Model Risk Has Become an Infrastructure Problem

Take four hundred firms, each with a well-run model-risk function, each failing independently about four times in a thousand periods. Now let half of them depend on one common provider that fails twice in a hundred periods. Hold the expected number of failures fixed. Under those assumptions, the probability that at least five per cent of the system fails in the same period is 2.0 per cent with the shared dependency, and did not occur once in 200,000 simulated periods without it.

That is the whole argument, and Figure 1 is its arithmetic. The mean is identical along every pair of lines by construction. Only the tail moves.

Section 01Four vulnerabilities, one theme

In November 2024 the Financial Stability Board identified four AI-related vulnerabilities with potential financial-stability implications: third-party dependencies and provider concentration, market correlations arising from common models, cyber and fraud exposure, and issues of model quality and explainability. None of the four is about whether a model is clever. Each is about what happens when many firms quietly come to depend on the same components.

Conventional model-risk management begins and ends inside the institution: specify the model, challenge its assumptions, validate its data, control its use. That discipline estimates whether a model is fit for a firm's purpose. It is silent on whether the same error can arrive at many firms at once, because it never leaves the firm.

Section 02What shared dependency does to a tail

There is no public dataset of which financial firms depend on which model, data and cloud providers. Drawing a network diagram out of invented firms would look empirical and be nothing of the kind, so Figure 1 does not attempt it. It asks the question that can be answered from assumptions alone: given shared dependency, what happens to the distribution of simultaneous failures?

The construction matters more than the numbers. For each level of shared dependency, the simulation computes the expected failure rate and then draws a comparison system with that same expected rate and no correlation at all. The two systems have identical means by design. Everything separating the solid lines from the dashed ones is correlation.

The result has a threshold quality that is worth noticing. Once the common-provider share passes roughly the failure threshold being measured, a single provider event alone breaches it, and the probability of a simultaneous failure stops depending much on how many more firms join. Adding the four hundredth dependent firm changes little; the damage was done somewhere near the twentieth.

each firm fails in a period if either
    (a) an independent failure occurs,  probability r
    (b) it depends on the common provider AND the provider fails,
        probability q, drawn ONCE per period for all dependent firms

parameters (assumptions, not estimates):
    N = 400 firms      r = 0.004      q = 0.02
    share s of firms dependent, from 0 to 1
    200,000 periods per point, NumPy PCG64 seeded 20251001

benchmark: a binomial system calibrated to the SAME mean failure rate,
           with no shared dependency. Mean identical, correlation zero.

result at s = 0.5:
    P(at least 5% fail together)   shared      0.0200
                                   independent 0.0000  (0 of 200,000 periods)

Section 03Correlation is the exposure

The correlation channel cannot be observed from one institution's inventory. The FSB notes that common models and data may increase correlation in trading, lending and pricing, with possible amplification through liquidity strain, and that reliance on specialised hardware, cloud services and pre-trained models has increased third-party dependency in a market with a small number of significant providers.

Supervision has been moving from the firm to the network in response. The FSB followed its 2024 assessment with a monitoring framework in October 2025, aimed at tracking adoption and provider concentration across the sector rather than assessing individual models. That is the right unit of analysis for the mechanism in Figure 1, because the quantity that moves the tail, the share of firms sharing a dependency, is not visible from inside any of them.

Validation answers whether your model is wrong. It cannot answer whether everyone is wrong together.

The resulting risk inventory has two layers. The first is familiar: model purpose, assumptions, validation, limits, monitoring. The second maps dependencies, substitution time, data lineage, provider concentration and failure propagation. A firm can complete the first layer immaculately and still be one of four hundred holding the same single point of failure, and nothing in the first layer would tell it so.

What Figure 1 does not do is estimate any of this for a real market. The failure probabilities are stated assumptions chosen to be legible; the dependency share is the axis rather than a measurement; and the conclusion is conditional throughout. Whether the financial sector is at ten per cent shared dependency or eighty is exactly the question the FSB's monitoring work exists to answer, and it is not answered here.

This research is analysis and commentary for general information. It is not investment advice, an offer, or a solicitation, and it contains no price forecasts. Figure 1 is simulated under declared parameters; descriptions of official work summarise the publications cited.