Customer churn is a major problem in the insurance industry. The financial consequences of customer churn are huge and it is not something that insurance companies can ignore. It is important to understand the causes of customer churn, so that you can take steps to mitigate it.

Customer churn is a big problem for the insurance industry. This is because it is a lot more expensive to get new customers than to retain them. It is important for insurance companies to be able to identify which customers are likely to leave. They should then use this knowledge as an opportunity for customer retention by offering incentives or discounts in order to keep them from leaving.

We’ve already explained in the past how AI can be used to provide solutions in the insurance sector as well as fintech, and we explain more in our AI case studies bible (along with case studies from nearly 30 industries). We’re firm believers that data has the capacity to transform the way that the financial services industry conducts business.

The Tesseract Academy team worked with leading insurance provider based in London in order to predict and prevent churn through the use of machine learning and predictive analytics. In this post we explore this case study.

customer churn

The problem: Churn, data and insurance

Our client is one of the biggest insurance companies in the space of electronic devices. The company resells policies to partners, who then provide these policies to their customers.

Customers might choose to buy insurance for their device for different reasons, a common one being that their device is new and expensive (e.g. a new iPhone). They also decide to leave for various reasons, such as a claim not being covered, or not really using the insurance.

With regards to customers leaving, there are two types of churn: active and passive.

Active churn is when someone cancels their policy before its expiration date. Passive churn exists when someone simply decides to not renew their policy.

Unfortunately, it wasn’t possible to get customer demographic data, due to data privacy regulations. However, we had data on things like:

  • the device type and technical specs
  • when the customer joined
  • country

Plus some other proprietary datapoints.

Modelling customer churn using data science

There are various ways to model customer churn. Two popular ways to do it are to use survival modelling and classification.

A classification model can tell you the probability of a customer churning at a given point in time. The benefits of a classification model is that it is easy to interpret, since many classification models can provide you with a probability, which can be interpreted as a risk score. This means that the higher the probability of someone churning, the higher the risk.

A more effective and interesting class of models for this problem are called survival models. Survival models are popular in medicine, where they are used, as the name implies, to model survival of patients. A survival allows us to directly model relative risk (that is the risk of one customer vs another customer), and also model this risk over time. A example of this is shown in the figure below, where you can see how the risk is different for each contract type. A common characteristic amongst all three curves is that they tend to go down over time, indicating a lower probability of survival. What this means in simple terms is that as time goes by, the probability of quitting the service increases.

survival curves for customer churn prediction

Predicting customer churn: The results

The conclusion of our work consisted of two core deliverables:

  1. A determination of factors that contribute to churn.
  2. A predictive model that predicts which customers are at higher risk of churn and when they are about to churn.

First of all, we were able to determine which factors place the customers at higher risk of churning and rank these factors accordingly.

feature importance predicting customer churn

Then, we used this knowledge to build a machine learning pipeline. Using this pipeline we were able to predict about 89% of the customers that churn with a precision of up to 90%, which means that we can target about 4 out of 5 customers that will churn.

machine learning churn prediction results

Finally, using our model and domain knowledge we are able to rank customers according to when they are most likely to churn, so the business can prioritise contacting those at higher risk.

data science customer churn prediction risk score

Predicting customer churn using AI: You can start now

It’s clear that predicting customer churn is a hugely valuable proposition for any company. The Tesseract Team has proven that it is possible to predict customer churn. While this particular case study is from the insurance industry, there is nothing preventing us from applying the same principles to other industries such as retail.

If you want to know more what we can do for your business, then make sure to get in touch. We specialise in delivering solutions in AI and data science, but also educating decision makers on topics such as data strategy and data maturity.