Hacking Insurance AI algorithms with the French regulator

Hacking Insurance AI algorithms with the French regulator

On 30th of June and 1st of July, Zelros tried the first original Tech Sprint organized by the French insurance and financial services regulator (ACPR) and Banque de France. The main goal was to explain the behavior of a credit risk scoring black box model. Twelve teams were participating at this hackathon which was both challenging and friendly. The technical level was high and we learnt a lot from every participating team. Zelros thanks the organisation which was a great success and all the teams that were very strong!
AI explanation is not easy and we will describe how our time together was.  

Why a Tech Sprint around the explainability of financial services algorithms?

First of all, regulation of Artificial Intelligence algorithms in Insurance is a rising topic.
In the European Commission latest AI regulation proposal on the 21st of April this year, the credit risk modelization has been flagged as within the highest risk, with the most strict controls to apply to. 

Also, many regulators and associations are releasing specific guidelines for the Insurance industry, like for example ACPR “Governance of Artificial Intelligence in Finance”, or EIOPA “ethical and trustworthy artificial intelligence in the European insurance sector”

With this spotlight, no surprise our French bank and insurance regulators are keen on exploring tools for more transparency and explainability.

Preparing the tech sprint

Zelros participated with a multi-disciplinary team: Marie, Julien, Antoine (Data Scientists), Nicolas (Software Engineer) and Isis (Product owner). According to us, it was a good choice. Indeed, we shared our knowledge and we created an answer that was technically interesting with a creative business approach. 

ACPR organization provided a very complete Analyst guide, which allowed us to prepare.
We box checked what we needed for D-day: 

  • The black box model was to be provided through an API > we got it covered by our Software Engineer, aka master of curl and requests
  • The consumption credit was something new to us on a Business view-point > we interviewed an experimented bank advisor, who provided us with valuable insights (such as ratio of customers getting their credit request refused: we used that information to put a threshold on the prediction feedback)
  • As data generation was mentioned in the Analyst guide as useful, we checked how to adapt an internal tool to do so
  • We prepared alternative plans, such as working on white-box surrogate models trained on generated data with associated prediction by the black-box API (at that time we expected to have only a few samples for the model)

Functionally, we went through all the documentation which was well written and insightful. Besides, we spent time focusing on the main topic: explanations. A very challenging topic which was aimed at a very diverse population. There were three levels of explanations expected: customers, advisors, auditors.

The D-Day! 

D-Day was exciting. Several models, 9 to be precise (while we had been expecting only one per team :p), needed to be analyzed and a very detailed report needed to be completed. Beyond the technical challenges that had to be met, stepping back to meet the business needs was just as necessary and challenging. We have deepened our knowledge on the explanation of AI, which certainly requires a significant technical aspect, but also an ethical and almost philosophical aspect concerning the explanation.

Here are a few takeaways from this experience: 

  • Never underestimate the contribution of a diverse skills team
  • Never forget the importance of the “human style” explanations
  • Never take yourself too seriously, stay humble.

After all, Zelros ended up 2nd out of 12 teams. We learnt such important things about different aspects of AI, much more than we expected.  
We are extremely happy to have had the opportunity to compete in this challenge. 
It was an amazing experience that we recommend.
We will definitely do it again!

Congratulations to all the participants: Quantmetry, H2O.ai, DataRobot, Tinubu Square, NukkAI, MAIF, La Banque Postale, Crédit Mutuel, BNP Paribas, GPFP, Crédit Agricole.
And special thanks for the perfect welcome and the preparation marathon done by the black box models conceptors (BPCE, Crédit Mutuel Arkéa, Société Générale et Younited Crédit) and the ACPR organization!