kestrel@atheron:/scores$
// Scores
Every test, kept.
Honest, day by day.
Each time Kestrel is tested, the result goes in this log, good or bad. You see its net return after real costs, its worst dip, how sure we can be, and what changed. Nothing is edited or removed, and so far nothing has passed.
// Strategies
Every test, in money
Each strategy Kestrel tested on days it never trained on, with every trade sized at 100. What it won, what it lost, and what was left after real fees, spread, slippage and funding. Paper only.
At real costs: taker fees both ways, half the spread, slippage and funding as recorded.
Expected: most candidates fail. The search exists to find the few that pass. This is not the model's result.
None yet. Only a strategy that passes the full gate counts as a result.
Passed means it made money after real costs on unseen days and stayed within the drawdown cap. Beating holding is shown, not required.
- Lost
- Won
- Net after costs
Net return after fees, over time
Each mark is one evaluation: Kestrel's net return after fees. Above the dashed line it made money, below it lost money. Hollow marks are runs found invalid later.
Spot is judged by what it makes after real fees and by its worst dip, not against holding: on spot, fees mean trading will rarely beat simply holding, so that comparison is not a fair test. Leverage is still compared with holding.
bps means basis points: 100 bps is 1%. A period is how long the specialist keeps a trade.
No evaluation has been logged yet.
Kestrel reader
The Kestrel reader reads the market alongside Kestrel's own model and leans long or short, with an honest confidence. Only a read that is confident enough, with a move worth its costs, becomes a signal. Each read is scored once its time is up, after fees. A read can go short, as a leverage trade does, so it is also compared with simply holding.
Since Oct 13, 2026, reader version twin-read-v1
| Date | Reads scored | Reads that traded | Reader after fees | Simply holding | Against holding | Running total against holding | Worst dip |
|---|---|---|---|---|---|---|---|
| Oct 25, 2026 | 200 | 200 | −85.7 bps | −682.8 bps | +559.6 bps | +2598.1 bps | 0.0 bps |
| Oct 24, 2026 | 223 | 223 | −4.7 bps | −7.1 bps | −15.5 bps | +2038.5 bps | −15.9 bps |
| Oct 23, 2026 | 294 | 294 | +37.4 bps | −236.6 bps | +16.7 bps | +2054.0 bps | −0.4 bps |
| Oct 22, 2026 | 191 | 191 | −23.4 bps | −16.5 bps | −17.1 bps | +2037.3 bps | −17.1 bps |
| Oct 21, 2026 | 676 | 676 | −69.0 bps | −163.0 bps | +53.3 bps | +2054.4 bps | 0.0 bps |
| Oct 20, 2026 | 423 | 423 | +7.7 bps | +112.2 bps | +28.6 bps | +2001.1 bps | 0.0 bps |
| Oct 19, 2026 | 1,871 | 107 | +49.3 bps | +273.6 bps | +31.0 bps | +1972.6 bps | 0.0 bps |
| Oct 18, 2026 | 760 | 33 | +9.8 bps | +13.8 bps | +186.4 bps | +1941.6 bps | 0.0 bps |
| Oct 17, 2026 | 31 | 1 | −1137.9 bps | −371.2 bps | −106.9 bps | +1755.2 bps | −106.9 bps |
| Oct 16, 2026 | 80 | 4 | +231.5 bps | +327.7 bps | +275.9 bps | +1862.1 bps | 0.0 bps |
| Oct 15, 2026 | 220 | 10 | +237.3 bps | +351.8 bps | +317.9 bps | +1586.3 bps | 0.0 bps |
| Oct 14, 2026 | 102 | 5 | +184.9 bps | −286.6 bps | +667.0 bps | +1268.4 bps | 0.0 bps |
| Oct 13, 2026 | 139 | 11 | +103.0 bps | −441.2 bps | +601.4 bps | +601.4 bps | 0.0 bps |
The locked six-month exam
Not yet used. Six recent months are kept away from all training. A version sits this exam once, only after it passes every other test.
Confidence (PSR) says how sure we can be that a result is real and not luck. It is shown beside each result and no longer decides a pass: a pass needs a net return above zero after 1.5 times the real costs in every backtest, on unseen data, with the worst dip within the cap. Leverage must also beat holding.
Gate v2 is stricter. Its confidence (DSR) counts every design we tried, so trying many is no shortcut, and it checks Kestrel against simple strategies such as always staying out.
Every entry
Newest first. Nothing here is edited or removed. A run found invalid stays, marked with the reason, and a correction is a new entry.
No evaluation has been logged yet.
These are test results on past data, not a promise of future results. Paper trading only, no real money.
The log reads from our servers and updates within a few minutes of a new test.
// How it is judged
The bar a version has to clear
A specialist passes only when all of these hold, on history it never learned from. Spot is judged by what it makes after real fees and by its worst dip, not against holding: on spot, fees mean trading will rarely beat simply holding, so that comparison is not a fair test. Leverage is still compared with holding.
Beat cash after real costs
In every backtest, its net return after 1.5 times the real costs (trading fees on both sides, half the spread, slippage and funding) must be above zero, so it does better than parking the money in stablecoins. It is judged over at least 26 stretches of 14 days, and every result is shown again at twice the costs and at Kraken's fees.
Keep the worst dip within the cap
Its worst dip, the furthest it falls from its high, must stay within our cap, 30% for now.
Leverage must also beat holding
A leverage specialist must also beat simply holding the same coins, after the same costs. Spot is not held to that test.
Luck and simple rules, shown
Its deflated Sharpe ratio, which counts every design we ever tried, and how it does against simple strategies such as always staying out, a trend rule and random picks are shown beside every result. They no longer decide a pass.
One locked exam
Six recent months stay locked away. A version sits that exam once, after it passes everything else.
How we keep it honest
Every test, at least daily
Every evaluation is written down when it finishes, and a daily check is added even on days with no new test.
Nothing is rewritten
The log only grows. The database refuses any edit or deletion, ours included.
Mistakes stay visible
A run found invalid later stays in the log, marked invalid with the reason. The correction is a new entry.
Paper trading only until launch. ATON Kestrel is a tool, not financial advice, and nothing here promises a return.