7 surprising stock correlation effects revealed by machine learning

Sorry Mark, Facebook and Google are the same.

Standard risk models assume that once you account for common factors like industry and size, the remaining company-specific (idiosyncratic) risks should be uncorrelated between different companies. But what if that assumption is wrong? Our research shows that ignoring these hidden correlations means you could be dangerously over- or under-estimating your risk.

This should concern anyone managing equity assets. For example, it is critical for hedge funds to identify the idiosyncratic correlation between seemingly different companies for accurate risk prediction and effective hedging. For wealth managers, this data can be used to better hedge concentrated positions or for finding tax-loss harvesting substitutes. 

What exactly are these hidden correlations and why do they matter? 

Consider this: You have two portfolios, one containing only Mastercard and the other only Visa. Of course, due to their overlapping risks from common factor exposures (similar size, same industry, etc.), factor models should predict a relatively modest tracking error between the two portfolios. In practice, however, due to their very similar business models, the actual tracking error will be even lower because the model assumes that the idiosyncratic risk correlation is zero. And if they are in the same portfolio, the actual idiosyncratic risk will be much higher than the standard model predicts.

Our approach to capturing this additional risk is through a technique called Comparable Company Specific Covariance (CCSC), which builds upon our established Issuer-Specific Covariance (ISC) process. While ISC already links securities from the same issuer, CCSC extends this concept by using machine learning (specifically hierarchical agglomerative clustering) to identify which seemingly unrelated companies exhibit surprisingly correlated idiosyncratic movements.

Instead of adding numerous unstable sub-industry factors, CCSC dynamically identifies meaningful linkages while maintaining distinct clusters, giving portfolio managers a more accurate picture of their true risk exposure.

So which stocks end up being comparable companies?

We applied the CCSC process, available in the latest Axioma Worldwide Equity Factor Risk Model (v5.1) and analyzed over one billion possible security pairs which revealed these fascinating insights:  

Insights from different risk model variants  

We also examined the clusters from the short horizon (1-2 months) model alongside the statistical model. The short horizon produced similar clusters but fewer in number, while the statistical model produced more clusters and often different ones (for example, pairing Intel with Micron, and putting Nvidia and AMD with the other equipment & services companies like KLA and LAM). For the statistical model, this is to be expected as it has a very different factor structure from the fundamental model. Indeed, lacking explicit industry factors, the statistical model essentially recreates more of them through CCSC, with an extra 5,000 company pairs and over 500 additional clusters.

By finding stocks that behave similarly, portfolio managers can improve their forecasts, and ultimately, their performance.  

CCSC is available to users of the latest version (5.1) of the US and Worldwide Axioma Equity Factor Risk Models. The company linkages are dynamic and vary across time, adapting to market circumstances and are updated quarterly. For more information, contact us below.