Abstract
The theory of Narrative Economics suggests that narratives present in media influence market participants and drive economic events. In this chapter, we investigate how financial news narratives relate to movements in the CBOE Volatility Index. To this end, we first introduce an uncharted dataset where news articles are described by a set of financial keywords. We then perform topic modeling to extract news themes, comparing the canonical latent Dirichlet analysis to a technique combining doc2vec and Gaussian mixture models. Finally, using the state-of-the-art XGBoost (Extreme Gradient Boosted Trees) machine learning algorithm, we show that the obtained news features outperform a simple baseline when predicting CBOE Volatility Index movements on different time horizons.
Introduction
Traditional economic models often struggle to account for the psychological impact of media on market behavior. The theory of Narrative Economics posits that stories circulating in the media are not merely passive reports but active drivers of economic events by influencing investor belief systems. However, extracting actionable data from these narratives is difficult because natural language is inherently messy and filled with noise. This research addresses this gap by investigating whether quantified news narratives can predict movements in the CBOE Volatility Index (VIX), commonly known as the market's "fear index."
Methodology: Quantifying Narratives
The core challenge of this study was transforming unstructured news metadata into predictive signals. The researchers compared two distinct paradigms for topic modeling:
- Latent Dirichlet Allocation (LDA): The canonical approach, which treats documents as mixtures of latent topics, where each topic is defined by a distribution over words.
- Doc2Vec + Gaussian Mixture Models (GMM): A more modern approach utilizing neural embeddings. Doc2Vec learns vector representations of documents to capture semantic context. To handle the reality that a news story might belong to multiple themes, the team applied GMMs for "soft clustering," assigning probabilities that a document belongs to various latent themes.
To enhance these embeddings, the authors introduced a technique of ordering keywords by their total corpus frequency. This helped the model understand hierarchical relationships—placing specific terms closer to their general parent terms. These narrative features were then fed into an XGBoost model to predict VIX movements across horizons of 1, 2, 4, 6, and 8 trading days.
Data and Experimental Setup
The study utilized a novel dataset from the Financial Times News API, consisting of approximately 26,000 news articles from 2010 to 2019. Rather than using raw text, each article was represented by curated financial keywords (e.g., "Central Banks," "Austerity Europe"), which bypassed the noise of natural language and focused on high-level concepts. The target variable was the historical end-of-day closing price of the VIX.
Key Insights and Results
The experimental results provided strong empirical support for the influence of narratives on market risk:
- The 4-Day Peak: Predictive performance generally peaked at a 4-day horizon. This suggests a "permeation period" where it takes several days for media narratives to fully influence market sentiment and reflect in implied volatility.
- Embeddings Outperform LDA: The Doc2Vec + GMM approach consistently outperformed the traditional LDA model, indicating that capturing semantic neighborhoods is more valuable than simple word co-occurrence statistics.
- Performance Gains: While a baseline using only past VIX movements achieved 51.5%–56.1% accuracy, the narrative-augmented models reached up to 59.1% accuracy at the 4-day mark.
- Transient Impact: Accuracy declined significantly by the 8-day mark, suggesting that news narratives influence market participants only for a brief window before being superseded by newer information.
Conclusion
This research demonstrates that quantifying news narratives through modern embedding techniques can significantly improve market risk predictions. While the study identifies a strong correlation, the authors note that further work is needed to establish direct causality and explore the impact of sentiment analysis. By moving from raw text to structured metadata, researchers can more effectively bridge the gap between Narrative Economics and quantitative finance.
Summary updated