Speaking at a recent Forum of Insurance Lawyers’ (FOIL) webinar, Allianz’s Chris Wyard discusses the pros and cons of decision trees 

With the volume of data available to the insurance industry, there is an opportunity to personalise products, enhance risk management and automate business processes.

This was the view of Chris Wyard, chief data officer at Allianz, who was speaking at a recent Forum of Insurance Lawyers’ (FOIL) webinar, titled ’The Ethics in AI (Artificial Intelligence) in Insurance Part Two’.

Typically, a decision tree attempts to take years’ worth of data and research, chunk it down and limit the biases.

But Wyard stressed: “Not every problem can be solved through machine learning. We need to be thinking about what kind of data we have available, the suitability and indeed some of the [inherent bias] that could exist within it, which needs to be resolved before we reorganise and reimagine our businesses.”

For example, the world created in the data might not actually exist between the customer and insurer, as context is always changing.

“The model is only going to be as good as the training data it has been created by,” he added.

Decision trees

Machine learning is a broad topic, around highlighting patterns and boundaries in data – one being decision trees.

The trend of data being used across industries to make decisions has often been referred to as a “fourth industrial revolution”.

Wyard described decision trees as a model used to understand the outcomes in a data set, using “if, then” statements.

He gave the example of his mother’s “GP at home” book, which included a ‘flow to’ list that encouraged the reader to answer several questions, such as ’why does my head hurt?’ And this would, in turn, guide the reader to an outcome.

He cited NHS Online as another example of this. “You will invariably know that the model takes you to a place where it says, ‘you need to call 111’.”

Wyard warned that decision trees need to operate at a level that is explainable and easily represented in order to avoid taking millions of transactions down to the individual point.

He questioned – at what point is it appropriate to aggregate data out?

“Because at that point you have over-engineered your model,” Wyard said. Although he added that one of the benefits of using decision trees, is to “easily prescribe and represent what’s happening within data”.