A personal blogby Fiona Edocket
Short reads on emerging and current news.
Can a bank adopt a credit bureau's new score before Germany's scoring rules change on 20 November 2026?
From 20 November 2026, German law prohibits the use of address data, age and other categories in credit scores, and gives lenders new duties when they decline an application. Banks rely on scores they did not build, yet remain accountable for them.
A Python pipeline: an evidence register of 49 sources, a tested points calculator, decision rules applied in code and an outcome test kit run on 303,949 loans.
Not yet approvable, with a clear path to approval. Two pieces of evidence decide the outcome, and the tests to apply to them are already built.
AI powered cycle tracking for earlier PCOS risk awareness. CycleSignal brings together a wearable, a short monthly check-in and two machine learning models to spot a concerning pattern early, explain why in plain words, and suggest when to see a GP. It is not a diagnosis.
Polycystic ovary syndrome affects about 1 in 10 women of reproductive age, and the WHO estimates up to 70% are undiagnosed. Each symptom is easy to explain away, and a GP sees a snapshot rather than months of cycles.
A Streamlit prototype that combines a wearable cycle model, a risk model trained on self-reported answers, and a Claude-based assistant that explains the score and suggests when to see a GP.
Test ROC-AUC of 0.92 and recall of 0.81. An ovulation detector failed its pre-set checks, so temperature is shown as context only. The app never diagnoses.
"Our best model reached a test ROC-AUC of 0.92, but most of our design time went on deciding what the app should not claim."From the team's project write-up
The same discipline as a model review: pass and fail criteria written down before testing, a failed detector demoted rather than shipped, model quirks labelled instead of hidden, and an assistant that states what the app cannot see.
I work in AI Governance, Technology Risk and Regulatory Compliance, with a background in financial services.
I started in banking compliance and accounting, where I got interested in how organisations manage risk and keep up with regulation. Since then I’ve trained as an ACCA accountant and completed an MSc in Economics and Financial Technology (FinTech) at the University of Galway.
I’m now Co-Founder of AIRIS.ai, a platform that helps organisations meet their obligations under the EU AI Act.
My work covers the EU AI Act, DORA, GDPR, NIS2 and standards such as ISO 42001. I’m most interested in how organisations turn regulatory requirements into governance processes that actually work day to day.
To me, regulation isn’t the enemy of innovation. Done well, it’s the middle ground where new technology earns people’s trust, and I want to help organisations get there.
Everything here is my personal view, written to explore ideas rather than to advise.
Short updates on regulation and model risk, between larger projects.
What a bank must demonstrate before adopting a credit bureau's new score, and how much of that the public record supports.
On 20 November 2026, a new provision of Germany's Federal Data Protection Act, §37a BDSG, takes effect. It prohibits the use of age, gender, name, bank account transaction data and address data, among other categories, in credit scores. At the same time, an amended §30 BDSG requires lenders that decline an application on the basis of bureau information to tell the applicant more about that decision. In October, the Federal Court of Justice also clarified when a bureau must explain an individual score.
Most banks do not build their own credit scores; they license them from a bureau. That raises a practical question for any lender: what must it be able to demonstrate before adopting a new bureau score, and how much of that can be verified from public information alone? This case study answers that question for one score.
In March 2026, a large German credit bureau replaced its existing scores with a single score based on twelve published criteria, on a scale of 100 to 999. Every lender that uses it must decide whether to adopt it and how to recalibrate its cut-offs.
The review uses a fictional lender, Mainkai Direktbank: a Frankfurt-based direct bank that would use the score to approve, price and size unsecured personal loans across roughly 120,000 applications a year. The bureau is real but is not named.
Under German supervisory rules, responsibility for a vendor model stays with the bank. MaRisk AT 4.3.4 extends model governance to models developed by third parties, with requirements on suitability, data quality, validation and explainability. BTO 1.2 Tz. 5 further requires a bank to form its own view of credit risk rather than rely solely on external assessments.
This bureau is unusually transparent: it publishes every criterion and the points assigned to each. That makes it a useful test of a common assumption in vendor oversight, that a transparent model is also an adequately evidenced one. The bureau publishes no model card, validation report or outcome data, and the review examines what that gap means for a lender.
The review was built in four parts.
§37a(2) No. 1(d) prohibits the use of address data ("Anschriftendaten") in scores. One criterion, time at the current address, contributes up to 94 points. On a broad reading, residence time is derived from address data and falls within the prohibition. On a narrower reading, the provision targets location-based scoring, and residence time is a measure of stability. No court has yet ruled on the point.
Time-dependent criteria account for 280 of the 999 points. An applicant with a clean payment record and no mortgage can reach at most 743 after one year in the German credit file, which falls in Class 2. Reaching Class 1 takes between two and five years; the published table does not support a narrower estimate.
X axis: years in the credit file. Dashed lines: Class 2 from 709, Class 1 from 776.
Lower scores for short credit histories are expected by design. The relevant question is whether the score is calibrated within these groups: whether a given score implies the same default risk for younger applicants and recent arrivals as for the wider population. No public evidence addresses this.
In the calculator, a 60-month Mainkai loan reduces the borrower's score by 78 to 88 points through three criteria. This has implications for top-up lending and for how customers are told about the effect of loan term on their credit standing.
On the Bondora data, rank ordering failed at the riskier end of the scale, where an automatic decline cut-off would typically sit. In 2025, grade F (17.6% default rate, 1,309 loans) defaulted less often than grade E (19.6%). Seven of eight grades deteriorated against the reference period, while the overall default rate held broadly stable (8.3% against 8.6%) because the mix shifted towards the best grade. Across customer segments, observed defaults ranged from 0.72 to 2.34 times the level implied by the grade.
