A large insurer sought to shift their model for selling long term care insurance by pivoting from third party brokers to direct sales. In order to do profitably, they needed to decrease their lead-to-policy cost. This, paired with the volume of data available, presented the opportunity to apply machine learning in a proof of concept which would help the company determine those leads most likely to convert. The UDig team evaluated thousands of model variations before identifying a successful artificial neural network variation as a model that identified patterns indicative of purchasing behavior.
With the goal of increasing product profitability around the client’s specific insurance product, and in light of a changing marketplace, our client sought to eliminate the use of third-party brokers. Instead they sought to pivot to a direct sales approach focusing on organic lead generation.
Our client combined their existing data sources with a third parties rich market research data. This created a very broad data set to identify trends in customer conversion.
The pivot in approach lead to an interesting challenge: while they could describe historic behavior for this product well with their data, it was not directly applicable to their new approach. Indeed, the data presented a very biased view of customer acquisition due to the nature of brokerage services.
This model will lead to a lead prioritization system driving customer acquisition efforts.
As such, our initial effort focused on profiling and preparing the data for machine learning. Once in a conducive state for machine learning, we proceed with model selection. Here, we investigated the application of various models for the task of lead-to-policy classification. Analyzing the behavior of everything from linear regression classifiers to artificial neural networks provided deep insight into problem space.
Settling on a focus of decision trees and neural networks as target models we were able to optimize and grade model performance to seek out the optimal model. This wound up being an artificial neural network, with a close second of a decision tree ensemble.
Meanwhile we worked on reducing the feature set necessary. Through a variety of approaches, we were able to reduce the overall feature set by about 80%, from around 1600 to just over 300. These features represent the core of what it is to be a customer converting from a lead to a policy.
The result was a model capable of scoring lead conversion probability. Along with additional efforts undertaken by the client, this model will lead to a lead prioritization system driving customer acquisition efforts. In addition, through the identification of bias between the old and new approaches the model has been made to better generalize to the true target market.