Abstract
Predicting stock market trends from news streams is a complex task due to the non-linear, temporal, and cross-topic nature of information flow. Traditional models often rely on simple word associations or sentiment analysis, failing to capture "hidden" causal relationships separated by time. We present Word Influencer Networks (WIN), an unsupervised framework that utilizes Lasso Granger Causality to map directed influence between words in massive news corpora. Our results demonstrate that WIN identifies highly specific predictive indicators and achieves stock price prediction errors two orders of magnitude lower than existing causal graph baselines.
Introduction
Stock market prediction is a classic challenge in financial forecasting. While many modern approaches incorporate financial news or social media sentiment, they typically rely on known factors (e.g., earnings reports) or simple co-occurrence associations. These methods often miss subtle, long-range temporal relationships—such as how a political shift might precede market volatility weeks later. The challenge lies in moving from mere correlation to unsupervised causal reasoning across massive, unstructured news streams.
Methodology
The WIN framework employs a multi-step pipeline to construct a sparse, directed graph of word influence:
- Informative Word Selection: The system filters noise by retaining named entities, significant bigrams (e.g., "fuel price"), and "trigger words" (e.g., "flood," "election").
- Time-series Representation: News data is transformed into normalized daily frequency signals for each word.
- Lasso Granger Causality: This is the core engine. While standard Granger causality tests if one time-series forecasts another, the WIN framework uses Lasso regression to handle a high-dimensional vocabulary (approx. 39,000 terms). This efficiently selects only the most predictive "influencer" words for a target, resulting in a sparse, directed graph representing the flow of influence.
- Topic Influence Compression: To enhance interpretability, Latent Dirichlet Allocation (LDA) clusters words into topics, allowing the model to calculate influence strength between broad categories like "Politics" and "Economics."
Performance Evaluation
The efficacy of the WIN framework was tested against price data for 10 major stocks. The model demonstrated exceptional accuracy in trend prediction, significantly outperforming the baseline CGRAPH method.
As shown in the table above, WIN achieved an RMSE of approximately 0.02, which is two orders of magnitude lower than the baseline errors (ranging from 1.96 to 5.25). This performance gain is attributed to the model's ability to avoid overfitting through Lasso-driven sparsity, focusing only on the most informative causal links.
Qualitative Analysis and Indicators
Beyond quantitative performance, WIN identifies highly specific "prediction indicators" that align with real-world contexts. For example, the model discovered that "Politics" was a primary influencer for both "Crime" and "Economics."
For individual companies, the influencers are remarkably intuitive: Facebook (FB) is influenced by terms like "unfriended" and "troll," while Amazon (AMZN) is linked to "HBO" and "cloud." Manual validation confirmed that 67% of the model's causal links corresponded to direct textual evidence, with the remainder appearing as short two-step paths.
Conclusion
Word Influencer Networks provide a robust, unsupervised method for uncovering the hidden causal structure of news. By prioritizing causality over simple association, the framework bridges the gap between disparate topics and timeframes, offering a powerful new tool for financial forecasting and trend analysis. Future work will focus on integrating semantic word embeddings and applying these networks to more granular econometric models.
Summary updated