While severe weather events like Hurricane Katrina or Hurricane Harvey dominate the news for weather-related disasters, the majority of insurance claims for weather-related events come from much more mundane, small-scale weather phenomena.
Insurance companies have noticed that existing weather-claims models do not adequately capture why this is so and are looking at way to remedy this so they can set appropriate premiums. More accurate models would also benefit local governments and businesses by identifying the most at-risk areas for flooding.
A team led by Christian Rohrbeck of UK’s Lancaster University studied a decade of data from three Norwegian cities to understand the relationship between the weather and insurance claims and why existing models could not adequately capture that relationship.
Rohrbeck and his team noticed that some of the verified claims of flooding were on days when the weather was mild and dry, which was surprising. So the team tried out a temporal clustering algorithm “which brings together insurance claims made over a small number of days, which the researchers believe are likely to be associated from just one weather event. In contrast, existing models separately count claims made on consecutive days.”
Rohrbeck told Insurance Business:
“By using a clustering approach, which has not been used in insurance before, we are getting a better insight into the association between weather conditions and claims, which helps to more accurately predict how many properties will flood,” he said. “The improved information from our model should help insurance companies, and authorities, to offer improved information to their customers, and residents, when weather conditions are likely to result in substantial localised flooding.”
Temporal clustering is a new subfield of data mining that collates information from sensors that record data and note time and location in real-time. It identifies time frames where the model gauges the likelihood that one event will trigger the occurrence of a similar event at a probability greater than random chance.
The use of a clustering approach sidesteps the daily reportage shortcomings of existing weather models that insurance companies rely on and represents a better approximation of the relationship between the weather and insurance claims.
