🎯 Goal: Identify which customers are statistically likely to stop renting — before they actually do — so the business can intervene while there's still a relationship to save.
Imagine you manage 1,600 business customers who rent machinery. Some of them rent every month. Some rent once and disappear. And some — the dangerous ones — are quietly drifting away without sending any signal.
By the time a customer stops calling, it's too late. The contract is gone, the relationship has cooled, and winning them back costs far more than keeping them would have.
This model was built to answer one question: which customers are showing early warning signs of leaving, right now?
The model is a machine learning classifier — a type of algorithm that learns patterns from past data and uses them to predict future behaviour.
It was trained on 4 years of real transaction data (2021–2024) from 1,621 B2B customers in the machinery rental sector.
Rather than looking at customers in isolation, the model combines six signals that together paint a behavioural picture of each customer:
| Signal | What it measures | Why it matters |
|---|---|---|
| Recency | Days since their last contract ended | A customer who hasn't rented in 6 months is more at risk than one who rented last week |
| Frequency | How many times they've rented in total | High-frequency customers have a stronger relationship with the business |
| Monetary Value | Total revenue they've generated | Higher spend = more invested relationship, but also more to lose |
| CV of Interpurchase Time | How irregular their rental gaps are | Customers who used to rent regularly but have become erratic are a warning sign |
| CV of Contract Duration | How variable their contract lengths are | Shifting from long-term to short-term contracts often precedes departure |
| Erlang-k Parameters | Mathematical shape of their purchase rhythm | Derived from the CVs above — captures the statistical structure of customer behaviour |
💡 The last three signals — the CV metrics — are what make this model more sophisticated than a basic RFM analysis. They capture how a customer is changing over time, not just where they stand today.
A customer is considered churned if their last contract ended more than 90 days before the end of the observation window (December 31, 2024). This threshold reflects the typical re-rental cycle in this industry.
Two models were run — one on the full customer base, one focused on the top revenue customers.