Model migration, a 12-point loss ratio gap and a pricing decision
You are a senior actuary at a UK private medical insurer. The pricing team is replacing a 12-year-old claims model with a new platform. The new model projects a 90% loss ratio for the main PMI product, versus 78% from the old model. The commercial team wants to increase prices by 15% at renewal, but the CEO is worried about losing 8% of policyholders. During validation you find: 8% of claims lack diagnosis codes; 2% are duplicate records; provider coding changed in 2022; and 2020–21 experience is distorted by COVID-19. The new model uses a three-year weighted average, while the old model used a five-year average with explicit COVID adjustments. The board meeting is in 10 days. The chief medical officer says the new model is ‘black box’ and wants a simple trend rate. You must recommend which model to use and how to communicate the result.
1. What validation and reconciliation would you perform before accepting the new model?
Read the worked answer
I would start with data reconciliation: compare claim counts, amounts, dates, provider codes and diagnosis completeness between the old and new data sources. I would remove or flag duplicates, investigate the 8% missing diagnosis codes, and understand the 2022 coding change so that it is not mistaken for a real trend. I would run both models on the same cleaned data set, with the same cohort and the same treatment of COVID-19, to isolate model differences from data differences. I would back-test both models on earlier years, compare predicted and actual loss ratios, and examine the sensitivity of the 90% result to the three-year weighting and to the COVID adjustment.
I would also check the new model’s assumptions, inputs and calculations, and ask for independent review or peer challenge. If the new model cannot be explained or reconciled, I would not use it for pricing. The old model is not automatically right either: a five-year average may be stale after the 2022 coding change. The aim is a defensible central estimate with a range, not a choice between two unexplained numbers.
2. How would you communicate the model uncertainty and pricing implications to the board?
Read the worked answer
I would present the loss ratio as a range, for example 82% to 92%, rather than a single 90%, and explain the drivers: data gaps, coding change, COVID distortion, model choice and trend uncertainty. I would show the old model’s 78% as a lower estimate that may understate current experience if the coding change or medical inflation is real. I would separate the technical estimate from the commercial decision: the board can choose a 15% increase, a phased increase, or a lower increase with a plan to monitor, but it should understand the margin consequences.
I would recommend a moderate initial increase, such as 8–12%, with a commitment to review after six months, rather than a single 15% shock that may cause the feared 8% lapse. I would also quantify the lapse sensitivity: a 15% increase with 8% lapses may produce less revenue than a smaller increase with lower lapses. I would avoid ‘black box’ language by explaining the model in plain terms and showing the key assumptions. The board paper should include the data limitations, the range, the recommended price and the monitoring plan.
3. Recommend the governance and monitoring arrangements for the model after launch.
Read the worked answer
The model should have a named owner, documented assumptions, version control, change log and independent validation. The board should approve the pricing recommendation and the acceptable range, not the model code. I would set up monthly monitoring of claims frequency, severity, diagnosis mix, provider coding and lapses, and quarterly monitoring of the loss ratio against the pricing assumptions. I would define triggers for revalidation, such as a material coding change, a loss ratio outside the range, or a change in medical inflation. I would also require a post-implementation review after 12 months and a plan to improve diagnosis completeness.
The chief medical officer’s concern should be addressed through transparent explanation and clinical input, not by ignoring the model. A simple trend rate may be useful as a challenge or a fallback, but it is not a substitute for a validated model. The governance should make clear who is accountable for the data, the model and the pricing decision, and how disagreements are recorded.