Optimizing Pricing in the Insurance Industry : Towards transparent Machine Learning

In the last two decades, Big Data has brought revolution to the long-standing approach and increase its efficiency by ten-fold. The availability of large-scale datasets, and the advent of artificial intelligence (AI) and machine learning (ML), are offering powerful capacity to classify risks and produce more accurate predictive pricing models. However, the opacity of current ML algorithms, often described as ‘black boxes’, is posing a challenge to the industry. “Conventional ML techniques rarely provide explanation on why a client is considered as lower risk than another. This makes them less reliable, less flexible, and less actionable”, explains Guillaume Beraud-Sudreau, Head of R&D from AXA Global Direct and now leading the actuarial team at Akur8. Together with Prof. Alexandre d’Aspremont, from École Normale Supérieure, they have decided to carry out a collaborative project to design robust and ‘transparent’ algorithms capable of addressing the complex subject of insurance pricing.

The previous JRI focused on prediction models regarding the likelihood of a website visitor to sign a contract based on different factors, namely behavior on the site, answers to a questionnaire, and proposed price. “People have been making predictions like this for a long time. But in our case, we wanted to see what the drivers were to the decisions. This is crucial because we wanted to make a ML system whose decision is easy to understand, so that actuaries can actually use it and incorporate their expertise in the model”, explains Prof. Alexandre d’Aspremont. “A key component of the project was to test and implement these new algorithms that could be included in a core open source ML library called SCIKIT-LEARN. This was an important academic contribution. It also has direct industrial applications as AXA uses this software.” Thus the new JRI aims to convert the insight gained from these new prediction models into the design of new pricing models.

Opening the black boxes

The current research focuses on integrating these advances into a longer and operational value chain to exploit them for actual pricing and overall profit and loss. “What we did in the first JRI is going to be a block of what we do in the second one. The algorithms we’ve developed so far are intrinsically local in the sense that they can only model the probability of one prospective client to sign the contract. Optimizing prices solely based on individual decisions is of little interest for an insurer. Our objective now is to understand how to optimize all price proposed to a pool of prospects, while managing other key aspects of contracts and enforcing advanced constraints on overall sales, risk, profit, and interpretability”.

The added value of this partnership will be to test the theory on real-life data. “At AXA, we hold a lot of client information, so we have the resource for accurate predictions. It’s really gratifying for an actuary like myself to have the opportunity to contribute to the actuarial theory ”, comments Guillaume Beraud-Sudreau. For Prof. d’Aspremont too, the bilateral approach provides something of great value: “witnessing how deep theory research is translated into applications is always very gratifying.”

Today it is pretty standard to make predictions with machine learning. But making the predictions explainable is highly innovative. Transparency inherently offers more reliability, flexibility and thus, more actionability. In this sense, the accumulated output of both the JRIs mentioned above will contribute to unlocking the potential of machine learning in the insurance sector, as well as for the policy holders, with the promise of more accurate and fairer premiums.

* Kamet Ventures is a €100m InsurTech incubator dedicated to conceptualizing, launching and accompanying disruptive products and services for insurance clients.

Optimizing Pricing in the Insurance Industry : Towards transparent Machine Learning

Related articles

Addressing Misinformation on Natural Hazards in Social Media (NatHazFake)

Université Paris Cité

Confronting the Challenges of AI-Generated Misinformation

University of Western Australia

Trust in the Fediverse: Leveraging Community Protocols and Automation to Combat Online Harms

Queensland University of Technology