Prime Meridian Direct (PMD) is a private, short-term motor insurer. PMD was pricing policies using legacy deterministic (i.e. if-this-then-that) logic. They found their manual system fraught with risks: it’s impossible to scale such a system to great volumes, and the human operators may bring their subconscious biases into the equation. Their loss ratios were too high, and they felt their pricing was inaccurate and not market-related.

PMD wanted to differentiate by pricing risk accurately and objectively. They intended to attract good risk and avoid bad risk by pricing and underwriting in a more scientific, data-driven manner. PMD sought to generate more quotes per unit time, more accurately, and to make this functionality available as an API endpoint.

PMD partnered with us to implement custom ML algorithms for accurate pricing over their diverse portfolio using existing claims data. We enabled PMD to incorporate domain expertise into the pricing pipeline by embedding the ML model in a simulation framework. This framework maximises the bottom line and minimises the probability of non-conversion.

Working with the PMD ExCo, we developed domain-specific pricing rules that were integrated into the pipeline with relevant competitor information.

Motor Insurance Pricing Model

The model is a data- and AI-driven pricing tool. It overlays the competitive landscape (i.e. the market’s prices for the same risks) to ensure accuracy and competitiveness of PMD’s pricing. PMD’s data includes the following information:

• Rating factor details, including demographic info, driving experience, and driving statistics per client
• Vehicle details, including type, age, colour and mileage
• Claim details, including excess, claims history, and insurance product type

Data is used to train ML models to generate prices. The models integrated actuarial techniques to ensure pricing accuracy. The model was exposed as an API and made available in a user-friendly interface for salespeople. During a call with a lead, salespeople use the prices that the model outputs to inform the client of the premium in question.

The model sends emails to ExCo when results begin to drift when compared to a control / test sample.

Highlights and Results

• Six months to build the solution using four data sources
• Ingested 5–10 million rows of data
• Automated pipelines regularly trigger model updates
• PMD is now able to target a loss ratio per portfolio
• The model and underlying data are monitored autonomously
• Generates thousands of prices per second and is hosted on an AWS instance