09:50-10:00: Globally consistent natural hazard maps using deep neural networks
by Roland Schöbi
Insurance companies are keen to model their natural hazard-related risks accurately and consistently around the globe. The increased availability of globally consistent open-source datasets in combination with existing in-house data sources enables the creation of hazard maps that are globally consistent and reflect the company’s internal risk view. The ability of deep neural networks to model complex physical phenomena (i.e. nonlinear behaviour) facilitates the mapping between the open-source datasets and internal risk view. The proposed approach is discussed on the application of an earthquake and windstorm hazard map, which are currently in production for pricing and underwriting risk selection in the broader company.
10:00-10:10: Natural Catastrophe Insight - Deep Learning and Satellite Imagery to the Rescue
by Patrick Jayet & Paul Feng
How can we leverage Computer Vision and the ever-growing availability of aerial imagery to help perform an automated assessment quickly after a natural catastrophe? In this presentation we will share the work we have done on automated flooding detection using Deep Learning applied to satellite images. We will present various approaches that we tried for tackling this problem, show their results, pros and cons as well as share various problems we faced along the path.
10:10-10:20: How machine learning can support disability claims assessors: Two real-life examples
by Nora Leonardi
When an employee can no longer work due to illness, disability insurance kicks in to replace their wage until they are fit to return to work. As an insurer, how do you identify which work absences may develop into lengthy ones where claimants could benefit from early support, and how do you identify which long-running absences to review for continued validity and which ones not? We present two use cases where we leveraged machine learning to support disability claims assessors: For a UK insurer we predicted a submitted claim's likelihood to be of long duration using non-traditional behavioral factors to explain variability beyond diagnoses and demographics and to triage claims into risk categories. For a German insurer we segmented their in-force claims portfolio according to the likelihood of a claim terminating and matched each claim with a low to high-touch review action.
10:20-10:30: Health Progression Modelling
by Antoine Lagrange
In this presentation, we would like to present our work for forecasting mortality rates. Instead of using a time series approach (top-down), we use a bottom-up approach: first, we used machine learning algorithms to learn the health progression of individuals (changes in BMI, smoking status, diseases, death etc.). Second, we use these models to simulate the health progression of a seed population, to derive incidence, prevalence, mortality rates at different level of granularity (individuals, cohort, population).