Did our claims come true?
This site publishes three kinds of dated claim, and each was graded on its own page in its own vocabulary. This page states one rule and applies it to all three, so the records can be read against each other — and, more importantly, so the claims we could not decide are counted in public instead of disappearing.
The abstention count is the point. A hit rate whose denominator is chosen by the same code that reports it is not a measurement — any grader can look excellent by quietly declining the hard cases, and nothing in the output would show it. So every claim we refused to grade is counted above and itemised below, with the reason.
The rule
A claim is graded only when all three hold. If any one fails it is abstained, with a stated reason, and it stays in the denominator.
- Pre-registered. Recorded, with its due date, before the outcome was knowable.
- Unambiguous. It says something that could turn out false.
- Outcome observed. The deciding fact exists and was not inferred, interpolated or back-filled.
The registers
Claim: A price level, recorded before expiry, where option holders lose most.
Rule: Graded a hit when the final pre-expiry reading landed within 1% of the expiry-day close.
648 graded settlements, worth about 289 independent ones; 95% interval 39-50% (resampled by expiration date, 63 dates, because settlements on one date move together; how)
An expiration with no expiry-day observation is abstained rather than settled against the nearest day we did see.
Claim: A direction — higher or lower — recorded on the day it was asked, due 30 days later.
Rule: Graded a hit when the 30-day price move went the way the call said.
46 graded windows, worth about 9 independent ones; 95% interval 22-78% (derived from the windows’ measured overlap and the tickers’ measured co-movement; how)
- 34 abstained — the claim did not commit to anything checkable
All 46 decided calls said higher, so the hit rate is the share of those windows that closed higher: it measures the market, not the model. An "unclear" answer is recorded and abstained, never re-asked until it commits. 40 calls not yet due.
Claim: A date window for a company's next report, projected from its own filing cadence.
Rule: Graded a hit when the filed report date fell within 7 days of the projection. Walk-forward: only dates filed before the one being scored are visible to the projection.
4,727 graded reports, worth about 1,311 independent ones; 95% interval 89-92% (resampled by company, 135 companies, because one company's quarters move together; how)
Measured live: 39 reports graded on 39 companies, 76.9% inside the window published the day before the report and 23.1% outside it (95% interval 64.1% to 89.7%). 140 projections still stand. The live record
- 563 abstained — the register cannot produce a gradeable claim of this kind
Scoring starts at a company’s fifth report, because the projection needs three dates and two usable gaps before it can say anything. Those earlier reports happened and we hold their dates, so they are abstained rather than excluded. That warm-up is also why the low tier carries zero scored points: this backtest cannot test the bottom of its own scale, and reports that rather than printing 0%. The replayed projections were not recorded before their outcomes, so they do not meet the first test of the rule; the live line grades the projections we logged as we published them.
Claim: A cluster on a liquidation map — the largest, and the nearest on each side — recorded with its distance from the mark the day it was drawn, due 14 days later.
Rule: Reached when the venue's own daily candles traded into the band within 14 days. The same test is run on a mirror band at the same distance in the opposite direction, where no cluster sits.
1,612 graded claims in 119 coin-fortnight blocks; 95% interval -0.2 to +7.8 points (resampled by coin and fortnight; 62 coins over 3 fortnights, and coins that move together inside a fortnight are not modelled, so the true interval is wider; how)
- 1 abstained — the deciding fact was never observed, and we do not interpolate it
851 claims not yet due. A cluster reached no more often than its mirror would mean clusters do not attract price, and we would publish that. Resampled by coin and fortnight, the interval on the gap spans zero, so this record cannot yet tell clusters from their mirrors. Rates by distance, and by which card the claim came from, are on the liquidation map methodology page: a pooled rate over mixed distances hides the only variable that matters.
Why we can publish this
We issue no positions and sell no signals, so a bad number here costs us nothing but the chance to look better than we are. A product that sells the calls has the opposite incentive, which is why so few publish a settled record: the same page that proves the method would price the product. That asymmetry is the only reason this page can exist, and it is worth more than any individual figure on it.
None of this is a forecast or a recommendation, and a past hit rate does not carry forward. It is a record of how our own published claims turned out, including the ones we could not decide.