Method in brief
The percentage is not a count of favourable articles, a measure of confidence in the prose or a promise about the future. It is our current estimate that a precisely defined event will resolve as true by its deadline.
Resolution before calculation
Define the event.
Every forecast begins with a Boolean claim, a deadline and a resolution test. A demonstration is not commercial adoption. A nomination is not an election victory. If the boundary is vague, the percentage has no stable meaning.
The outside view
Begin with the base rate.
We first ask how often comparable events have happened under comparable conditions and within a comparable period. That reference class supplies the prior probability. Its definition is recorded because changing the comparison group can change the starting point.
Successive requirements
Map the route.
When an outcome requires several stages, later stages are estimated conditionally. We ask for the probability of each step assuming the preceding requirements have already succeeded.
P(H) = P(R₁) × P(R₂ | R₁) × P(R₃ | R₁, R₂)Ordinary headline probabilities are not multiplied together unless independence is justified. Alternative routes are represented as branches, and shared requirements remain shared rather than being counted twice.
Worked chain
- Regulatory approval
- 70%
- Production at scale, given approval
- 60%
- Adoption, given both requirements
- 80%
- Combined route
- 33.6%
Known data points
Update what the evidence touches.
Evidence modifies the requirement it bears on. Internally, the update is expressed on a log-odds scale: how much more likely would this evidence be if the requirement succeeds than if it fails?
logit(p updated) = logit(p base) + Σ αg log(LRg)- Common origin. Ten reports repeating one company statement form one evidence group.
- Limited strength. Source quality, uncertainty and dependence reduce how far evidence can move the estimate.
- Known requirements. A completed step becomes 100%. A failed necessary step closes that route. Neither is counted again as supporting evidence.
- Visible judgement. Where a likelihood ratio is estimated rather than measured, that judgement remains part of the record.
LR is the likelihood ratio for an evidence group. α is a discount between zero and one for uncertainty and dependence.
The correction layer
Calibrate against reality.
The requirement model produces a raw probability. As the archive of resolved forecasts grows, we compare those estimates with observed frequencies. If events forecast near 70% occur materially less often, later estimates in that range should move down.
Calibration is applied only from past, resolved forecasts and tested out of sample. Until the record is large enough, any calibration remains provisional. Published percentages are rounded to avoid pretending to a precision the evidence cannot support.
Verification
Keep score in public.
A method improves only if its failures remain measurable. We evaluate resolved binary forecasts with proper scoring rules and compare them with a base-rate benchmark.
(p − y)²The primary error measure. Lower is better; zero is perfect.
−[y ln(p) + (1 − y) ln(1 − p)]A secondary score that penalises confidently wrong forecasts.
reliability − resolution + uncertaintyWhether stated probabilities match observed frequencies, and whether they separate likely events from unlikely ones.
We retain the initial estimate, subsequent updates and the final pre-resolution estimate. A late correction should not erase the earlier history.