Is there such a thing as a ‘Crisis’ factor?
Authors
Arnab Banerjee PhD
Axioma Analytics Solutions
Robert D Stock PhD
Axioma Analytics Research
Uncovering hidden systematic structure in market turmoil
Following the publication of an article in Hedgeweek on ‘Detecting hidden crisis factors,’ we sat down with former Quant PM and current research lead Bob Stock and Axioma risk model product head, Arnab Banerjee, to dig a bit deeper into their findings on a transient crisis factor. The study challenges fundamental assumptions in equity factor modeling and demonstrates that there is hidden systematic structure that emerges during market stress periods.
At the heart of this research is Comparable Company Specific Covariance (CCSC), an existing Axioma risk model methodology that uses machine learning clustering to identify companies with highly-similar business models within the same industries. Traditional models assume that specific returns – the portion of stock returns unexplained by common factors – have no correlation with each other, but CCSC identifies where this assumption is violated. The team modified this approach to detect short-term patterns that other traditional models have failed to catch.
You challenge the basic premise of how in a crisis, risk models normally show that ‘correlations go to one’. Why are you not convinced?
Arnab: If that assumption were accurate during crises, we'd also expect to see residual correlations intensifying broadly across the market, causing the number of distinct stock clusters to shrink as everything converges into a single correlated group.
But you didn’t observe that?
Bob: No we didn’t. This is because conventional wisdom is oversimplified. Yes, correlations increase, but we actually observed more clusters of correlated stocks forming, not fewer. Instead of everything converging into one big cluster, we see crises creating multiple distinct groups with each internally correlated but behaving differently from other groups.
Can you walk us through what’s actually happening?
A: Think of it this way: during a thematic crisis like COVID, airline stocks move together, cruise lines cluster, and work-from-home technology stocks form yet another group. Within each cluster, the specific correlations spike as investors stop differentiating between similar companies. But airlines still behave very differently from tech stocks. So you get more clusters, not convergence.
You had to modify your existing CCSC methodology to detect these patterns, correct?
B: Yes, our standard CCSC uses three-year data windows and updates quarterly because it's designed for stability. In this study, we needed responsiveness to catch transient effects, so we shortened the correlation window to three months and computed weekly. This let us detect mild, short-term crisis factors that the standard approach would miss completely.
You studied ten specific crises. Which ones provided the clearest signals?
A: The Subprime Crisis was our strongest case: clusters increased 53%. Liberation Day tariffs showed a smaller but rapid 13% rise. COVID registered a 22% gain even in standard CCSC. But six of our ten crises showed increases above 20%, suggesting this is a real, systematic phenomenon.
This has significant implications for risk management. Can you quantify the practical impact?
B: When crisis factors emerge, your risk model's diagonal assumption for specific returns can be moving away from the truth. You might think you're diversified when you're actually concentrated in a hidden systematic factor. Conversely, if you can identify these patterns early, you can position accordingly.
How quickly do these patterns emerge and disappear?
A: That's the key insight: they're transient. We tracked changes over ten-week periods around events. Some patterns emerge rapidly, like the tariff response, while others build more gradually. The temporary nature means traditional long-horizon models will always miss them.
If you’re a fund manager, how would you operationalize these findings?
B: First, enhanced risk monitoring: when your modified CCSC signals spike above 20%, reassess your portfolio's hidden systematic exposures. Second, tactical positioning: You can amplify (or not!) these temporary correlations depending on your strategy. Third, you can look for pair trading opportunities within and across the emerging clusters.
Any final thoughts for risk managers and portfolio managers?
A: Residuals aren't random noise. They contain valuable systematic information, especially during volatile periods. Traditional factor risk models, while useful, can be incomplete, since, by design, risk models are parsimonious. Managers who can harness this hidden structure will have a significant edge in both risk management and alpha generation. The tools exist; it's about having the awareness to deploy them effectively.
You may also like
