top of page
  • Writer's pictureBeyond Team

Implementing Random Forest for Effective Attrition Modelling

Everyone knows that it costs more to recruit new customers than it does to keep existing ones. One of the most recent studies reveals that, on average, acquisition costs seven times the amount of retention.

In fact according to Gartner, if you reduce customer churn by 5%, an organisation can boost profit by at least 25%.

So, creating an attrition strategy makes excellent business sense, and is exactly the reason why we devised a new approach for one of our financial clients. A leading provider of commercial fuel cards, both in the UK and across other international markets, their customers range from sole traders such as tradesmen, small to medium enterprises (SMEs) with varying requirements and fuel usage, through to large logistics companies with high volume fleets.

The fuel market is highly competitive because it’s very much driven by price. In the current climate where the cost of doing business is at its highest level for over 15 years, for smaller businesses getting the best price is critical and just a fraction of a pence difference was proving to trigger customers with multiple fuel cards to switch between providers. Attrition amongst SMEs was becoming a problem and the business was keen to find ways to better identify behaviours which could signal that a customer was truly churning.

Herein lies the problem - different customers have very different behaviours. Just because someone hasn’t used their card for three weeks, doesn’t necessarily mean that they’ve swapped to another provider, perhaps their inactivity was actually down to the fact that they had taken some annual leave. If this was the case, triggering an intervention from a customer support agent could do more harm than good and result in eroding lifetime value.

Through the misdiagnosis of attrition behaviours it means that non-churning customers could be offered better rates or more favourable credit terms, which leads not just to direct depletion of the bottom line, but also indirect costs associated with the inefficiency of making the call. Essentially addressing false positives, whilst missing false negatives.

The attrition modelling solution

To reduce the attrition problem and improve confidence that customer support interventions were being offered to the right customers, we devised an approach using Random Forest, a powerful machine learning algorithm which combines multiple decision trees to make accurate predictions. The model is able to handle complex datasets with many features, including both numerical and categorical variables, which is necessary for attrition datasets. By using a large number of decision trees, each trained on a random subset of the data, the model can avoid overfitting and provide more robust predictions. This means that the interactions between different variables are identified. Additionally, Random Forest models can handle missing data and outliers, making them resilient to “noisy” data. This is critical, because to be effective the model had to be able to diagnose patterns of behaviours for each unique customer, rather than look at the customer base as a whole.

To put it simply, if customers A and B both typically pump every three days and then stop, with no fuel purchase being logged for a month, should they both receive a call from customer support offering them an incentive to continue pumping? If the model was just looking at pattens of inactivity, then yes they both should be called, but if the model takes into account individual behaviour and inactivity, then the answer is no. Only A should receive the call because their individual behaviour shows that this is a true anomaly; whilst B is a shift worker and the stoppage is merely denoting a period of inactivity, rather than churn.

The above example displays one of the major benefits of using random forest for an attrition model. The output is easily interpreted as it provides feature importance scores that can help identify which variables are most influential in making predictions. This means that the customers like A and B who were identified as having an attrition risk, are now ranked and prioritised and the customer support team make calls that target the right customer.

The result

Attrition has been reduced by 18% and the number of inappropriate lures-for-loyalty being offered to non-churning customers have also significantly dropped, demonstrating the double-whammy benefit of the approach.

Ultimately, financial businesses need to keep customers on the books for longer as longevity is directly correlated to loyalty, and loyal customers are proven to spend more, take less resource to service and often make referrals. Whilst customer acquisition is critical to any business, simultaneously stopping needless churn at the bottom of the funnel will, as the above demonstrates, pay dividends.



bottom of page