From 20 November, the amended §30(7) introduces two notification duties. Where a consumer loan is declined on the basis of bureau information, the lender must tell the applicant what that information was, which database it came from and which data categories were used. Where the decision is automated, the lender must also explain the right to human review and how to contest the decision. Failure to inform is an administrative offence under §43(1) No. 2. This finding is rated High.
On public evidence alone, the score is not yet approvable. That is not a judgement against the score. The review applies a ratability test in code: a core element can be rated only where a result is visible in the evidence, such as a reported metric, a published table or a stated review conclusion. A description of a process is not sufficient. Data, methodology, testing and external review each meet this test. Outcome analysis on the bank's own applicants does not, so the decision rules return "defer pending evidence".
The more useful output is what would change that outcome. Running the same rules on two hypothetical scenarios shows that the decision turns on two pieces of evidence:
Outcomes on the bank's own applicants cannot be rated.
Defer pending evidenceThe new score shows no material loss of discrimination against the incumbent score on the same applicants.
Approve with conditionsThe bureau keeps it, and the bank's counsel treats it as address data under §37a.
RejectBoth are specific and obtainable. The first is a standard request in a vendor model review: the bureau scores 24 to 36 months of the bank's past applications with both the new and the current score. The second is a written position from the bureau and a legal opinion from the bank's counsel. Both are set out in the accompanying evidence request, as items E02 and E01.
The outcome tests do not need to be designed once the retrospective score arrives. They are built and run end to end on the Bondora data, so the same code can be applied to the bank's own file:
When the evidence arrives, the review resumes from a defined point: run the tests, re-rate the outcomes element, and let the same rules return the decision.
| Initial approach | Revised approach | Rationale |
|---|---|---|
| A fixed threshold of 0.55 Gini on the bank's loans | No loss of more than 3 Gini points against the incumbent score on the same applicants | A Gini measured on one lender's applicants is lower than a bureau-wide figure by construction, so a fixed threshold can reject a sound score |
| Reference period 2023 Q3 to 2024 Q4 | 2024 Q2 to Q4 | The longer window combined two rating regimes and overstated drift by about a third |
| Figures entered manually in the report | All figures generated from the data at build time | An internal review found three figures that no longer matched the data |
The review is a reproducible Python project rather than a document. Every figure, rating and recommendation is produced by code from the evidence and the data, so a change to a source or a threshold flows through to the report automatically. The pipeline runs in six stages:
# src/assurance/models.py
@model_validator(mode="after")
def _ratability_test(self):
if self.core and self.ratable != bool(self.visible_result.strip()):
raise ValueError(f"element {self.element}: core elements are ratable "
"if and only if a visible result is recorded")
return self
[[ceiling_12m]] that are filled from the lab outputs at build time, and the build fails if any token is left unfilled.Vendor model review against MaRisk AT 4.3.4, SR 26-2 and PRA SS1/23; out-of-time testing; champion and challenger design; tolerance setting.
Evidence requests, an obligation matrix, an outsourcing assessment and a decision memo with conditions.
Working from statute and case law, weighing competing interpretations, and translating legal duties into testable controls.
Python, statistical testing, reproducible pipelines, automated tests and document generation.
Absolute thresholds are unsuitable for vendor scores. The bureau reports a Gini of around 0.6 across its full population. On a single lender's applicants the figure will be lower regardless of model quality, so the appropriate test is a paired comparison with the incumbent score on the same applicants.
Drift measures depend on a stable baseline. The original reference period combined two rating regimes, which became apparent only when the data were examined quarter by quarter.
The German changes take effect on the same day that the revised EU Consumer Credit Directive (Directive (EU) 2023/2225) begins to apply across member states, 20 November 2026. An Irish lender adopting a vendor score would face many of the same questions under a different supervisory framework. The method transfers unchanged; the legal anchors do not.
| Topic | Germany (this review) | Ireland |
|---|---|---|
| Data that may not be used | §37a BDSG: age, gender, name, account transactions, address data and other categories | CCD2 Art. 18: no special-category data and no social-network data, and the assessment may not rest on credit history alone. The address question is specific to §37a. |
| Explanation and human review | §30(7) BDSG notice duties when a loan is refused | CCD2 Art. 18: where the assessment is automated, a right to human review and a clear explanation |
| Status of the rules | In force from 20 November 2026 | CCD2 takes effect through national law. At the time of writing I could not confirm that Ireland's law has been published. |
| Credit data | Private bureau score, 100 to 999 | The Central Credit Register reports credit data but produces no score; any score comes from a private provider or the lender itself |
| Supervisor's rules | BaFin MaRisk AT 4.3.4, AT 9, BTO 1.2 | Central Bank of Ireland: revised Consumer Protection Code in force since 24 March 2026; outsourcing guidance and DORA for third-party and ICT arrangements |
| AI Act | Credit scoring is high-risk under Annex III 5(b) in both countries. Following the AI Omnibus, these obligations are reported to apply from 2 December 2027. | |
In an Irish version of this review, the address finding would fall away, the notification finding would be re-based on the directive's rights to explanation and human review, and the governance findings would be mapped to Central Bank expectations rather than MaRisk. The ratability test, the champion and challenger comparison and the segment calibration tests would apply as they stand.
Thanks for reading. Questions, or a different view? I would like to hear it on LinkedIn.