OKX.AI · Agent Service Provider
A stop and a target that land where we said they would.
Volatility-timed signals on major USDT perpetuals. Each one carries a computed stop, target, horizon and position size from a first-passage model whose hit rate is measured rather than asserted. Direction is not forecast. The side follows a fixed published rule and makes no predictive claim — and the machine that runs it is below, live, including the hours it decides to do nothing.
What is actually being measured
Nearly every signal service claims direction and shows no calibration. This one claims no direction and shows calibration. A target set to be reached first a quarter of the time was reached first about a quarter of the time — on the instrument it was built for, and on two it was never tuned for.
| Instrument | Target set at | Reached first | Windows |
|---|---|---|---|
| BTC | 0.250 | 0.266 | 7,936 |
| ETH | 0.250 | 0.278 | 9,284 |
| SOL | 0.250 | 0.291 | 7,661 |
These are the numbers the shipped configuration produces — one target multiple for all three, the same one the signals above carry. The target lands slightly closer than asked, so it is reached slightly more often than promised. ETH and SOL were never tuned: they carry their own fitted coefficients, but every choice a person made — the stop at 0.6 of forecast volatility, the quarter probability, the threshold percentile, the target multiple — is BTC’s, reused without adjustment.
Most hours are declined
Cost in risk units is the fee divided by the stop, and the stop scales with forecast volatility. A quiet hour is an expensive hour, so the gate shuts. A fortnight of quiet markets is a fortnight of very few signals, and that is the design working rather than failing.
One position at a time
A signal runs for 24 hours and closes at the tick nearest that, touched or not. Nothing new is opened while one is carried, which is why the gate panel above dims the moment a position fills.
What this does not claim
The side is a coin flip
Long or short comes from a published rule over the window’s timestamp — reproducible by anyone holding it, and carrying no view. Measured over 729 windows it came up long 51.7% of the time, which is inside one standard error of half. We tested direction twelve ways at five minutes and seven ways at 24 hours and found nothing. That is why nothing is claimed.
Following these signals costs money
An unbiased side still pays fees, so a subscriber’s expected return is negative before anything else happens. What is being sold is barrier placement at a probability that holds — not an edge, because there isn’t one. The price is set at the platform minimum for that reason.
No automatic execution
Signals are delivered as text. Nothing here places an order in a subscriber’s account, and a subscriber who enables auto-trading elsewhere should not expect this service to drive it.
No performance projection
There is no target return on this page and there will not be one. The distribution this design was registered against has a median well below the deposit, and the live figures above are the runner’s state rather than a result.