📈 economics

Measuring Algorithmic Monoculture in Credit Intermediation: Supply-Chain Concentration, Decision Convergence, and the Case for Reference-Portfolio Supervision

This paper addresses the lack of measurement frameworks for algorithmic monoculture in credit intermediation by distinguishing between supply-chain concentration and decision convergence, developing a model showing that shared models can destabilize credit supply when error correlation rises faster than variance falls, and proposing a reference-portfolio supervision approach to estimate error correlation while arguing against premature capital add-ons in favor of building robust measurement capabilities.

Original authors: Deepak Ranjan Sahoo, Vaishali Deepak Sahoo

Published 2026-09-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Deepak Ranjan Sahoo, Vaishali Deepak Sahoo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of banking, lenders have always faced a simple but dangerous truth: if every bank makes its decisions based on the same information, they will all make the same mistakes at the same time. This is not a new fear; it is the financial equivalent of a farmer planting a single crop across an entire continent. If a disease strikes that specific crop, the entire harvest fails, and the farmer has no backup. For decades, regulators have watched banks to ensure they do not all lend to the same companies or people, because if those borrowers fail, the banks fail together. But a new kind of risk has emerged, one that is harder to see. Instead of lending to the same people, banks are now starting to use the same computer programs to decide who gets a loan. These programs, often built on massive, shared artificial intelligence systems, can process millions of applications in seconds. The worry is that if these systems all rely on the same underlying code or data, they might all reject the same good borrowers or approve the same bad ones simultaneously, creating a sudden, system-wide freeze in credit that no single bank saw coming.

A team of researchers has set out to measure this invisible danger, which they call "algorithmic monoculture." Their work focuses on the gap between what regulators can see and what they actually need to know. Currently, supervisors can easily count how many banks use a specific cloud provider or a specific software vendor. They can see the supply chain. However, they cannot see the result. Two banks might use the exact same software, but if they tweak it differently or feed it different data, their decisions might still be unique. Conversely, two banks might use completely different software, but if they are solving the same problem with similar data, they might end up making identical errors. The researchers argue that counting the software vendors is a poor way to measure the real risk, which is not about who built the tool, but whether the tools are making the same mistakes at the same time.

To understand the stakes, the authors built a mathematical model of how credit flows through the economy. They imagined a world where banks use a shared, highly accurate artificial intelligence system. This system is better than the old, independent ones; it makes fewer mistakes overall. However, because it is shared, when it does make a mistake, every bank makes that same mistake together. The researchers found a critical tipping point. If the shared system is only slightly better than the old ones, the benefit of having fewer errors is wiped out by the danger of having all the errors happen at once. In their simulations, even a system that was fifty percent more accurate than the old methods could still make the entire banking system more unstable if it caused the banks' errors to become highly correlated. The gain in accuracy is not enough to pay for the cost of the shared failure. The system becomes more volatile, swinging wildly between lending too much and lending too little, simply because everyone is looking at the world through the same lens.

The paper then turns to the question of how to catch this problem before it causes a crisis. The authors propose a new way to measure the risk using data that regulators already collect, but in a new way. Instead of just asking banks what models they use, supervisors would give a group of banks the same set of hypothetical loan applications with known outcomes—essentially a test where the answers are already written down. By comparing how each bank's model scores these same applications, regulators could calculate how much the banks' errors move together. If the banks are all making the same wrong predictions on the same applicants, the system is in danger. The researchers discovered that this measurement is surprisingly difficult to perform accurately. It requires a large number of loan defaults to be visible in the data to be statistically useful. For small business loans or large corporate loans, where defaults are rare, the data is too thin to give a clear answer. But for everyday consumer loans, like credit cards, there is enough data to get a precise reading.

Despite having a way to measure the risk, the authors argue against the most obvious solutions. They suggest that regulators should not simply force banks to use different models or add extra capital charges based on how similar their models are. Forcing diversity could backfire, pushing banks to use worse, less accurate models just to look different, which would hurt borrowers who rely on these systems to get fair credit. Instead, the paper proposes a more practical approach. Regulators should focus on ensuring that every bank has a "fallback" plan. If a bank relies heavily on a shared, external artificial intelligence system, it must prove that it can switch to a different, independent method for a significant portion of its business if that shared system fails or starts making correlated errors. This ensures that the system has a backup that is not connected to the main risk.

The researchers are careful to note that this crisis has not happened yet. Recent surveys show that while many banks are using artificial intelligence, very few are using the most advanced, shared foundation models for their core lending decisions. The risk is prospective, meaning it is something that could happen as the technology becomes more common. The urgency of their proposal is not that the danger is here today, but that once banks fully adopt these shared systems, it will be impossible to know what the "normal" level of risk was before the change. If regulators wait until the problem is visible to start measuring it, they will have lost the baseline needed to understand how bad it has become. The paper concludes that the time to build the measuring tools is now, while the system is still changing, so that supervisors can see the shift in real-time rather than discovering the damage after the fact.

Technical Summary: Measuring Algorithmic Monoculture in Credit Intermediation

Problem Statement
Regulatory bodies (FSB, ESRB, Bank of England) have identified that the banking system's reliance on common foundation models, data vendors, and cloud platforms creates a pathway for artificial intelligence to exacerbate systemic risk through "algorithmic monoculture." However, a critical gap exists between this diagnosis and measurement. While supervisors can measure exposure concentrations (e.g., large-exposure limits), they lack a framework to measure "model shocks." Current policy conflates two distinct phenomena: concentration in the supply of model components (observable via third-party risk registers) and convergence in the decisions made by institutions (unobservable and the actual transmission mechanism for shocks). The paper argues that supply-side concentration does not necessarily imply decision convergence, and conversely, decision convergence can occur without high supply concentration. Furthermore, existing regulatory frameworks (e.g., SR 11-7, SS1/23) operate at the firm level and are blind to system-level vulnerabilities arising from correlated errors across institutions.

