The recent PRMIA webinar on the future of risk and performance generated a number of thoughtful audience questions, several of which there was not time to address in full during the live session. The following are extended responses to five of those questions, on topics that came up repeatedly during the panel discussion and in the Q&A.
From the discussion of the webinar, we can see the general trends fall into a few categories. AI is increasingly used to parse language-related data from various sources — news, announcements, regulatory filings, earnings transcripts. It is also being applied to the detection of non-linear factor patterns that linear estimation methods may miss. A third application is in input data quality control and in the flexible summarization of model output, which can make results more accessible to non-specialist users.
It is worth noting that the factor framework itself does not necessarily need to be replaced. The estimation methodology and the workflow around the model are where most of the meaningful integration is taking place.
Clearwater’s GR8 equity factor model includes an explicit crowding factor exposure. It is derived from the product of a stock’s relative trading volume over a period and the directional return of the stock over that same period. A strong volume increase combined with a strong price increase suggests the long trade has become crowded; a strong volume increase combined with a strong price decrease suggests the short trade has become crowded.
Once stocks with high crowding exposure are identified, the next analytical step is to examine whether they cluster around other factor exposures. If a meaningful set of stocks share both high crowding and elevated exposure to another factor in the model, this is an indication that the factor itself has become crowded. This shifts the analysis from a stock-level concern to a portfolio-level one.
This construction was chosen in part because it does not depend on detailed trade flow data, which is generally limited to U.S. portfolios. The volume-and-return product is observable across most markets and provides a workable proxy for positioning intensity.
Statistical models extract both factor exposures and factor returns from the investment universe. By construction, the resulting factors do not have identifiable characteristics that connect them to economic variables with observable meaning. If statistical models are used, a secondary layer of analysis is generally required to find observable economic variables whose patterns correlate with those of the statistically derived factors.
Cross-sectional models — including GR8 — take a different approach. The factors are explicit variables tied to security characteristics and observable market behavior from the outset. This makes interpretation more direct and, in environments where regulatory scrutiny of model explainability is increasing, easier to document and defend.
There is no requirement that one approach replace the other. But the choice of model class has implications for the workflow, the documentation burden, and the degree to which the model can be communicated to non-technical stakeholders.
It is correct that short-term or transient factors will make the asset covariance matrix less static than a fixed factor set will tend to indicate. Events such as the COVID disruption, tariff-related developments, and AI-driven sector dislocations can produce return patterns that fall outside the envelope of historical sampling and distort the short-term covariance structure.
The implication is not that covariance estimation is unreliable, but that the model framework should be open to incorporating transient factors and themes as part of its architecture, rather than treating the factor list as permanent. Long-term covariance tends to remain comparatively stable, provided the framework can absorb regime-specific signals as they emerge.
Risk models are most useful when they participate in a broader conversation among stakeholders about the risk-return profile of the portfolio, and the difficulty of connecting model output to measurable business outcomes is a real one.
In the practice of our Risk-aware portfolio construction with publicly traded instruments, this transparency is in fact close to complete and has been for several decades. It is embedded directly in the algorithms designed to generate risk-adjusted returns using risk models, where the connection between factor exposures, expected risk, and portfolio outcomes is explicit by construction.
The challenge tends to be more pronounced when the same approach is extended to a broader scale — across multiple portfolios, asset classes, or business lines. There, additional frictions and constraints often prevent the benefit from being more than a linear sum of single-portfolio improvements. These are typically the result of operating across organizational silos rather than any limitation of the modeling framework itself.
One of the practical advantages of factor models in this broader context is that the same factors will tend to appear across different portfolios. They function as a kind of “risk DNA” of the total portfolio holdings, which allows top-of-the-house risk management to identify shared exposures and manage aggregate risk more effectively than is possible by treating each portfolio in isolation.
A consistent set of considerations runs through these five questions: the importance of model transparency, the need for frameworks that can adapt to evolving market structures, and the role of factor models in providing a common language for risk across portfolios and stakeholders. These are not new themes, but the rate at which markets and analytical tools are changing makes them increasingly central to how risk frameworks are evaluated.
We expect to continue this conversation in subsequent research and in upcoming sessions. Thanks to those who attended and submitted questions.
Watch the PRMIA webinar replay
Take the next step and contact us to explore how your firm can navigate the future.