TL;DR
This paper introduces FAST, a novel hierarchical network that leverages financial news and tweets to predict stock trading profits. By modeling the temporal dynamics and diverse influence of textual data, FAST significantly outperforms existing methods in both cumulative and risk-adjusted returns.
What Problem Does This Paper Address?
Predicting stock market movements and achieving profitable trading is a notoriously difficult challenge. While traditional methods rely on historical price data, they often fail to capture the impact of real-world events like breaking news and social media sentiment. Existing NLP-based approaches, while better, often treat text data as uniformly spaced, ignoring the critical role of timing in how quickly markets react to information. Furthermore, most NLP methods optimize for prediction accuracy (classification/regression) rather than directly for profit, which can lead to strategies that are accurate but not financially optimal. This paper aims to bridge this gap by developing a model that explicitly optimizes for trading profit by considering the time-aware nature and varying influence of textual information.
Methods
The core of FAST is a hierarchical temporal attention mechanism.
- Intra-Day Textual Information Encoder:
- Text Embedding: Uses pre-trained BERT models to generate rich contextual embeddings for individual news articles and tweets.
- Time-Aware LSTM (t-LSTM): A modified LSTM that accounts for the irregular time intervals between texts. It applies a decay function to the short-term memory based on elapsed time, allowing it to better model the dynamic influence of information.
- Intra-Day Attention: An attention mechanism that weighs the importance of different texts within a single day, helping to filter out noise and emphasize more influential information. This aggregates the t-LSTM outputs into a single daily representation.
- Inter-Day Temporal Encoder:
- LSTM: Processes the sequence of daily representations from the intra-day encoder.
- Inter-Day Attention: An attention mechanism that weighs the importance of different days within a lookback window, capturing time-aware dependencies across days. This results in a final stock-level representation.
- Ranking and Optimization:
- The final stock-level representations are fed into a feed-forward network to predict a return ratio.
- The model is optimized using a joint point-wise regression and pairwise rank-aware loss function, directly aiming to rank stocks by their expected profitability.
Datasets
The paper evaluates FAST on two benchmark datasets:
- US S&P 500: Comprises 109,915 English tweets from January 2014 to December 2015, related to 88 high-volume stocks from the NASDAQ. Tweets were extracted using stock tickers.
- China & Hong Kong: Includes 90,361 Chinese financial news headlines from January to December 2015, covering 85 top-traded stocks in Shanghai, Shenzhen, and Hong Kong. News was aggregated from major financial websites.
Historical stock prices for all stocks were collected from Yahoo Finance. The datasets were split temporally for training, validation, and testing.
Key Insights and Ideas
- Timing is Crucial: The paper strongly emphasizes that the precise timing of financial news and tweets significantly impacts their influence on stock prices. Simply treating all text as equally spaced is a major limitation of prior work.
- Profit Optimization is Key: Directly optimizing for trading profit (via ranking) is more effective than optimizing for classification or regression accuracy, as demonstrated by FAST's superior performance.
- Hierarchical Attention Captures Nuance: The combination of intra-day and inter-day attention mechanisms effectively filters noise, identifies influential information, and models the complex temporal dependencies in textual data.
- Textual Data is Powerful: Leveraging financial news and tweets, even with their inherent noise and irregularity, provides significant advantages over methods relying solely on historical price data.
Key Results
- Profitability: FAST achieved significantly higher cumulative investment returns (over 8%) and risk-adjusted returns (over 10%) compared to all baseline methods across both datasets.
- Ranking Performance: FAST also outperformed baselines in ranking metrics like MRR and NDCG@5.
- Component Ablation: The study showed that each component of FAST (t-LSTM, intra-day attention, inter-day attention) contributed to improved performance, with the full model yielding the best results.
- Time-Awareness Impact: Performance dropped significantly when the time granularity of the t-LSTM was coarsened, highlighting the importance of fine-grained temporal modeling.
- Qualitative Analysis: Visualizations demonstrated how FAST's attention mechanisms correctly identified influential news headlines and days, leading to better stock rankings and profit predictions, outperforming baselines like RankNet and SN-HFA.
Limitations and Future Work
- Lookback Window Sensitivity: The study found that mid-sized lookback windows (around 5 days) performed best, suggesting that very short or very long windows can be detrimental. This indicates a sensitivity to the chosen lookback period.
- Parameter Sensitivity: FAST's performance varies with parameters like the lookback window length and the number of top stocks considered for trading.
- Future Work: The paper suggests that further exploration of the hierarchical temporal attention mechanism and its application to other financial forecasting tasks would be valuable. The authors also implicitly suggest that extending this approach to other forms of financial data or markets could be a fruitful direction.
Summary updated