Definition
Measures whether the telematics driving score's stated risk levels correspond to actual claims frequency and severity at each score band, revealing whether policyholders with "safe" scores are genuinely lower-risk or just lower-scored.
The Measurement Gap
Telematics programs promise to price individual driving risk more accurately than traditional rating factors. Carriers grant premium discounts based on telematics scores, assuming that high scores indicate genuinely lower risk. But the score-to-risk relationship may not hold uniformly across all score bands. A "90-score" driver may file claims at the rate the model predicts for score 80, meaning the discount is actuarially unjustified for that band. Legacy metrics cannot detect this band-level miscalibration. Loss ratio measures aggregate severity across the entire telematics book without isolating performance by score band. Claims frequency tracks how often losses occur in aggregate but does not compare actual frequency against predicted frequency at each score level. Retention rate measures whether telematics-scored policyholders stay, not whether their risk pricing is accurate. Discount penetration measures adoption of telematics programs, not whether the discounts granted are earning their keep. None of these metrics evaluate whether the score-to-discount mapping is actuarially calibrated at the granularity where pricing decisions are actually made.
Decision Supported:
Is the telematics discount accurately reflecting reduced risk at each score level, or are some score bands receiving discounts that are not actuarially justified, and should the score-to-discount mapping be recalibrated?
Formula
Plain language: Group policyholders by telematics score band (e.g., deciles). For each band, compare the actual claims frequency to the frequency predicted by the score. Perfect calibration means a "90-score" policyholder files claims at the rate the model predicts for score 90.
Notation:
Per-band: TSCI(b) = Frequency_actual(b) / Frequency_predicted(b)
Overall: TSCI = Sum of (n_b x |1 - TSCI(b)|) / Sum of n_b
Variables:
| Variable | Definition | Unit |
|---|---|---|
| b | Telematics score band (e.g., decile or quintile grouping) | Classification |
| Frequency_actual(b) | Actual claims frequency for policyholders in score band b | Claims per policy-year |
| Frequency_predicted(b) | Model-predicted claims frequency for score band b | Claims per policy-year |
| n_b | Number of policy-years of exposure in band b | Policy-years |
Per-band TSCI(b) of 1.0 indicates perfect calibration. Values above 1.0 indicate the band is underpriced (more claims than predicted). Values below 1.0 indicate the band is overpriced. Overall TSCI of 0 indicates perfect calibration; higher values indicate worse calibration.
Proposed formula, pending validation pilots.