Everybody on prediction-market Twitter shows you their wins. We grade every engine we publish — in public, losers included — and put the receipts on one page.
Here is the whole board. It is read live from the same database that powers our public track record, so it is current as of the date stamped on it rather than the day somebody last edited this article.
Our edge engines have settled 3,143 timestamped signals against live Kalshi market resolutions — 1,504 winners, 1,456 losers, a combined 51% win rate, and +$104.69 net buying one contract per signal (+3.3¢ average per contract). Data as of 2026-09-04, read live from the same source as /track-record.
| Engine | Settled | Win rate | Net P/L | Avg / contract | Avg Brier |
|---|---|---|---|---|---|
| Bitcoin Edge | 825 | 60% | +$48.53 | +5.9¢ | 0.1489 |
| Mispricing Scanner | 633 | 29% | −$24.31 | −3.8¢ | 0.2132 |
| Oil Edge | 564 | 47% | +$7.60 | +1.3¢ | 0.1834 |
| Gold Edge | 493 | 62% | +$35.47 | +7.2¢ | 0.1599 |
| Silver Edge | 401 | 59% | +$43.35 | +10.8¢ | 0.1464 |
| Weather Edge | 194 | 34% | +23¢ | +0.1¢ | 0.2004 |
| Oracle Gas Edge v1 | 25 | 8% | −$5.94 | −23.8¢ | 0.3877 |
| Arb Scanner | 8 | 50% | −24¢ | −3.0¢ | — |
Every signal is recorded before settlement and graded automatically against the market resolution. Net P/L is the sum across every settled signal, buying a flat 1 contract each time; Avg / contract is that sum divided by the signal count. Win rate excludes pushes. Brier score measures calibration — lower is better, 0.25 is coin-flip guessing. Negative engines stay published; that's the point. Plays of the Day are hand-picked rather than algorithmic, so they are graded separately on /track-record and are not in this table.
How to read this table
Start with the column most sites never show you: net P/L.
Win rate on its own is close to meaningless in prediction markets, and it is trivially easy to inflate. Buy heavy favorites at 90 cents all day and you will post a win rate north of 80% — right up until the ninth trade takes back everything the first eight paid. A contract that costs 90 cents to win 10 pays you a dime and charges you ninety.
The reverse is just as common in the other direction. Oil Edge sits below 50% and is still net positive, because the positions it takes are cheap enough that being right less than half the time still pays. That single row is the argument for reading a scoreboard properly.
So the three columns work together:
- Win rate — how often the call was correct. Necessary, not sufficient.
- Avg / contract — what one contract actually returned, on average, across every settled signal. This is the number that decides whether an engine made money.
- Avg Brier — whether the stated probability was honest. Lower is better; 0.25 is coin-flip guessing.
An engine that clears all three is working. An engine that clears one is a marketing claim.
Why Brier is the column we care most about
A Brier score grades calibration, not hit rate. When an engine says 70%, calibration asks whether that outcome really happens about 70% of the time — not whether it happened this once.
That matters because every sizing decision downstream depends on the probability being true. Kelly sizing takes your estimated edge and tells you how much of the bankroll it justifies; feed it an overconfident probability and it will confidently tell you to over-commit. A well-calibrated 61% is worth more than a poorly-calibrated 80%, every time.
You can see the same probabilities the engines publish on any of the live boards — Bitcoin Edge, Gold Edge, Silver Edge — and grade them yourself as they settle.
The honest part
Three engines in that table are losing money: Mispricing Scanner, Weather Edge, and Oracle Gas Edge. They are not hidden, not footnoted, and not quietly retired — they sit in the same table, in the same font, as the ones that work.
That is deliberate. A track record that only contains winners is not a track record, it is a highlight reel, and it tells you nothing about what the next signal is worth. Publishing the failures is the only thing that makes the successes checkable.
Where an engine is being rebuilt, we say so and pull sizing while the model is reworked. We would rather tell you than let you find out with your own bankroll.
Want the rest of the toolkit?
The graded engines above are one slice of what we run. The full set — EV Calculator, Kelly sizing, Probability Converter, Bayes Updater, Arb Scanner and more — lives on the free tools page. Free shows direction; Pro shows the signed edge in pp on every strike.
The same engines are also exposed to AI agents over our MCP server, if you would rather ask Claude or ChatGPT for the read than click through a board.
Frequently asked questions
Which 7 Oracles engines are profitable?
Bitcoin Edge, Silver Edge, Gold Edge and Oil Edge are net positive across every settled signal. Mispricing Scanner, Weather Edge and Oracle Gas Edge are net negative, and they stay published in the same table as the winners. The live figures for all of them are in the scoreboard on this page and on our public track record.
Is a high win rate proof a prediction-market tool works?
No. Buying heavy favorites at 90 cents will post a win rate above 80% and still lose money, because the rare loss costs nine times what a win pays. Read per-contract P/L and the Brier calibration score next to the win rate — a tool is only working if all three hold up.
What does the Brier score measure?
The Brier score grades calibration rather than hit rate: whether a stated 70% probability is right about 70% of the time. Lower is better, 0 is perfect, and 0.25 is what coin-flip guessing scores.
Trade responsibly. Markets are speculative. Past tool performance does not guarantee future results.