Commercial insurance underwriting has traditionally been done on a case by case basis resulting in inefficiencies and lengthy communication between insurers and clients before the final policy is issued. This is a problem which AXA has been actively looking to solve to make it easier for our clients to do business with us.
Growing access to richer datasets has given rise to statistical techniques that enables, machine learning where our systems are able to make data-driven predictions based on repetitive learning algorithms.
AXA has assessed that machine learning is estimated to reduce the underwriting process significantly for customers and save valuable man hours of underwriters to deal with more complex risks.
We set about this journey by asking ourselves if it was indeed possible to replicate pricing decisions traditionally undertaken by underwriters and automating this process using machine learning models and on demand datasets accessed via APIs. These predictive underwriting models would help to improve efficiencies and free up underwriters from having to take pricing decisions for standard risks enabling them to focus on higher value opportunities.
Behind the customer front end lies a complex machine learning engine developed on AXA’s Cloud-based platform where Python scripts transform raw data into transversal tables used to create the training set needed to develop our machine learning model that is used in the predictive underwriting.
This data, which includes information about the policy (duration, coverage type, risks insured, sum insured for the different risks, contract type, number of risks insured, etc.), the company (age of the company, activity sector, number of officers, etc.) and the claims history of the company if any (occurrence date, sum claimed) are used to predict the most accurate premium based on the company’s risk.
During the training of the model, AXA’s underwriting team provided the “confidence level” for each suggested premium predicted by the machine learning models. This confidence level is defined as the “certainty index” for every prediction made by the models and helps to train the model to provide more accurate premium predictions.
The above figure shows the results of the machine learning model. The red dots are the accepted premiums predicted by the model after checking the confidence level. Blue dots are rejected predictions and green dots are the real premiums of the Bond policies. This confidence level allow us to reduce the risk to predict a wrong price. After applying the confidence level with a threshold at 0, the error on predicted premiums decreases significantly.
Aided by this new tool, AXA has been able to reduce the time and effort spent by our underwriters to accurately price risk premiums for our commercial customers.
The full whitepaper on how the machine learning model was developed can be viewed here.