“Predicting claims is an important measure and useful, but it is only a part of the picture.”
Insurers have used statistics and actuarial science for a long time. Digital sales bring a new range of data that most insurers do not yet use fully, and it needs modern machine-learning methods. This case study shows the difference, and it is significant.
When people think of analytics in online insurance, they usually think of pricing on the probability of a claim. Claims matter, but they are only part of the picture. To sell insurance online profitably, you also need to predict which leads to buy, which channels work, who will renew, who will cancel, who is gaming the quote, and who will default on instalments. The customer journey should adapt to each customer as well.
The client
A major insurance company, focused on sub-prime and near-prime personal motor insurance in the UK and continental Europe. It sold policies underwritten by partner insurers. Most customers arrived through price-comparison sites, shopping for the best quote. So every answer had to be right, and immediate.
For each customer arriving at the website or app, the client needed answers before quoting or buying the lead:
Acquisition and conversion
- Do we want to compete for this customer, and buy this lead?
- Is the customer likely to buy?
- If the journey is abandoned, when and how should we follow up?
- How should the journey adapt?
Risk and fraud
- What is the probability of gaming or fraud?
- Of default on instalments, and when?
- If the customer defaults, how likely are we to collect?
- What are the expected claim frequency and severity?
Pricing and underwriting
- Which products and add-ons should we offer?
- Should we take this business, and at what price?
- Do we need more data first?
Lifecycle and value
- How likely is the customer to cancel, or to renew, and how many times?
- What is their likely lifetime value?
- What is the next best action?
The client used proven claims and policy systems and its own loss history. To compete online, it needed to add new data and answer these questions at the moment of quoting.
The problem
As a first phase, the client chose to improve cancellations, gaming and fraud, renewals, claims, default, add-on sales, conversion, pricing and profitability. It wanted to use real-time analytics and new digital data, and measure the effect on each of these.
How the work was done
Zoral started with a short proof of value, to measure what the Zoral tools could add:
| Question | Answer |
|---|---|
| Does the client’s data contain more predictive value than it currently uses? | Yes |
| Can this be improved further by adding third-party data? | Yes |
| What is the likely scale of improvement? | Significant |
| What is the best path to a high-return pilot? | Build the data foundation, apply Zoral Decision Engine and tune the model library |
Zoral then:
- built the data foundation in Zoral Data Fabric;
- validated and tuned the Zoral model library on production data;
- connected Zoral Decision Engine to the client’s real-time pricing and rating workflow;
- trained the client’s team to use the engine and the models.
Internal and external data were brought together, including bureau and aggregator data. The tuned models covered default, gaming and fraud, renewal and cancellation, conversion, add-on sales and claims. They were benchmarked against the client’s own production models, where these existed. The benchmark was promising. It also showed data-quality problems, which Zoral Data Fabric fixed before go-live. The models then fed the client’s pricing engine at quote time.
The pricing system in production
Channels
Claims & policies
Web logic
RTP integration
Decisioning
Data
Insurance broker storage
Results by model
Default prediction: 13.9% improvement
More than half of the customers paid for their policies in instalments, while the client paid the full premium to the insurer at the start. So the client carried part of the default risk, and its profit depended on measuring that risk well. The tuned models predicted bad debt considerably better, using non-linear methods and a wider range of internal, bureau and aggregator data.
Bad debt probability prediction before and after ZDE / Model Library
Bad debt scores: ROC curves before and after Model Library
Cancellation prediction: 11.8% improvement
A cancelled policy cuts the customer’s lifetime value and can turn a profit into a loss. So the client needed to estimate the chance of cancellation at quote time and reflect it in the price. The cancellation model was tuned on the client’s data plus additional sources, including unstructured data such as call-centre notes.
Cancellation probability prediction before and after ZDE / Model Library
Cancellation scores: ROC curves before and after Model Library
Renewal prediction: 20.4% improvement
New customers are expensive to win, and first-year premiums were often discounted. So renewal drove each customer’s lifetime value and the profit of the business. The renewal model was applied through Zoral Decision Engine. Unstructured data from live chat and call-centre notes was added as a further input.
Prediction of insurance renewal with ZDE / Model Library
Renewal scores: ROC curves
Prediction of insurance renewal based on unstructured data with ZDE / Model Library
Claims prediction
Claims prediction is central to the price of a policy. Better prediction at quote time also helps the client negotiate prices and terms with its partner insurers. The new model separated risks more clearly, especially in the middle range, where the existing models had become unreliable.
Claim rate before and after ZDE / Model Library
Claim scores: ROC curves before and after Model Library
Gaming and fraud prediction
Some customers change their details to bring the premium down. They also cancel more often, and perform worse on credit and renewal. Finding these customers let the client improve its pricing further.
The fraud team had 22.5% less gaming and fraud to chase. So the same fraud and operations team could handle a growing portfolio. That lowered the cost of running the portfolio and raised its profit.
Gaming prediction with ZDE / Model Library
Fraud / gaming Model Library score: ROC curve
Add-on sales: 17.3% increase in add-on profitability
A lower price to win a customer can be offset by add-on sales. The add-on model estimated, at quote time, how likely each customer was to buy specific add-ons.
Add-on cross-sell prediction before and after ZDE / Model Library
Add-on scores: ROC curves before and after Model Library
Conversion: 47.1% increase
The conversion model raised conversion among low-risk customers by offering them progressively better terms. It raised profit among high-risk customers by charging them a higher premium.
Conversion probability prediction before and after ZDE / Model Library
Conversion scores: ROC curves before and after Model Library
+44 20 3963 2002