This sessional will walk through the working party’s recently highly commended paper for the 2024 Brian Hey prize. The meeting will show how it has applied Google’s federated learning algorithm (2016) for text prediction on smart phones to insurance modelling.

This concept enables the direct training of machine learning models on users’ devices, such as smartphones. It eliminates the need to share or transfer potentially sensitive data to a centralised server.

Unlike traditional machine learning methodologies, federated learning adopts a model where the algorithm is brought to the data, rather than transferring the data to the algorithm. It will be hugely important as AI becomes more commonplace.

In our paper and sessional we show how insurance companies can collaboratively develop a neural network model to predict claims frequency specifically. We achieve this using the Flower package in Python along with PyTorch. We show that if companies cannot share customer data, they can achieve near double their model predictive performance by using federated learning while still keeping their customer data secure.

The working party is part of the IFoA Data Science and AI practice area.