Machine learning
Can past behaviour be used to predict and influence future outcomes?

28th February 2018

Historically, the biggest question in artificial intelligence has been 'can a machine think?'. Nowadays, however, the more pertinent question may be 'can a machine learn?'. Often viewed as a subcategory of AI, machine learning is the field that is currently yielding the most promising results when it comes to creating tools with applications that will change the world we live in.

In insurance, machine learning is already being used to better target online advertising campaigns by the likes of Zurich, and Cambridge-based InsurTech start-up Cytoria are using machine learning algorithms to help insurers better evaluate risk. It's set to transform every aspect of the industry from customer service to claims processing. At CDL, our InsurTech Incubator has just shown exciting developments using machine learning software to predict people's propensity to purchase policies at different prices.

Machine learning isn't a new discipline in computer science. The basic concepts have been applied since before computers had transistors. Computers are programmed to make statistical predictions about (or 'classify') real-world things or phenomena based on 'training' data about pre-selected 'features'. These classifications are honed using statistical tools, such as networks of decision trees or models that can incorporate data about hundreds of different features.

Organic inspiration
Some machine learning techniques have been inspired by how the human mind works. 'Natural neural networks' are loosely based on the networks of neurones in the human brain that take chemical signals and process them to deliver outputs to other neurons. Artificial neural networks process data, rather than chemical signals, according to a set of rules, and the outputs are probabilities. The addition of a feedback loop in this system enables 'learning'.

It's these artificial neural networks that have enabled the creation of computers that could beat grand masters at chess, effectively determine if emails are spam and predict what you want to say when texting.

Despite being a relatively old discipline in computer science, it's only recently that we've been able to fully realise the full potential of artificial neural networks as a result of increased computing power and the advancement of graphics processing units (GPUs). This has allowed for the creation of artificial neural networks with many processing layers and enabled what’s known as 'deep learning'.

With deep learning algorithms, it's possible to create programmes that are capable of learning to understand natural language, recognise human faces and make predictions about human behavior, including anticipating what you may like to watch on Netflix. In the future, it will be machine learning that enables cars to drive safely on roads by themselves and computers to diagnose medical conditions more effectively than doctors.

New tools for developers
The pace at which we see machine learning play a meaningful role in everyday life is accelerating. We can expect the pace to increase as developers get their hands on tools that allow them to relatively quickly and easily build, train and deploy machine learning models.

CDL's InsurTech Incubator has recently created and trained a machine learning algorithm that is capable of accurately predicting people's propensity to purchase based on historical purchase data. The goal was to create a plug-in for CDL's Real Time Pricing (RTP) engine that could apply a discount if the algorithm predicted it would result in an otherwise unrealised sale.

Sixteen features were incorporated into the model, including the age of the person requesting the quote, their no-claims bonus, marital status and the price at which they purchased their policy, and it was trained using data from 168,000 instances of customers requesting a car insurance quote through Strata.

When tested, the model proved impressively accurate. In 68% of the 24,000 cases used to test the data, the algorithm correctly guessed if a person would purchase a policy or not. This is hugely encouraging given the relatively small sample of data used to train the model.

Next steps for insurance
It hardly needs saying that, once finely-tuned, this will be an invaluable tool for brokers and insurers, allowing them to maximise revenues by realising sales opportunities that otherwise would have been missed.

The next steps for CDL will be to further train the model using more data and experiment with other data fields and their weightings to see if this will improve the accuracy of the predictions, and the possibilities for using similar models for predicting other behaviours are endless.

In other areas, machine learning is a promising avenue to explore when it comes to developing enhanced natural language processing capabilities. As machines become better able to recognise intention and sentiment in people's voices, this will make our vision of the virtual call centre an ever-closer reality.

Although computer scientists today would say that machine learning is a set of techniques that sits inside the more ambitious goal of AI, we're betting on machine learning being one of the fields that has the biggest impact on making insurance more intelligent in the future.