Illustrative figures based on logged runs
Most "AI bots" are a model fitted to the same data it is judged on. This one was built the other way round: trained on one season, thresholds chosen on a second, tested once on a third, and then given a pass-or-stop bar before it touched money.
The model
A gradient-boosted tree on nine features read from the book and the spot feed: the ask, the spread, how many times the favourite flipped, whether spot agreed with the book and by how much, the hour, the second in the window, and the coin. It outputs a win probability; the bot fires when that probability beats the ask plus the fee by a margin, and takes the best second in the window.
The checkpoint
Before launch the bar was written down: after 50 live fills, stop if the outcome count is below expectation at the one percent level. After 89 fills the tally was 68 wins against 68.3 expected, plus 3.65 points per dollar against a backtest of 3.64. It passed, and it kept running.
The week it paid
The same model, fired from a server moved into the exchange's region with a websocket book mirror instead of polling, returned about 8 points per dollar over six days with every day positive. Illustrative: $500 at 10 percent stakes, about $1,100 after that week.
The week after
The next fortnight was flat. The pre-registered prediction about latency held; the market did not stay still. Per-coin retrains, a 15-minute version and a "switch to the best model of the last 12 hours" scheme were all tested and all lost to simply leaving the original model alone.
What to copy
- Split by time, not at random. Train, validate, and one untouched holdout you open once.
- Score against the executable ask at the true fee, one trade per window, never pooled across price levels.
- Write the stop rule before the first order. A checkpoint you set after seeing the results is not a checkpoint.
- Re-run the whole search on a zero-edge market and report that p-value, not the single-rule one. A 6-by-3 search on pure noise returns "+3.4 points" at the 95th percentile.