We grouped the “equipped status” of each core ADAS feature into three distinct levels: Yes (equipped), No (not equipped), and Unknown. Due to the underlying differences in how ADAS technology affects the claim outcome for a given coverage, we built separate models for bodily injury, property damage, and collision coverage.
We then adjusted the underlying claim frequency to account for covariates such as policy year written, liability limit, vehicle symbol, vehicle age, and past claims. In addition, we used a proxy score to control for the main effect of credit history. Only vehicles where the equipped information was definitive (no Unknowns) were included in the final model. This adjustment allowed us to compare the true effect of core ADAS features on claim frequency and reduce “noise” in the modeled predictions.
Certain business consideration constraints were applied to improve the predictive performance of the model. Core features with both a warning only and mitigation version (eg. forward collision) are a great example. Holding all other factors constant, the raw predictions were smoothed to ensure that mitigation class was no worse than the warning only class. Our team adjusted the model predictions by using an iterative smoothing approach to account for such considerations.
In a decision tree, the estimated effect of a combination of core ADAS features on underlying claim frequency can be quantified by following a decision path. We can track the relative frequency difference at each decision node based on the core ADAS features that are equipped on a vehicle.