Written by Marino Merzouk San Lorenzo
Model complex non-linearity
AI/Deep learning is being successfully used in the insurance industry. One example is the application of Neural Networks (NN) in pricing insurance contracts. They perform well at modeling complex non-linearity and interaction between continuous and discrete features (driver age, mileage limit, geographic postcode, coverage type and so on…). Why are NN best suited for the job? Because they are able to model complex interactions whilst also been able to generalize.
Extracting relevant features and enhancing pre-processing through embeddings
No pain no gain though, ad-hoc data pre-processing is still required. Using embeddings layers significantly improves the model accuracy and speeds up feature selection (the process of choosing which data points should be used in the model). The NN will learn what the most relevant features from policyholders’ information are in an unsupervised manner. By changing representations, it will simplify the presumably complex relationships that exist in the dataset whilst best approximating your initial input data:
As an example, the Peugeot Partner and Citroën Berlingo look alike although a naive statistical model would treat them as two sparse separate vectors. The NN will understand that they do not differ by considering them within same component whilst decreasing noise.
Learning the right risk exposure of the policyholder
Moving back to our initial objective: establish whether a policy is less or more likely to generate a claim. To discriminate and “classify” our policies by risk level, the NN will regress a counting Poisson parameter by policy. The loss function, which it will aim at minimizing, is thus a Poisson deviance given the exposure of the policyholder (period of the contract in force). However, the neural network’s flexibility allows learning a risk exposure specific to the policyholder.
The intuition behind is the following: a self-employed or transport driver relying heavily on his company car faces different exposure towards traffic risks compared to a retired driver or even an “occasional user”. Again, the NN will understand these effects and modify the initial exposure accordingly.
Refine risk understanding and learning from the network
Actuaries can overcome interpretability drawback by using Tensorflow API, benchmarking results and variable importance with other statistical models and draw on their insurance business knowledge. AI can help us refine our risk understanding the same way Alpha Zero or Deep Blue allow go and chess players to discover new tactics and strategies to bring their game to the next level.
AI is (hopefully) coming… for actuaries as well
To conclude, NN are powerful and flexible algorithms that allow the most relevant features in your portfolio to be modelled and identified. Moreover, NN can be applied to pricing other Lines of Business’ contracts. For all the above-mentioned reasons, more and more actuaries are starting to explore the use cases for Deep Learning. Data, computing power and algorithms are getting better and better, dramatically improving risk assessment and model accuracy using NNs.
References on the model used:
· Data Analytics for Non-Life Insurance Pricing, Mario V. Wûthrich, 2018
· Case Study: French Motor Third-Party Liability Claims, Alexander Noll, Robert Salzmann & Mario Wüthrich, 2018
· Insights from Inside Neural Networks, Andrea Ferrario, Alexandre Noll & Mario V. Wûthrich, 2018
· AI in actuarial Sciences Ronald Richman, 2018
Iconography: