We start with the decision, not the technique. The method is designed so that a commercial reader can follow the reasoning, and a data team can audit it.
Analytical Flow
-
Commercial framing
Establish which decision the work has to serve: a budget reallocation, a scaling call, a retention investment. Everything downstream is judged against whether it moves that decision. -
Reconciliation
Bring acquisition, billing and behavioural sources onto one definition of a customer, a conversion and an active month, and report where the sources disagree instead of silently picking one. -
Data assessment
Check whether the reconciled data is complete enough to support the decision. Where it is not, say so before any analysis is presented, not after. -
Cohorting
Group customers by acquisition source, timing and behaviour, so performance is compared like for like rather than averaged into a blended figure that describes nobody. -
Measurement
Establish what actually happened: realised revenue per cohort, acquisition cost against it, payback timing, the shape of the value curve, the pattern of lapses. -
Projection and attribution
Where a forward-looking answer is needed, project it from behaviour already observed, then attribute the result back to the specific behaviours driving it, so a ranked list of causes comes out rather than a single unexplained score. -
Validation
Test against customers and periods held back from the analysis, check that the ranking is stable, and identify where the evidence is thin. -
Recommendations
Connect the evidence to the commercial decision, with the uncertainty stated in the same breath as the number.
What Makes It Different
Most revenue analysis stops at a figure: a blended CAC, a churn rate, a lifetime value with no error bar. Attribution runs the other way. For every projection, the contribution of each underlying behaviour is measured and reported, so what comes out is a ranked set of causes with evidence attached to each line. Your team can interrogate it, disagree with it, and check it against what they already know about the business.
The same discipline applies to reconciliation. When the aggregator report and the settlement file disagree, the discrepancy is quantified and shown rather than resolved quietly in favour of whichever number is more convenient.
Under The Hood
The pipeline is deliberately unglamorous, and we will walk your data team through all of it:
- Validate and join the customer, event, billing and acquisition data, reporting what fails validation rather than dropping it silently
- Engineer behavioural features from event history: recency, frequency, trend and change-point signals across engagement, payment and usage
- Fit and compare candidate models on those features, selecting on performance against customers the model never saw rather than on preference or habit
- Attribute each prediction back to per-driver contributions, so the output is a ranked driver list rather than a score
- Aggregate billing and acquisition cost into observed value curves per cohort and channel, and report payback from what was actually billed
- Report where accuracy is weakest, rather than quoting one headline number
The techniques are standard and available to anyone. Choosing the right definitions, reconciling sources that were never meant to agree, and knowing when the data will not carry the conclusion, that is the part that decides whether the output is worth acting on.
Confidence Matters
Every recommendation should be traceable to evidence. When confidence is limited, the report says so.
That makes the work useful even when the answer is not certain: your team can act on the strong signals, treat the weak ones cautiously, and fix the measurement where the current data cannot support a decision at all.
