Series · Game analytics · 8 parts
Cohort LTV Forecasting
Spend to CPI to installs to actives to spenders to revenue per user to LTV. Seven links, each with its own estimator, and a chain whose weakest point is not where anyone expects it.
Parts
Forecasting a Monthly Cohort, End to End
Spend to CPI to installs to actives to spenders to revenue per user to LTV. Seven links, each with its own estimator, and a chain whose weakest point is not where anyone expects it.
The Price of a Player: Forecasting CPI
Forecasting cost-per-install from a UA programme that ended. Fifteen aggregate models and a structural family, backtested over seventeen rolling origins on a real account: the textbook log-linear trend finishes near the bottom, a weighted average of cell-level prices wins, and the edge is the media plan, not the granularity.
Weekly or Monthly? What the Grain of a Forecast Is For
The same structural CPI model rebuilt at weekly grain on eighty-five rolling origins. The ranking replicates, weekly wins the quarter-ahead plan, only the weekly monitor detects a regime break, and the whole result reverses on cohort curves. Not a contest between grains but a division of labour.
Running the Whole Chain: What a Revenue Forecast Is For
Every stage had been validated on its own. Run in series over forty origins, the composed chain loses to a drift line at one month and beats it by half from three months on, calls direction far better than level, and turned out to have been running a model nobody selected.
Each Cohort a Little Worse Than the Last
Successive acquisition cohorts decline in quality, and a composed forecast over-projects because of it. Five investigations across two titles: a shape drift wired to the wrong cohorts, a leading suspect that was innocent, a media-mix hypothesis that was a time trend in disguise, and the one correction that replicated.
Where the Uncertainty Actually Was
A posterior predictive band for the composed revenue forecast, and what it found: the default empirical band labelled 80% covered 61%, a shared calendar shock dominates band width at every horizon, and cohort behaviour — the thing four investigations had tried to improve — is the smallest term.
Does the LTV Forecast Work on Real Data?
The synthetic answer, the real one, and the gap between them. Error falls predictably with the observation window on both titles, which is the result that matters. But the fitted curve is only a modest improvement on a ratio a studio can compute in a spreadsheet, and a synthetic benchmark overstates the model, not the problem.
When a Cohort Stops Dying
The deep-age regime, a handover that inverted, and a grid that answered a different question. Past the first year a cohort fluctuates rather than drains; per-age decay beats survival at depth; selecting one specification per metric halves held-out error; and no age-dependent rule survives out of sample at all.
A Correction That Worked Until the Model Under It Changed
The calendar-period term. A constant level correction that helped by twelve percent, then doubled long-horizon error once an upstream fix changed the shape of the bias it was correcting. Short, because the result is a reversal, and the reversal is worth more than the result would have been.