How G7 Tackles Truck Underwriting Risk: Modeling Challenges & Solutions

Based on G7's IoT insurance algorithm team's early practice in truck underwriting risk control, this article examines modeling characteristics and difficulties from an algorithmic perspective and shares feasible alternative solutions.

Data foundation and overall goal

Complete truck operational risk description should be based on multiple dimensions (driver, vehicle, cargo, road), but due to limited data acquisition, currently only vehicle static information and movement data (GPS trajectories, highway camera counts) are used. The overall goal is to predict claim amounts for the next policy period using three to six months of onboard device data combined with other available data.

Modeling scenario characteristics, difficulties, and stage thoughts

Identifying scenario characteristics and difficulties guides solution design. Two main requirements: 1) Recognize accurate scenario traits and challenges based on business and data; 2) Build suitable solutions using existing resources.

Data sample and modeling target dimension

Focuses on the relationship between observed data and the modeling target, clarifying data support for different targets and their data needs.

Feature: Long chain from data to final target

Current risk modeling uses vehicle driving behavior as observation and claim amount/rate as the ultimate target. The claim process involves many intermediate steps (driving risk → accident risk → accident → insurance claim), introducing target diversity and strong randomness.

Difficulty 1: Diversity of modeling direction

Different modeling directions (driving risk, accident risk, claim risk) require different data, e.g., driving risk needs driver, vehicle, cargo, environment data; claim risk also needs insurer claim policies and injury standards; long‑chain data are hard to standardize across insurers.

Difficulty 2: Strong randomness of claim data

Even with real claim data, the amount shows strong randomness relative to underlying driving risk, making direct claim‑based risk assessment unreliable.

Stage thought

Answers: 1) Model vehicle driving risk as the primary target, with accident and claim risk as auxiliary; 2) Align claim data across insurers and use combined loss rate and group claim rate to mitigate individual claim volatility.

Feature construction dimension

Goal is to extract patterns from observation data and quantify them. Long‑term prediction requires short‑term behavior (3–6 months) to forecast risk over a year.

Feature: Long‑term applicability

Three approaches to use short‑term behavior for long‑term prediction are illustrated

Difficulty: Temporal variability of vehicle behavior

Driving patterns change over time due to season, cargo, etc., posing challenges for stable modeling.

Stage thought

Three strategies: 1) Emphasize stable long‑term behavior features; 2) Model temporal variability as an auxiliary target; 3) Aggregate similar vehicles into groups to capture collective patterns.

Overall model structure dimension

Scenario diversity and long‑tail distribution increase modeling complexity, making a single model insufficient. Regionalized models incorporating local risk factors improve discrimination between high‑ and low‑risk vehicle groups, as shown in the comparison

.

Future directions include enriching data sources (e.g., dual‑prevention devices), integrating spatio‑temporal behavior mining

, and enhancing model interpretability for risk explanation.

Conclusion

The article provides a broad overview of underwriting risk modeling characteristics and G7's current approach, highlighting challenges and potential solutions, while acknowledging that many issues remain at the conceptual stage.