Churn-prediction model
We build a churn-prediction model: we estimate the probability that a customer will leave, so you see the risk zone in advance and work with it before the customer churns. So you retain proactively, not after the fact. Honestly upfront: the model gives a probability, not a verdict; it does not prevent churn by itself and makes mistakes — retention work is done by you.
Churn-prediction model — overview
A churn-prediction model is an ML project (with refresh): on historical data (behavior, activity, purchases, churn signals) we train a model that assigns customers a leave-risk score, highlights key risk factors and forms a list of at-risk customers for proactive work. Honestly about the nature, this is key: the model outputs a PROBABILITY, not a fact — some 'risky' ones will stay, some 'safe' ones will leave; there are false positives and misses, and accuracy is limited by data and the nature of the task; it is a prioritization tool, not a prediction of fate. Honestly about data: history with enough churn cases and correct 'left/stayed' labeling is needed — without it or on small numbers the model is unreliable; dirty data = wrong forecasts (garbage in, garbage out). Honestly about aging: behavior changes, the model degrades — it needs retraining and fact reconciliation. Honestly about the essence, important: the model FORECASTS churn but does NOT PREVENT it — retention (calls, offers, fixing product problems) is separate work done by you or your team; the score itself retains no one, and it matters not only to flag risk but to react. Honestly about factors: the model shows features associated with churn (correlations), not proven causes. Requirements: access to data and churn labeling. An important boundary: this is churn forecasting, not retention work (separate), not cohort analysis (descriptive — separate) and not a guarantee of retaining customers. If there are few customers/churn cases, the model is premature. Picture this: instead of 'we learn of a departure when the customer is already gone' you see the risk zone in advance and have time to act (if you act). The base price starts from 30,000 ₽ per project; it depends on data and model complexity.
Problems we solve
- You learn of a customer's departure when they have already left.
- There is no prioritization of whom to work on retention first.
- You do not understand which signals are associated with churn.
- Retention is reactive, not proactive.
What's included in the Churn-prediction model service
- Preparing history and churn labeling
- Training a leave-risk scoring model
- Accuracy assessment (errors, false positives/misses)
- Key risk factors (correlations)
- A list of at-risk customers for proactive work
- A report with limitations and recommendations
- Reviewing results with you
- Model-refresh recommendations
What you get
- The churn risk zone is visible in advance (as a probability)
- Prioritization of retention work
- Churn-associated factors are clear (correlations)
- A base for proactive retention (retention itself — separately)
How the work goes: steps
- We clarify the churn definition, data, horizon; collect access
- We prepare data, train and check the model, assess errors
- We compile a report and risk list, review with you
Why PDV Expert
- Fixed price and timeline — no surprises on the invoice.
- Report and recommendations in plain language — clear without a technical background.
- In touch at every step and answering questions about the result.
FAQ
Will the model prevent churn?
No. It forecasts the probability of leaving and helps prioritize, but it does not retain by itself. Retention (calls, offers, fixing problems) is separate work done by you or your team; it matters not only to see risk but to react.
Can the forecast be trusted as a fact?
No. The model gives a probability, not a verdict: some 'risky' ones will stay, some 'safe' ones will leave; there are false positives and misses. It is a prioritization tool with limited accuracy, not a prediction of fate.
I have little churn data — can you build it?
History with enough churn cases and correct labeling is needed. On small numbers or with dirty data the model is unreliable (garbage in, garbage out). If data is scarce, we honestly say it is too early.
About the provider
The «Churn-prediction model» service is provided by PDV Expert — a team specialising in «Conversion & analytics». We work under contract and deliver a written report with recommendations.