By utilising machine learning and numerical text processing techniques, Swiss Re has been able to generate a “predictive view” of motor frequency developments in several markets.
In a recent conversation with Nikita Kuksin, Hhead of modelling within Casualty R&D , Miriam Hook, vice president Global clients and Surbhi Gupta, assistant vice president, casualty R&D at Swiss Re, it was explained to us how these alternative approaches were able to provide added granularity to existing data.
“We intended to develop an alternative to traditional actuarial calculation methods that would give us an “external perspective” on claims frequency within our motor portfolio and allow us to predict motor frequency developments in several motor markets,” said Kuksin, who leads the modelling team within the casualty research and development department at the Swiss Re Institute.
Gupta, who prior to her current role served at Swiss Re for three years’ as a data scientist, explained how these methods were brought into fruition by first checking the status quo of frequency developments against external data, before then explaining motor frequency using external data to generate factors that could be projected into the future.
“These are complex objectives, requiring solid data sets and robust analytics,” Gupta explained.
“We achieved our aims by using established machine learning algorithms, which were innovatively put to use by our experienced team using the right data.”
The team started by analysing claims frequency in Swiss Re’s US motor portfolio.
With the National Highway Traffic Safety Administration (NHTSA) database they could rely on a dataset containing details of every accident in the US, creating a complete picture across more than 30 years.
“Descriptive information within the NHTSA database is available in text form, or ‘unstructured data’ to use machine learning jargon”, Hook explained.
They then used a text mining process to “read” this data so that it could be utilised in a numerical context.
It was found that the NHTSA data indicates that approximately 85% of losses in relation to motor liability insurance stem from three general scenario groups, one of which is “accidents during turning.”
“We also identified strong regional differences in frequency within the USA and in turn could adjust our pricing models to capture this for more granular analysis,” Gupta added.
“This evaluation gave us an external benchmark to understand whether our own pricing assumptions were correct.
“We incorporated this information into our motor strategy, allowing us in particular to use the above findings to enhance the frequency assumptions that we use in our models.”
In order to analyse future claims frequency, the team looked to develop a predictive view of future claims frequency for third party motor insurance in various EMEA markets – Germany, UK and France – to improve Swiss Re’s portfolio steering and to validate existing projections.
“We again used machine learning algorithms and external data to complement our traditional modelling approaches for this analysis: key to our approach was to identify (external) data/factors that impact the value one is trying to establish, in this case the claims frequency of motor business,” said Hook, who developed several client use cases resulting from this work together with her two colleagues.
To find these factors, the team looked at a number of different datasets from economic indicators to infrastructure and weather data, etc. to establish whether these parameters on their own or in combination significantly explain motor frequency developments as it is perceived in any given motor market.
“For instance, in Germany we identified four factors that have a significant impact on claims frequency. One of those factors is the regional investment in infrastructure, most likely due to the fact that improved road conditions have a positive effect on the number of accidents and therefore on claims frequency.”
Ultimately, combining the four factors in a generalised linear model allowed Swiss Re to establish a reasonably accurate picture of historical claims frequency.
“We backtested the model over a longer historical timeframe and achieved good results. Based on the combination of best factors, we modelled the frequency development over time in a series of one-step-ahead forecasts.
“Although these initial results are not yet perfect, they did allow us to correctly predict all directional claims frequency developments one year in advance.”
Nikita Kuksin added that, if Swiss Re can now forecast these four key factors for the next few years (based primarily on external data), it will be able to determine the expected future claims frequency in the motor third party liability sector.
“The two mentioned areas of application illustrate how alternative modelling methods, such as machine learning, can be used in the motor sector,” Kuksin concluded.
“Similar questions may arise in other insurance branches. New technical possibilities for analysing data allow, with relatively little effort, comparable models to be created or assumptions to be tested that can be used to complement traditional actuarial models.
Swiss Re believes this strategy offers great potential to improve the models used in the underwriting process for risks or portfolios.