Methodology
The paper employs a three-stage approach: conceptual differentiation, analytical modeling, and practical estimation.

Conceptual Framework: The authors deconstruct the "model dependency stack" into four layers: Infrastructure (L1), Base Models (L2), Adaptation (L3), and Deployment (L4). They argue that concentration in L1 and L2 (cloud, pre-trained models) has low direct transmission to decision errors, whereas L3 (fine-tuning, data feeds) and L4 (cutoffs, overrides) determine error correlation. Crucially, they posit that concentration in supply is neither necessary nor sufficient for convergence in decision errors.
Analytical Model: A signal-extraction model is developed where $N$ $N$ banks assess a common pool of borrowers. Each bank's assessment error ( $\epsilon_i$ $ϵ_{i}$ ) is decomposed into a common component ( $\eta$ $η$ ) and an idiosyncratic component ( $\nu_i$ $ν_{i}$ ).
- The variance of aggregate credit supply is shown to depend on the product of the error variance ( $\sigma^2_\epsilon$ ) and the correlation of errors ( $\rho$ ).
- The model establishes a threshold condition: migration to a shared model is destabilizing if the proportional increase in error correlation ( $\rho$ ) exceeds the proportional reduction in error variance ( $\sigma^2_\epsilon$ ).
- The paper derives a "technology frontier" where accuracy gains (reduced variance) are purchased at the cost of increased correlation. It demonstrates that the privately optimal choice (maximizing accuracy) often diverges from the socially optimal choice (minimizing systemic volatility) because private actors ignore the externality of correlated errors.
Estimation Strategy: The paper proposes an estimator for the error-correlation parameter ( $\rho$ $ρ$ ) using supervisory benchmarking machinery already mandated under Article 78 of the Capital Requirements Directive (CRD).
- Instead of hypothetical portfolios, the proposal requires outcome-labelled reference portfolios (anonymized historical exposures with realized defaults).
- The core statistic is the pairwise correlation of prediction residuals ( $r_i = y - p_i$ ) across institutions.
- A Null-Normalised Monoculture Index ( $M$ ) is introduced to correct for the bias in raw residual correlations caused by shared outcome variation (the "shared outcome term"), which is particularly severe in low-default portfolios.

Key Results

Threshold Dynamics: Analytical results (Table 2) show that even substantial accuracy improvements (e.g., 50% reduction in error variance) may fail to offset systemic risk if the resulting error correlation exceeds a specific break-even point (e.g., rising from 0.15 to 0.30).
Simulation Evidence: A simulation of 30 banks over 160 periods demonstrates that a shared technology with a 64% accuracy gain (reducing error variance significantly) still increased aggregate credit volatility by 22–25% due to the rise in correlation. Tail risk (exceedance counts) increased by a factor of 2.5.
Measurement Feasibility: The raw residual correlation statistic is found to be "badly contaminated" by the shared outcome term, rendering it uninformative for low-default portfolios. The proposed Monoculture Index ( $M$ ) recovers the parameter but requires a minimum of approximately 80 expected defaults in the reference portfolio to achieve a standard error of 0.05. This implies the metric is viable for retail portfolios but effectively blind to wholesale/large-corporate portfolios where systemic risk is often highest.
Policy Interaction: The paper identifies a perverse interaction where AI compliance costs (fixed costs for documentation, governance) may inadvertently increase supply concentration, as smaller institutions are forced to purchase compliant third-party systems rather than building internal ones.

Significance and Claims
The paper makes several modest but critical claims regarding the current state of banking and the path forward:

No Current Monoculture: The authors explicitly state that the empirical case for existing algorithmic monoculture in banking is weak. Recent surveys indicate foundation models account for only 17% of AI use cases, with fully autonomous cases at 2%.
The Case for Measurement: The primary argument is not that monoculture is currently a crisis, but that the measurement apparatus must be built before adoption becomes widespread. Waiting until the exposure is visible will leave supervisors without a pre-adoption baseline, making it impossible to quantify the increase in risk.
Policy Recommendations:
- Against Capital Add-ons: The paper argues against capital add-ons scaled to model commonality, as the necessary parameters are currently unmeasurable in key portfolio segments, and such measures might penalize accuracy.
- Against Diversity Mandates: Hard mandates to use different models are rejected due to distributional costs (harming borrowers with thin files who benefit from advanced models) and the difficulty of enforcing diversity without degrading accuracy.
- For Reference-Portfolio Supervision: The paper advocates for repurposing existing Article 78 benchmarking exercises to use outcome-labelled portfolios to measure the Monoculture Index.
- For Fallback Capacity: A narrower requirement is proposed: institutions relying on shared external models must maintain a verified, uncorrelated fallback channel capable of handling a defined share of volume (e.g., 25%) to mitigate correlated failure risks.

Conclusion
The paper concludes that while the sign of the effect of model commonality depends on the trade-off between accuracy gains and correlation costs, the asymmetry (correlation can multiply, accuracy cannot improve infinitely) suggests a structural vulnerability. The immediate priority is not to ban monoculture but to operationalize the measurement of error correlation using existing supervisory tools, ensuring that regulators can detect the transition to a monoculture state before it becomes irreversible.

Technical Summary: Measuring Algorithmic Monoculture in Credit Intermediation

More like this