Churn & demand prediction
Using your own history to estimate what happens next — who leaves, what sells.
Prediction sounds like the most advanced thing on this page. In practice it is the most ordinary: a model looks at what happened before and estimates what happens next.
For churn, it learns the shape of a customer drifting away — orders getting smaller, gaps getting longer, support contacts changing tone — and scores who is at risk while there is still time to pick up the phone. For demand, it learns seasonality and trend and estimates next month, so reordering stops being a guess.
Nothing exotic is required. On the data volumes a typical SME has, gradient boosting or even a well-built regression usually beats anything more fashionable. Anyone proposing deep learning for a thousand customers is selling you something.
You find out a customer has left when someone notices they have not ordered in a while — usually a quarter too late to do anything about it. Reorder quantities are set by whoever is on shift and their memory of last year.
Monday morning: twenty accounts ranked by risk, each with the reason. Your salesperson calls the top five. Reorder suggestions arrive before the shelf empties instead of after.
How we actually build it
-
01
We look at what you already have
Two years of transactions is plenty. Less than one, and we will say so rather than build something that guesses confidently.
-
02
We agree what we are predicting
"Churn" has to mean something specific — no order in 90 days, say. Vague targets produce useless models.
-
03
We build and test honestly
Trained on older data, tested on newer, so the score you see is the score you get. We report what it misses, not only what it catches.
-
04
We put it where someone will act
A ranked weekly list with reasons attached, in the tool your salesperson already opens. A model nobody acts on is an expensive report.
- You have at least a year or two of transaction history
- Losing a customer is expensive and you find out too late
- You are guessing at reorder quantities and paying for it in stock or stockouts
- Too little history. A model cannot learn a pattern from a handful of examples
- The data is unreliable — clean it first, or the model learns the mess
- Nobody will act on the score. A prediction nobody uses is an expensive report
- Transaction history, ideally 18–24 months
- Reasonably consistent customer records
- Someone who will act on the list
3–5 weeks including the data cleanup that always turns out to be necessary.
What people ask about this
How accurate is it?
For SME churn, catching 60–70% of leavers early is a realistic and genuinely useful result. Anyone promising 95% either has extraordinary data or is describing a test they overfitted.
My data is messy.
Almost everyone's is. Cleaning it is usually the first half of the project, and it makes your reporting better whether or not the model ships.
Often paired with
Would this work for you?
The first call is free, and we will say plainly if this is the wrong tool for your problem.
Book a free call