Every studio does this and almost every studio does it differently.
You have market intelligence on comparable titles. You have soft launch data from two or three small geos. The two disagree, sometimes by a factor of two. And you need one number to put in a plan that someone will hold you to.
What usually happens next is a baseline plan and an aspirational plan, and a number picked somewhere between them because it feels defensible. Then the launch happens, the number is wrong, and nobody can say which of the twenty assumptions inside it was the one that broke.
There is a better way to do this, and it is not more precision. It is being explicit about which of your two data sources you are trusting for which variable, and why.
The two sources measure different things
This is the part that gets skipped. Market intelligence and soft launch data are not two estimates of the same quantity. They are estimates of different quantities, and blending them numerically produces something that is neither.
Market intelligence tells you about a market. Comparable titles, in the geos you care about, at scale, with a mature feature set and a marketing budget behind them. What it cannot tell you is whether your game is one of those titles.
Soft launch tells you about your game. Real players, real retention, real conversion. What it cannot tell you is how any of that behaves at a hundred times the spend, in a different country, with the features you have not shipped yet.
So the question is never "which number do I believe". It is "which variable does each source have authority over".
Splitting the variables
Work through the model one input at a time and assign each to a source. A rough allocation that holds for most launches:
Soft launch has authority over:
- Early retention shape. D1 to D30 is your game. It is the most reliable thing you own and it is measured on real players doing real things.
- Conversion to payer. Also your game. The proportion of installs that ever spend is a product property more than a market property.
- The early monetisation curve. What a payer is worth in their first month.
Market intelligence has authority over:
- Long-tail retention. Your soft launch is eight weeks old. It cannot tell you D180. Comparable titles can, approximately.
- eCPI at scale. Your soft launch bought installs in cheap geos at low volume. Neither of those survives a global launch.
- Ceiling effects. What the best-performing titles in your genre actually converge on, which bounds how optimistic your extrapolation is allowed to be.
Neither has authority over:
- The features you have not shipped.
- Your organic install loop at scale, because it barely exists at soft launch volume.
- Geo mix, which changes every revenue and cost assumption at once.
That last category is where forecasts actually go wrong, and it is the category people spend the least time on because there is no data to point at.
The gap nobody models
Here is the specific failure. A soft launch in a cheap tier-two geo, running a few thousand dollars a day, should generate almost no organic installs. There are not enough players to produce word of mouth, store ranking movement, or the visibility that brings in users you did not pay for.
That is reasoning rather than measurement, and it is worth holding as an assumption rather than a fact. If your soft launch is producing meaningful organics at low volume, the rest of this section matters more to you, not less, because your global number is further out than you think.
So the soft launch model has an organic contribution of roughly zero, and that gets carried into the global forecast as a small flat uplift, if it gets carried at all.
At global launch scale that is wrong in both directions at once. Organics become a meaningful share of installs, which makes your paid spend look better than your attribution tool says. And the size of that contribution depends on your retained player base, which means it compounds, which means a flat multiplier cannot represent it.
The honest version is that your retention is an acquisition channel. Model it as one, and the payback number changes materially. Model it as a constant, and you will be wrong in a way that is invisible until the quarter closes.
How materially: on the calculator's default inputs, with no organic contribution, payback lands on day 207. Move the organic share to 25 per cent and change nothing else, and it lands on day 122, because each paid install is now carrying a third of another install with it and the true cost per install drops from $1.10 to $0.83.
Open that comparison and move the organic slider yourself. Those are the calculator's published defaults with one input changed, not a customer result.
Bounding rather than guessing
Once each variable has a source, the next move is not to produce one number. It is to produce a range and then find out which input the range is actually made of.
Run the model three times. Once with soft launch authority applied everywhere it is defensible and conservative market assumptions elsewhere. Once with the optimistic reading. Once with what you actually believe.
Then rank the inputs by how much each one moves the answer.
Almost always, one or two assumptions account for most of the spread, and they are frequently not the ones the room has been arguing about. Teams will spend an hour debating eCPI to two decimal places when the forecast turns almost entirely on D30 retention and the organic coefficient.
Knowing which input owns the uncertainty tells you what to do next, and that is the actual output of a forecast. If the answer turns on D30 retention, extend the soft launch. If it turns on eCPI at scale, buy a small test at real volume in a real geo before committing the budget. If it turns on a feature you have not shipped, the forecast is a product question wearing a marketing costume.
What to put in the plan
One number, with the three things that would change it named underneath, and the measurement that would settle each one.
That is defensible in a way that a single point estimate is not. It also answers the question your CFO is actually asking, which is not "what will happen" but "how wrong could this be and how would we find out early".
And it sets up the thing almost nobody does afterwards, which is going back six months later and asking whether it worked. A forecast with named assumptions can be audited. A number picked between a baseline and an aspiration cannot, which is why so few studios can tell you why their last plan missed.
Running it
Kadar models this directly: retention, conversion, monetisation and spend day by day per cohort, including the organic installs your retained players generate. Changes carry dates, so a feature or a monetisation switch goes in on the day you would ship it. And it ranks which single assumption is driving your payback day, which is the step this piece is really about.
If you have no cohort data yet, it ships fitted defaults across five genres and three monetisation models, so a pre-launch title starts from a curve rather than a blank sheet.
The free calculator does the first part of this: three retention points, a year forward, your payback day and the organic loop. The dated changes and the sensitivity ranking need an account, which is free for one product.
Run your own numbers in the free calculator. No sign up.