The prediction-market news radar's recorded calls: edge or noise after fees?
Does buying the side each recorded call named, at the logged quote and held to settlement, make money after the exchange's fee, and does it beat simple rules that ignore the radar on the same markets?
The record cannot tell an edge from noise. Each of the radar's 200 recorded calls (15 October to 11 November 2026, UTC) was priced as one contract of the side it named, bought at the logged quote and held to settlement. Of the 144 that had resolved, on 86 events, 59 matchups and 26 call days, the mean was +4.35 cents a contract after fees, with a 95% interval of -6.67 to +14.28 cents: that spans zero, so the record cannot tell +4.35 cents from nothing, and the study's own word for the result is underpowered. Buying YES on every market the radar flagged made +11.59 cents on the same calls, more than the radar's +4.35, so the radar's choice of side added nothing measurable (difference -7.23 cents, interval -21.74 to +7.07); counting each game or dated question once, the mean was -2.87 cents. 56 of the 200 calls (28%) had no outcome yet, and which ones are open depends on the outcome, so the settled set is tilted toward early YES results and the mean is biased up. Nothing was pre-registered: this is a retrospective record, written on 21 November 2026 after every result was seen, and it changes none of them. One quantity was judged, so the family is 1; that is not a pre-registered N.
The report's label is exploratory because nothing was pre-registered. The study's own word for the result is underpowered: its 95% interval, -6.67 to +14.28 cents a contract after fees, spans zero. Publication records that as descriptive, not as a pass or a fail of a pre-registered test, and the study judged one quantity, so the family is 1.
Backtested by QS Research
Pre-registration
There was no pre-registration. This is a retrospective record, written on 21 November 2026, ten days after every result was seen; it records the question, data, rules and judging exactly as the committed code and write-up show they were run, and changes none of them. Nothing was fixed in advance, so no result of this study counts as a pre-registered test. The date that sits with this summary is the date the record was written, not a registration date. The rules as run: each call is one contract of the side it named, bought across the spread at the logged quote (YES at the ask, NO at one minus the bid), held to settlement, paying the taker fee 0.07 x P x (1 - P) per contract at entry, rounded up to the next cent. Intervals are 95% percentile bootstraps that resample whole events, matchups or call days, 4,000 resamples, seed 13, and the verdict reads the widest of the three. Controls buy the same calls at the same quote and fee: always YES, always NO, always the favourite and the other side. A 14-check audit of the record looks for lookahead and selection. The write-up judged one quantity, the settled calls' mean after fees against the controls, so the family is 1; that is the one judged quantity, not a pre-registered N. No holdout exists for this source, and a separate later forward test of the same radar is not part of this record.
Registered 2026-11-21Trials · 0 trials · 0 passed · judged at N = 1 trials; 2965 trials run in total
| Trial | Verdict | |
|---|---|---|
| always YES | RESULT | not counted |
| always NO | RESULT | not counted |
| always the favourite (the side priced above 50c at the mid) | RESULT | not counted |
| the other side | RESULT | not counted |
| full | RESULT | not counted |
| lean | RESULT | not counted |
| NO | RESULT | not counted |
| YES | RESULT | not counted |
| 1 first half | RESULT | not counted |
| 2 second half | RESULT | not counted |
| first call per matchup only | RESULT | not counted |
| leave one matchup out | RESULT | not counted |
| entry 60 s after the call | RESULT | not counted |
| entry 300 s after the call | RESULT | not counted |
| league-table mix, BEFORE fees | RESULT | not counted |
| entry quote precedes the call | RESULT | not counted |
| market unresolved at the entry quote | RESULT | not counted |
| call follows its news | RESULT | not counted |
| repeat stories | RESULT | not counted |
| market open at the call | RESULT | not counted |
| result matches the exchange | RESULT | not counted |
| one result per market | RESULT | not counted |
| source P&L reproduces | RESULT | not counted |
| settlement censoring | RESULT | not counted |
| record window | RESULT | not counted |
| repeated exposure | RESULT | not counted |
| one fixed rule | RESULT | not counted |
| no-skill control | RESULT | not counted |
| fills | RESULT | not counted |
Limitations
- Nothing was pre-registered: this is a retrospective record written on 21 November 2026, after every result was seen, and no result here counts as a pre-registered test. The verdict thresholds (30 settled calls, 10 call days, 25% open, 6 hours between repeat calls) were committed in the same commit as the result, so the record cannot show that they were chosen before the numbers were seen.
- The sample is the newest 200 calls the radar's service returned, gathered on 11 November 2026: a tail of the history cut by date, not the whole record. The market data and the scanner log are not committed, and only a 16-character prefix of the record file's hash is written down, so the inputs cannot be checked byte for byte and nothing here was re-run. The binary that produced the result is not recorded; the commit that added the code and the result together is assumed.
- 56 of the 200 calls (28%) had no outcome, above the 25% bar, and which calls are open depends on the outcome: a by-date market that closes early has resolved YES, so the settled set holds the early YES outcomes and the open set holds the rest. The mean is therefore biased up, for the radar and for always YES alike, and the study's own censoring check fails.
- 144 settled calls sit on 59 games or dated questions, and one game holds 19 of them. The intervals resample whole clusters, but a percentile bootstrap on a few dozen clusters still under-covers, and the stated win-rate interval treats calls as independent and is too narrow.
- The scanner's bar changed during the record: 137 calls on its lean bar and 63 on its full bar, the first lean call on 30 October 2026. The record carries no scanner version, so this is the record of two rules, and a change of thresholds inside a bar cannot be seen.
- No order was ever placed: every call is a model fill at the logged quote for one contract, so depth, queue and the delay between the quote and a real order are untested. Fees are the exchange's taker fee rounded up per one-contract order; settlement is free.
- The audit's 'no-skill control' row says no naive control clears zero, but always YES was +11.59 cents and its lower bound was -0.01 cents, and it beat the radar. The row is true only on the strict test and should not be read as the radar showing skill; the write-up says the same in its own reading.
- No family was declared and no correction for multiple looks was applied: the 7 breakdown tables, 5 sensitivities and 14 audit checks are descriptive and no cell is a finding. The one judged quantity is not deflated against the 2,965 trials QS Research had run when the report was built.
- The one cell labelled hindsight in the write-up is the time-to-close breakdown: the exchange's close time today moves for a market that closed early because its question was decided, so a short time to close moves with the outcome and is not a filter. That table is not among the published ones.
The numbers: every trial's checks, metrics, curve and yearly results, and the provenance
1 published version · current 201a7e369e5a0f66, published 2026-11-21