We propose, in this paper, a comprehensive study on risk assessment related to car insurance, based on claim and telematics data, collected from a dataset of voluntary drivers. Our work addresses experimental settings not covered before in the state of the art, such as the collection of telematic data within a period significantly after the claim reporting one, and coping with much more unbalanced data owing to the rarity of the claim class. To address these issues, we propose weighted XGBoost models that model the telematic-based features, represented as tabular feature heatmaps and mitigate the class unbalanced problem. Our heatmaps encore not only speed-acceleration distributions, but also speed-jerk (acceleration derivative) distributions, not considered in previous studies. To ensure model interpretability, we assess the importance of each feature in order to reveal the telematic features that are most discriminant for claim detection. Owing to rarity of the claim data, we propose also one-class machine learning models, namely 1-class SVM and Isolation Forests, trained only on the most dominant class, the samples from the underrepresented one being assessed only in the test phase and considered as outliers or anomalies. In addition, we propose a novel unsupervised machine learning strategy consisting of a two-stage clustering scheme that not only allows to infer a driver signature and estimate, through sound informationtheoretic measures, how stable a driver behavior is, but also uncovers the correlation between driver groups (clusters) and claim distributions. The first clustering stage uncovers behavioral clusters, i.e. monthly-based heatmap prototypes, irrespective of drivers, while the second uncovers yearly based driver clusters with similar behavior w.r.t the first-level clusters. We obtain promising results given the adverse conditions considered and we provide guidelines in the conclusion for developing more effective driving risk assessment based on telematic data.