A Portfolio-Anchored Frequency–Severity Risk Index for Trip and Driver Assessment Using Telematics Signals

Battle Count: 50

This paper is entirely focused on telematics-based driver risk assessment for insurance applications. While it uses statistical methods (mixture models, Bayesian updating, wavelet transforms) that have analogs in quantitative finance, the domain, data, and objectives are unrelated to trading, portfolio management, or market prediction. The Poisson-Gamma conjugacy and sequential updating concepts are broadly applicable but are not leveraged here for any financial market application.

The framework involves multiple interconnected components: (1) MODWT decomposition with Daubechies D4 filters and across-level aggregation, (2) hierarchical sampling with within-trip thinning, (3) Gaussian-Uniform mixture fitting via the novel MU-MEMR algorithm with constrained EM (scale, tail-separation, coverage, and monotonicity constraints), (4) isotonic regression for probability constraints, (5) Poisson-Gamma conjugate modeling with winsorized empirical Bayes initialization, and (6) sequential Bayesian updating for driver-level profiles. The grid search over (M-, G, M+) and gamma tuning adds further complexity. Proper implementation requires careful handling of numerical stability, constraint enforcement, and computational efficiency.

The UAH-DriveSet dataset is publicly available. The methodology is described in extensive mathematical detail including the MODWT filters, MU-MEMR algorithm pseudocode, Poisson-Gamma conjugacy, and severity weight construction. Model selection criteria (BIC) and tuning parameters (d=0.5, alpha=0.05, delta=1.96, q=12, p=10, gamma=1.7) are explicitly stated. However, no code repository is provided, and some implementation details (e.g., exact thinning procedure, grid search specifics) would require careful reimplementation.