Finance & Insurance

Client Context

A mid-sized insurance company wanted to improve its claims forecasting model. Their existing approach relied on simple year-on-year comparisons and was producing unreliable reserve estimates. They had 10 years of claims data across motor, property, and liability lines of business, and needed a statistical model that could project future claims with defensible confidence intervals.

The Challenge

The data exhibited several complexities: long-tailed liability claims with significant development after the accident year, changes in reporting practices that created structural breaks in the time series, and strong seasonality in motor claims. The company also needed the results presented in a format suitable for their actuarial reserving team and the Financial Conduct Authority (FCA).

Our Approach

We applied the Mack chain-ladder method as a baseline and then extended it with a generalised linear model (GLM) framework that incorporated accident year, development year, and calendar year effects. We tested for calendar-year trends and identified a structural break at year 7, which we modelled with a dummy variable. Bootstrap simulation (10,000 resamples) provided predictive distributions for outstanding claims. We also implemented a Bornhuetter-Ferguson model as a sensitivity check. All analyses were performed in R with reproducible scripts.

Results

The GLM-based model reduced the mean absolute percentage error of reserve estimates by 34% compared to the chain-ladder baseline. Bootstrap prediction intervals gave the actuarial team defensible 95% confidence bands for the first time. The structural break analysis identified the claims reporting change, allowing the team to adjust historical data appropriately. The FCA reviewer accepted the reserving methodology without challenge.

Client Testimonial

"AnalyticsScholar gave us a model we could actually explain to the board and the regulator. The confidence intervals alone were worth the investment."

— Chief Actuary, Meridian Insurance

[Explore More Case Studies → /case-studies/]

[Discuss Your Finance Project → /contact/]