Financial markets are notoriously chaotic. To predict where a stock is headed, traders don't just look at price charts; they scan the news for CEO scandals, track supply chain disruptions, and monitor global economic shifts. However, building an AI that can synthesize all these "signals" is incredibly difficult.
In a paper presented at ACL 2022, researchers Gary Ang and Ee-Peng Lim from Singapore Management University introduced GAME (Guided Attention Multimodal Multitask Network), a sophisticated model designed to navigate this complexity by treating financial forecasting as a holistic, interconnected puzzle.
Addressing the Complexity of Financial Markets
Most existing financial AI models suffer from "tunnel vision." They often focus on a single company in isolation or rely on a single type of data (like just prices or just news). The authors identify three major hurdles:
- Data Heterogeneity: Stock prices are updated daily, but news arrives at irregular intervals. Fusing these different "cadences" is technically challenging.
- Global vs. Local Noise: A macro-economic event (global) affects a company differently than a report about its specific quarterly earnings (local). Disentangling these effects is difficult.
- Inter-company Ripples: Companies are interconnected. If a major supplier goes bankrupt, its customers' stocks will likely tumble. Most models fail to capture these shifting relational dynamics.
The GAME Architecture
The core of the GAME model is its "Guided Attention" architecture, which consists of several novel modules designed to handle the multi-faceted nature of market data.
- Guided Latent Cross-Attention: Instead of simply concatenating text and numbers, the model projects both into a shared "latent space." It uses one modality to "guide" the attention of the other, helping the model ignore noise and focus on relevant features across different time scales.
- Dynamic Graph-Guided Learning: The researchers built a graph of companies based on real-world relationships (e.g., ownership, industry peers, supply chain). Crucially, GAME doesn't treat these links as static; it learns to dynamically weight the importance of these relationships based on current news and market conditions.
- Guided Global-Local Attention: The model uses a company’s specific local information to "query" a massive pool of global news, extracting only the macro-events that are actually relevant to that specific firm.
- Multitask Learning: Unlike models that only predict if a stock will go up or down, GAME predicts three things: mean returns, volatility (risk), and correlations (how stocks move relative to each other).
Datasets and Experimental Setup
The researchers evaluated GAME using a massive dataset spanning 2015 to 2019, including over 1.5 million news articles from Investing.com and Benzinga. Market data included daily price and volume for over 2,000 companies across the NYSE and NASDAQ. Inter-company links were extracted from Wikidata, covering 57 different relationship types.
Key Results and Insights
GAME consistently outperformed state-of-the-art baselines (such as TST and RSR) across all four datasets.
- Forecasting Accuracy: The model achieved significantly lower error rates (RMSE and MAE) in predicting returns and volatility.
- Synergy through Multitasking: Training the model on three tasks at once (returns, volatility, and correlation) actually improved performance on each individual task. The tasks act as regularizers, preventing the model from over-fitting to noisy price data.
- Risk Management: When used to calculate Value-at-Risk (VaR), GAME resulted in significantly fewer "breaches" (instances where actual losses exceeded predicted risk) compared to other models. For the IN-NY dataset, GAME had only 6 breaches compared to the next best model's 28.
Conclusion
GAME represents a significant step forward in multimodal financial forecasting. By using local company data to filter global news and updating relationship weights dynamically, the model captures how shocks propagate through the market in real-time. While future work could incorporate additional data sources like earnings call audio or social media sentiment, GAME already provides a powerful tool for investment support and risk management.
Summary updated