Explainability: The attention weights provided insights into which events were most influential for specific predictions, often highlighting negative words and specific figures for downward movements.
Limitations and Future Work
- Pre-training Data Timing: A potential concern is whether the pre-training data for BERT/ALBERT might overlap with the period used for testing, potentially biasing results. The authors address this by creating a new test set from 2020, which is after the release of ALBERT.
- Computational Cost: Processing millions of events and using large models like BERT/ALBERT is computationally intensive, especially with long event sequences.
- Global Markets: The current work focuses on U.S. stock markets. Future work could extend this to global markets (e.g., Nikkei, TOPIX, HSI).
- Integration with Financial Models: Integrating textual information with traditional financial asset pricing models (like CAPM, APT) is suggested as a promising direction.
Summary updated