Predictive churn model drives over £1m in EBITDA

A machine-learning churn model built by QuantSpark helped a software-as-a-service provider identify its highest-risk customers with four times the accuracy of random selection, underpinning an estimated £1m-plus benefit to EBITDA. The business used the model's scores to prioritise retention outreach on the accounts most likely to leave, and turned the model's own insight into a strengthened customer training programme designed to make its customer base stickier.

The provider, a software-as-a-service business, was losing value from within its existing customer base. Rather than spreading retention effort evenly, or reacting only once a cancellation notice had already landed, it needed a way to identify high-risk customers early enough to intervene, and to focus that intervention where the financial stakes were highest. In subscription software, retaining an existing account is typically far cheaper than winning a new one, so a working early-warning signal for churn was worth building even before a single pound of benefit had been proven.

QuantSpark's response combined a tight, two-week analytical sprint with a delivery mechanism the business could actually use day to day. The team pulled together two and a half years of historical revenue data and engineered features from it: variance, gradient (the direction and pace of change) and averages, alongside other customer attributes that might signal risk. Rather than feeding every available variable into the model, the team identified the parameters carrying the most predictive weight and used those to tune a churn-scoring model. That discipline, keeping the feature set to what mattered rather than what was merely available, is what let the model be built and tuned inside two weeks.

A model is only as useful as the access people have to it. QuantSpark built a secure web portal specifically to disseminate the model's outputs across the business, so stakeholders across the business could reach churn scores directly rather than waiting on ad hoc data requests. That distribution step is often what separates a model that lives in a data science notebook from one that changes how a business actually spends its retention effort.

Benchmarked against random selection, the model correctly flagged likely churners four times more often, a clear, measured lift over the baseline the business had previously been working from. That accuracy translated into action: the business used the churn scores to prioritise retention outreach on its highest-risk accounts rather than treating its customer base uniformly. QuantSpark's own account of the engagement puts the resulting benefit at an estimated £1m-plus in EBITDA, explicitly flagged as a projection built on the model's outputs rather than an audited, realised outturn.

The engagement produced a second, less obvious dividend. Understanding which drivers the model leaned on most heavily gave the business insight into why customers were at risk in the first place, and that insight fed back into commercial strategy. It contributed to a strengthened customer training programme, on the logic that better-trained customers extract more value from the product and are less likely to leave. In effect, the churn model did not just triage the existing book, it pointed the business towards a structural fix upstream of the risk it was measuring.

The estimated EBITDA benefit follows a straightforward logic, even though the underlying figures behind it are not disclosed: use the model to identify the accounts at highest risk, prioritise retention effort on them, then model the revenue that would otherwise have churned through to its EBITDA effect using the business's own margin assumptions. That is a defensible way to size the opportunity a churn model creates, but it remains a modelled projection rather than a confirmed, after-the-fact result, and QuantSpark's own materials are careful to label it as such. What is verifiable from the public record is narrower and more solid: a model that, on benchmark, was four times more accurate than chance at flagging the customers worth calling first.