Methodology
What counts as a prediction
A public statement by a named person or organisation about something that will or will not happen, with a date or a window by which it can be checked, and a source anyone can read. Opinions, values and claims about the past are left out.
The record holds the predictions collected so far, not everything a figure has said. Famous misses are easier to find than quiet hits, so each scorecard shows how many predictions it rests on, and a figure with fewer than 3 resolved predictions gets no verdict.
How one prediction is scored
Each prediction scores the probability the figure gave to what actually happened.
| The figure said | It happened | It did not |
|---|---|---|
| It will happen (a flat call) | 1 | 0 |
| 70% chance it will happen | 0.7 | 0.3 |
| Even odds | 0.5 | 0.5 |
A prediction marked partly scores 0.5: the claim had two readings and the outcome matched one of them. Each one says which in its note. A prediction marked void could not be checked and is not scored.
Credibility and the half-life
Credibility is the weighted average of those scores, starting from a prior worth 2 predictions at 50%: one right, one wrong. Older predictions count for less. A prediction's weight halves every half-life, so with a 10 years half-life a call that resolved 10 years ago counts half as much as one that resolved today.
credibility = (1 + Σ weight × score) ÷ (2 + Σ weight), weight = 0.5age ÷ half-life
When nothing new resolves, every weight shrinks and the score drifts back toward 50%, a coin flip. That drift is the line between the marks on each chart. The half-life is yours to choose on every page; 10 years is the default, and "No decay" weighs every prediction equally.
Example: one right call, five years ago, with a 5-year half-life. Its weight is 0.5, so credibility is (1 + 0.5 × 1) ÷ (2 + 0.5) = 60%.
Verdicts
| Credibility | Verdict |
|---|---|
| 65% or more | Mostly right |
| 55% to 65% | Leans right |
| 45% to 55% | Coin flip |
| 35% to 45% | Leans wrong |
| below 35% | Mostly wrong |
| fewer than 3 resolved | Too few to score |
| less than half of one fresh prediction's weight left | Record faded |
A faded record has decayed so far that the score mostly reflects the 50% starting point. The all-time figure on each scorecard, which uses no decay, shows how that record looked in full.
Brier score
Shown for figures who stated probabilities. It is the mean of (stated probability − outcome)² over predictions that came out right or wrong, where the outcome is 1 or 0. Lower is better: 0 is perfect, and always saying 50% scores 0.25.
Dates
Dates are shown as precisely as the source gives them. For the arithmetic, a month means its 15th and a year means 1 July. Scores are computed as of the date shown on each page.
Sources and corrections
Every prediction links to a source. If a claim is misquoted, misdated or wrongly resolved, or a figure's record is missing a prediction, write to us with a link. Corrections change the score the next time the site is rebuilt.
What this is not
Not investment, political or safety advice, and not a measure of anyone's character. It measures one thing: how public, checkable predictions turned out.