Enterprise AI Analysis

The methodology involves constructing and evaluating clusters based on different similarity measures and algorithms:

• Similarity Measures:
• Correlation of Historical Returns: Uses daily stock return co-movements from Compustat, computed over rolling time windows.
• Human Insight (GICS): The de facto standard classification by S&P and MSCI, categorizing stocks into 11 sectors based on business activity.
• LLM Embeddings of News Headlines: Utilizes OpenAI's text-embedding-3-large model on RavenPack news headlines, creating 3072-dimensional embedding vectors to capture semantic relationships.
• Random Baseline: Stocks arbitrarily assigned to clusters as a performance floor.
• Clustering Algorithms:
• k-means clustering: Partitions firms into k clusters by minimizing the within-cluster sum of squares (WCSS).
• Hierarchical clustering (Agglomerative): A bottom-up approach merging clusters based on smallest inter-cluster dissimilarity.
• Evaluation Framework: Each clustering method generates 11 clusters. These are converted into synthetic factor models using the Arbitrage Pricing Theory (APT) framework. Daily return series are constructed for each cluster (equal-weighted average). A linear model is estimated weekly with one month of daily data. Out-of-sample prediction errors (RMSE and MAE) are then computed.

The empirical results, derived from rolling-forward experiments on S&P 500 constituents from 2022 through 2024, provide clear insights into the predictive performance of different clustering methods.

• Dominant Performance of Price-Based Clustering: Price-based clustering methods consistently outperformed both LLM embedding methods and GICS.
• Quantified Superiority: Price-based clustering reduced the Root Mean Squared Error (RMSE) by 15.9% relative to GICS and 14.7% relative to LLM embeddings.
• Statistical Significance: A paired Student's t-test comparing the best-performing returns-based clustering with the best-performing LLM-based clustering yielded a p-value of 0, indicating that the difference in performance is statistically significant.
• Reinforcement of EMH: The findings reinforce the view that the dominant information content for equity pricing is already embedded in co-movements of returns, supporting the semi-strong Efficient Markets Hypothesis.

Human judgments and sophisticated NLP extractions from news text did not translate into superior short-term predictive factors compared to market-driven price signals.

The research has significant implications for both financial practitioners and academics:

• For Practitioners: The proposed framework offers a practical diagnostic tool for monitoring evolving sector structures. It allows for computing rolling correlation clusters, constructing equal-weighted synthetic portfolios, and tracking their incremental explanatory power relative to existing sector definitions. This can guide portfolio construction, risk modeling, and strategic asset allocation by identifying the most effective drivers of cross-sectional return variation.
• For Academics: The methodology provides a robust framework for testing alternative hypotheses about how quickly markets absorb information and what types of information are most efficiently priced. It opens avenues for further research into market efficiency and the utility of alternative data sources in financial modeling.
• Support for EMH: The empirical evidence strongly suggests that for short-horizon returns, the market efficiently incorporates public information, making price-based signals highly effective.