Illustrative figures based on logged runs
Every entry rule tested before this one fired at 3 or 4 minutes into a 5-minute window, because that is where the books looked deepest. Nobody had measured the last 30 seconds. When someone did, the sign flipped.
The rule
At 285 seconds into a 5-minute crypto window, buy the favourite if its ask is between 50 and 80 cents and the spread is 5 cents or less. Hold to settlement. Optionally require the spot feed, read two seconds behind the book, to agree with the favourite's side; that filter roughly doubled the margin on the largest coin and held out of sample on four others.
What the book does near the close
The favourite is underpriced only in a narrow band before the close: the book carries more doubt than the oracle average will resolve. Before 270 seconds the fee eats the edge; after 290 the remaining asks are dust.
How it was checked
- Selected on one season, applied unchanged to two others; positive on all three.
- Search-corrected: the whole scan that found it was re-run on a synthetic zero-edge market 200 times, and the real result beat all of them.
- Four coins the search never touched were tested once with the frozen rule: all positive.
- Two refinements that looked strong on one coin (weekday only, wide books only) reversed on the others and were dropped.
What it returns
Illustrative: $1,000 at 2 percent stakes, about 70 signals a day across seven coins, 60 days: roughly $2,700. Larger stakes hit the top of book, which is $3 to $300 a side depending on the coin.
Why most bots cannot trade it
And it decays: within a week of the rule going live, the share of windows with any late ask at all went from 23 percent to 84 percent empty. That story is in what happens after you find an edge.