How to read a simulated number
These are not forecasts.
A synthetic population is more agreeable than a real one. It reads what is in front of it and does not get interrupted the way people do. So when Prefigure says 41% reached value, the useful claim is not 41. It is that roughly two in five got there, and that the losses are where the report says they are.
Four ways to use it well
- Compare, do not forecast. Run version A, change one thing, run version B. The gap is the finding.
- Read the quotes before the chart. The explanation of why holds up far better than the size of the number.
- Trust behaviour over stated intent. Where someone left is evidence. Whether they say they would pay is not.
- Check the biggest prediction against reality. Ship the cheapest fix to real users and see whether the direction held.
What each number is worth
Where people leftHighCounted from the behaviour log, not asked for.
Reached the point of valueHighA session either got there or it did not.
Came back on day 1, 3, 7MediumA later session exists only if that person chose to come back. Real boredom is still under-modelled.
Could explain the promiseMediumAsked at the end, in their own words. Good at spotting confusion, imprecise as a score.
Would payLowStated intent overstates payment in every method ever devised. Rank versions with it; never budget against it.