The aggregate net interest margin for significant EU/EEA institutions reached 1.69% in the first quarter of 2024, the highest reading recorded by the European Banking Authority in its published supervisory data series [1]. That figure represents a rise of 53 basis points from the 1.16% recorded in the first quarter of 2022, when policy rates across the eurozone sat at, or near, their lower bound. Eighteen months of ECB tightening did what it was supposed to do: it rewarded liability-sensitive balance sheets and punished banks that had loaded up on long-dated fixed-rate assets during the era of quantitative easing. The aggregate, however, now tells the wrong story.
By the second quarter of 2025, the aggregate had compressed to 1.58%, a decline of 11 basis points from the peak [2]. The ECB's rate-cutting cycle, which began in June 2024, is the proximate cause: as policy rates fall, the repricing advantage that deposit-funded lenders enjoyed during the tightening phase reverses. The EBA's Q2 2025 Risk Dashboard recorded that net interest income "declined further, reaching levels last observed in December 2023," even as return on equity, supported by fee income and cost efficiencies, edged up to 10.7% [2]. The headline NIM number is compressing. The institutions best positioned to absorb that compression are those that spent the past two years deploying machine learning in their credit pricing engines.
What the Aggregate Conceals
The aggregate figure obscures a dispersion that is, by any measure, extraordinary. In the first quarter of 2025, the EBA reported that country-level NIMs across the EU/EEA ranged from 0.90% in France to 3.37% in Slovenia [1]. At the individual institution level, the divergence is equally pronounced. ING Group reported a full-year 2024 NIM of 1.45%, down from 1.90% in 2023, as deposit repricing accelerated and the Dutch lender's relatively short-duration loan book offered limited protection against the cutting cycle [4]. BNP Paribas, whose retail operations span a broader geographic and product mix, reported a 2024 NIM closer to 2.4% [5]. Barclays UK, whose consumer lending book is concentrated in shorter-tenor products where rate resets are more immediate, sustained a retail NIM in the range of 3.0 to 3.2% through most of 2024, though full-year guidance was trimmed twice as the Bank of England's own cutting cycle gathered pace [6].
The spread between the best-positioned and worst-positioned institutions is not primarily explained by geography or balance sheet duration alone. It maps, with growing clarity, onto technology investment: specifically, onto the degree to which credit-pricing decisions are driven by granular machine learning models rather than legacy scorecard systems built on broad risk buckets.
The Machine Learning Entry Point
The ECB Banking Supervision's 2025 data collection, published in its November supervisory newsletter, documented "a strong increase in AI use cases among European banks between 2023 and 2024," with credit scoring identified as one of the two dominant applications alongside fraud detection [3]. Decision tree-based models are the prevailing architecture for credit scoring, the ECB noted, with neural networks more commonly deployed in fraud work [3]. This distinction matters for the margin question: gradient-boosted decision trees, the workhorse of most deployed credit scoring systems, allow institutions to price risk at a granularity that rule-based scorecards cannot approach. Where a traditional scorecard assigns a borrower to one of twelve risk bands, a well-calibrated ML model can generate a continuous probability of default that reflects hundreds of input variables simultaneously.
The practical consequence, visible in the margin structures of early adopters, is a bifurcation within the loan book. Algorithmically confident borrowers, those for whom the model assigns a high-certainty, low-risk probability, are priced more competitively than a legacy scorecard would permit: the institution can afford a tighter spread because it has genuinely less risk. Borrowers at the model's uncertainty boundary, where prediction intervals are wide, attract a wider spread that reflects that uncertainty. The headline NIM figure, computed across the whole portfolio, may not change dramatically. The economic value of getting the within-portfolio allocation right, however, accumulates over cycles.
Where a traditional scorecard assigns a borrower to one of twelve risk bands, a well-calibrated ML model generates a continuous probability of default from hundreds of variables simultaneously.
The EBA's own analysis of AI adoption data found that banks deploying ML credit models reported "extended lending, more effective risk assessments, better risk avoidance and lower default rates, ultimately contributing to improved profitability" [3]. The direction of the effect is not in dispute. The magnitude is harder to isolate from the rate environment: banks that adopted ML credit scoring heavily between 2020 and 2023 also tend to be institutions with stronger technology governance and risk management frameworks in general. Attributing margin outperformance specifically to the models, rather than to the broader organisational capabilities that enabled their deployment, requires care.
The Compliance Counterweight
The margin benefit from ML credit scoring does not arrive without cost. The EU Artificial Intelligence Act, which classifies credit scoring systems as high-risk AI, imposes obligations that include rigorous documentation, risk management procedures, post-deployment monitoring, and continuous log-keeping [7]. The EBA's 2025 factsheet on the AI Act's implications for banking noted that these requirements apply to any AI tool used in credit decisions, regardless of the institution's size or the loan amount [7]. Compliance overhead is real, and for smaller institutions, the cost-benefit calculation of deploying a full ML credit pipeline may only become favourable at scale.
The ECB's supervisory data also identified transparency gaps: some institutions "lack full transparency regarding the internal processes of some AI models," creating what the regulator termed potential "black box" risks [3]. Only a minority of banks reported "effectively applying data management standards in practice," a finding the ECB identified as a critical vulnerability [3]. Poor data inputs, the regulator noted, "will inevitably lead to unreliable results." An institution that builds a margin advantage on a model trained on biased or incomplete data is not constructing a durable competitive position; it is accumulating model risk that will eventually surface in credit losses. The institutions that understand this, those that invest in data governance as seriously as they invest in model architecture, are the ones whose NIM outperformance is worth watching.
The aggregate 1.58% NIM figure in the EBA's Q2 2025 data is the headline. The story underneath it is a widening structural gap between institutions with sophisticated, data-governed ML credit pricing and those still operating on legacy scorecards. As the rate environment normalises further and the rate-cycle tailwind disappears entirely, that gap will become the primary explanation for diverging profitability outcomes across European banking.
- European Banking Authority. "Q2 2025 Supervisory Data Indicate Improvements in ROE Despite Continued Tightening of Net Interest Margins in EU/EEA Banks." EBA Press Release. 2025. eba.europa.eu
- European Banking Authority. "EBA Risk Dashboard — Q4 2024." EBA Supervisory Data. March 2025. eba.europa.eu
- ECB Banking Supervision. "AI's Impact on Banking: Use Cases for Credit Scoring and Fraud Detection." SSM Supervisory Newsletter. November 2025. bankingsupervision.europa.eu
- ING Group. "ING Posts Full-Year 2024 Net Profit of €6,392 Million." ING Press Release. February 2025. ing.com
- BNP Paribas. "Fourth Quarter and Full-Year 2024 Results Press Release." BNP Paribas Investor Relations. 4 February 2025. cdn-group.bnpparibas.com
- Barclays PLC. "2024 Results Announcement." Barclays Investor Relations. 12 February 2025. home.barclays
- European Banking Authority. "EBA Factsheet: AI Act Implications for the EU Banking and Payments Sector." EBA Publications. 2025. eba.europa.eu
- European Banking Authority. "First-Quarter of 2025 Supervisory Data Shows That the EU/EEA Banking Sector Remains Robust, Despite Increased Cost of Risk." EBA Press Release. 2025. eba.europa.eu