On March 31, Agio Ratings will release version 3.2 of its crypto exchange risk model, the most significant update since the CEX risk model launched in 2022. The Agio Ratings model scores centralized exchanges on their probability of default, using 17 risk indicators to produce a single statistical credit rating for each exchange. The changes in v3.2 make the model more predictive and better calibrated, reflecting three and a half years of new data on how exchanges actually fail.
Here's what's changing and why.
What Are Agio Ratings CEX Ratings?
Agio Ratings produces quantitative credit ratings for centralized crypto exchanges, estimating each exchange's probability of default over a one-year horizon. The Agio Ratings model combines on-chain data, operational metrics, and regulatory signals into a single score, applying the same statistical framework used in traditional credit risk assessment to the digital asset market. For institutions, trading firms, and risk teams evaluating counterparty exposure, the CEX risk model provides a standardized, data-driven measure of exchange creditworthiness where traditional ratings coverage remains sparse.
Recalibrated Default Rates
When Agio Ratings launched CEX ratings in July 2022, we were working with limited data and a market in severe downturn. We calibrated conservatively, setting an average default rate of 15.0% across rated exchanges.
The data since then tells a different story. Of the 100 exchanges we originally rated, 25 have defaulted, representing a real annualized default rate of around 7.7%. Our original ratings were roughly twice as conservative as actual experience warranted.
v3.2 recalibrates to a new average default rate of 10.1%, still above the observed 7.7%. We used the 95th percentile of our statistical estimate of the true rate, a standard assumption in traditional credit markets. The result is a meaningful reduction that still retains a margin of safety.
The practical effect: all rated exchanges come out safer under v3.2. Coinbase, for instance, moves from a 3.99% average default rate to 1.48%, closely aligned with S&P's BB- rating for the company, which implies a 1-year probability of default of around 1.18%. The model and the market are telling the same story.
Smarter Weights in the CEX Risk Model
The second change concerns how we combine the 17 indicators that feed into each rating: variables like on-chain balance, web traffic, employee headcount, trading volume, regulatory licenses, and asset volatility.
In v3.1, we derived weights by testing each variable's individual ability to predict observed defaults. This worked reasonably well but had a structural limitation. The small default sample could generate individual variable weights but could not adjust those weights to account for overlap between correlated variables.
v3.2 takes a different approach. Rather than asking which variable best predicts default on its own, we ask what underlying risk traits the 17 variables collectively capture. The method identifies five latent dimensions in the data: size, trajectory, volatility, age, and regulatory strength. The common thread across all five is credit risk, so we extract that shared signal from each variable and weight accordingly.
The practical advantage is that these weights are derived independently of any default outcomes, making them genuinely out-of-sample estimates. That gives the model a sturdier foundation when scoring exchanges it has never seen before.
Performance holds up. v3.2's discrimination accuracy, measured by the Area Under the ROC Curve, comes in at 0.864, marginally above v3.1, even though v3.1 had the advantage of being trained on the very defaults used in the test. Traditional credit models typically range from 0.70 to 0.90, so v3.2 sits comfortably at the upper end.
Wider Spread Between Strong and Weak Exchanges
Better weights also mean more confidence in the model's ability to separate safer exchanges from riskier ones, and v3.2 translates that confidence into the ratings themselves.
If a model had zero predictive power, every exchange would simply receive the average rating of 10.1%. As predictive power increases, strong credits get pulled further below that average and weak credits get pushed further above it. In v3.2, we have raised the assumed discriminant power modestly, reflecting the improvement in out-of-sample performance.
In concrete terms, the gap between an exchange like Coinbase and higher-risk names has grown meaningfully. An exchange with strong fundamentals now receives a lower default probability than it would have under v3.1, while a weaker exchange receives a higher one. The 25% cap on default rates that existed in v3.1 can also be dropped, as every currently rated exchange sits comfortably below it.
What Stays the Same
The underlying signal remains consistent with prior versions. Aggregate ratings continue to move in intuitive directions, rising during the stress of 2022 to 2023 and gradually improving as the market stabilized. Coinbase's rating still tracks its stock price with an R² of 91%, a useful real-world check that the model captures a genuine risk trajectory.
The 17 risk indicators remain unchanged, as does the data infrastructure. What has changed is the calibration, the weighting method, and the confidence placed in the model's discriminant ability.
What Comes Next
v3.2 delivers a crypto exchange risk model that is more realistic, better grounded in observed default data, and sharper at distinguishing between lower- and higher-risk counterparties. For anyone using Agio Ratings to evaluate exchange exposure, the update means default probabilities that better reflect actual market experience and a clearer separation between the strongest and weakest names.
Two near-term developments follow. First, we plan to extend this methodology to Broker Ratings. The v3.2 approach works well for entity types where default history is limited, since it does not require a large sample of defaults to construct a principled credit risk score. Second, a v3.3 update is in progress. Planned improvements include incorporating PageRank and social media sentiment data, refining variable formulations such as the time windows used for on-chain metrics, and expanding the default test set as more historical data becomes available.
FAQ
What is improving in the newest CEX risk model?
v3.2 produces default probabilities that better reflect actual market experience. The model's discrimination accuracy (AUC of 0.864) sits at the upper end of the 0.70 to 0.90 range typical of traditional credit models, and its out-of-sample predictive foundation is stronger because weights are now derived independently of observed defaults.
When is the new model going live?
Agio Ratings will release v3.2 on March 31. The 17 risk indicators and data infrastructure remain unchanged, so the transition affects calibration and scoring rather than underlying data collection.
How is the Agio Ratings model changing in v3.2?
Three changes define the update. First, the average default rate has been recalibrated from 15.0% down to 10.1%, bringing it closer to the observed annualized rate of 7.7% while retaining a margin of safety. Second, the weighting methodology now extracts a shared credit risk signal from five latent dimensions in the data rather than relying on each variable's individual predictive power. Third, the wider spread between stronger and weaker credits means exchanges with solid fundamentals receive materially lower default probabilities, while higher-risk names receive higher ones.
