Sepehr Zolfeghari: Early Behavioural Risk Assessment in Telematics Insurance: The Role of Heatmap Resolution and Observational Data Availability

Abstract: Motor insurance pricing has traditionally relied primarily on policyholder and vehicle characteristics for risk assessment and premium setting. Telematics-based insurance extends this framework by enabling personalised behavioural risk assessment based on how policyholders actually drive. While telematics data are rich and complex, one practical representation is velocity–acceleration (v-a) heatmaps, which have previously been shown to improve claims prediction beyond models relying solely on these traditional factors. However, an important unresolved question remains: how does the granularity of the heatmap representation, together with observational data availability, affect the predictive utility of such behavioural representations?
  This thesis addresses this question primarily through a controlled simulation study, where the underlying claim-generating mechanism is known. Three heatmap resolutions and three levels of observational sample size are considered in order to assess how predictive performance is affected under varying levels of observational data availability. Principal component analysis (PCA) and bottleneck neural networks (BNN) are used as feature extraction methods applied to the heatmap representations, with the resulting features incorporated into a Poisson generalised additive claims frequency modelling framework. A complementary real-data application using telematics data from Paydrive AB, a Swedish telematics-based motor insurer, is also included to assess practical relevance.
  The simulation results suggest a trade-off between heatmap resolution and predictive performance under differing data availability conditions. Lower-resolution heatmaps generally allow behavioural patterns to be estimated more reliably under limited observational data, though in some cases with slightly weaker predictive performance, while finer heatmaps capture richer behavioural structure but appear more sensitive to observational scarcity. Overall, the findings suggest that an intermediate heatmap resolution may represent a practical compromise between predictive performance and observational data requirements.
  From a practical insurance perspective, these findings are relevant not only for selecting deployable telematics model architectures, but also for the broader question of how early behavioural risk can be assessed reliably. Earlier identification of risky driving behaviour could improve pricing accuracy, portfolio management, and potentially enable earlier intervention to reduce future claims.