Cohort revenue forecasting
Cohort Prediction
See what this week's installs will earn through D365. Compare revenue, ROAS, ARPU, and payback while there is still time to move budget.
- Forecast horizon
- D365
- Blind backtest · D365
- 5% to 10% MAPE
- Protected method
- U.S. Patent 12,014,278 B1
- Revenue · D365
- $42,827
- Gross forecast
- ROAS · D365
- 131.4%
- Return on spend
- ARPU · D365
- $4.83
- Value per install
- Installs
- 8,860
- $32,593 spend
| Install date | Campaign | Country | D7 | D30 | D60 | D90 | D180 | D365 |
|---|---|---|---|---|---|---|---|---|
| 2026-07-06 | US Prospecting | US | $1,948 | $5,161 | $5,569 | $5,816 | $6,719 | $9,504 |
| 2026-07-13 | Android Scale | DE | $1,904 | $5,044 | $5,443 | $5,685 | $6,567 | $9,289 |
| 2026-07-20 | Broad Value | DE | $2,136 | $5,658 | $6,106 | $6,377 | $7,367 | $10,420 |
| 2026-07-27 | Tutorial Test | US | $1,197 | $3,171 | $3,422 | $3,573 | $4,128 | $5,839 |
| 2026-07-06 | Broad Value | CA | $954 | $2,528 | $2,728 | $2,849 | $3,291 | $4,655 |
| 2026-07-06 | Tutorial Test | JP | $640 | $1,694 | $1,828 | $1,909 | $2,206 | $3,120 |
Patented model
The model predicts each user, then totals the cohort
Lemon learns from early behavior and static acquisition context instead of extending one average curve across every install.
Two views of each user
- User features
- App, country, device, acquisition source
- Action sequence
- Event, time, value, and changing event parameters
Predict future value
Encoders learn from the user features and the action sequence, combine those representations, and produce a predicted value for that user.
Aggregate the decision view
Individual predictions roll up into cumulative cohort revenue, ROAS, ARPU, and payback for each selected horizon.
U.S. Patent No. 12,014,278 B1 · granted June 18, 2024
Method for automated prediction of user data and features using a predictive model
Blind backtest
See the model prove itself on your history first
Setup includes a forecast of mature historical cohorts without showing the model what those cohorts eventually earned.
5% to 10%
MAPE at D365
- 01
Hide the mature D365 outcomes from the model.
- 02
Forecast each cohort from its earlier user behavior.
- 03
Compare the forecast with the revenue that actually arrived.
| Cohort | Forecast | Actual | Error |
|---|---|---|---|
| Jan cohort | $104,800 | $100,000 | 4.8% |
| Feb cohort | $116,000 | $125,000 | 7.2% |
| Mar cohort | $152,740 | $140,000 | 9.1% |
See the cohort forecast backtest audit for the leakage, error, coverage, and budget-decision checks behind a useful result, then derive a D7 ROAS target from mature cohorts before turning the forecast into a scale, hold, or cut rule.
Managed setup
Start with your history and a blind backtest
Lemon maps the source data, configures the model, and validates the result before the forecast workspace is activated for your team.
- 01
Map the source data
Pseudonymous identities, events, revenue, and acquisition context.
- 02
Run the historical backtest
Forecast mature cohorts blind and inspect the D365 error on your app.
- 03
Activate the workspace
Open approved horizons, filters, breakdowns, and saved views for the growth team.
- Identity
- Pseudonymous tracker or MMP ID
- Behavior
- Timestamped user actions
- Value
- Purchase and ad revenue events
- Acquisition
- Source, campaign, creative, country, app
Names, email addresses, and phone numbers are not required. Exact history and volume requirements depend on the app and configured horizons, so they are determined during setup rather than represented by a universal minimum.
FAQ
Questions growth teams ask about cohort forecasts
What does Cohort Prediction forecast?
Cohort Prediction forecasts cumulative revenue, ROAS, ARPU, and payback through the configured horizon up to D365. The workspace keeps observed and forecast periods visibly separate, supports gross and net revenue, and can combine or isolate in-app purchase and advertising revenue.
How does the prediction model work?
Lemon combines early action sequences with static user and acquisition features to predict future value for each pseudonymous user. Those individual predictions are then aggregated into the cohort totals your growth team evaluates. The underlying automated prediction method is covered by U.S. Patent 12,014,278 B1.
How is forecast accuracy checked?
Before your workspace goes live, Lemon hides the mature outcomes of historical cohorts, predicts their D365 value from earlier behavior, and compares each prediction with the revenue that later arrived. This blind backtest is included in setup. The D365 backtest range is 5% to 10% mean absolute percentage error.
What data is required?
Setup uses pseudonymous tracker or MMP identifiers, timestamped user actions, revenue events, acquisition dimensions, and relevant static device or attribution features. Names, email addresses, and phone numbers are not required. The exact history and volume needed depend on the app and are determined during setup.
How much does Cohort Prediction cost?
Cohort Prediction is enabled with the Lemon team. Setup costs $1,399 once per account and includes data mapping, model configuration, and the blind backtest on historical cohorts. After setup, the capability costs $679 per month for each app using the forecast workspace.
Availability
See what your cohorts will earn
The one-time setup includes data mapping, model configuration, and the blind backtest on your historical cohorts. The workspace is then enabled per app.
- One-time setup · per account
- $1,399
- Per month · per app
- $679