Skip to content
KESTREL

kestrel@atheron:~$

// Our own AI model family

Built for trading.

Trained in the open, live.

ATON Kestrel is a small powerhouse we are building for trading, in the open. Our own market models, which we train on years of exchange data, will judge when a move is worth its fees, a second reader checks each call, and a small language model of ours explains it in plain words. No version has passed its tests yet, and you can watch every try.

Paper trading only until launch. ATON Kestrel is a tool, not financial advice, and nothing here promises a return.

// The score log

Latest results, honestly

Each specialist's newest fair test after real costs. Nothing has passed yet. When something does, it shows here first.

Latest scores

The newest valid test result for each specialist, after fees, on history Kestrel never learned from.

No evaluation has been logged yet.

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.

Kestrel reader
Average daily gap to holding
+204.7 bps
Worst dip
−106.9 bps
Reads scored
5,168

Since Oct 13, 2026, reader version twin-read-v1

These are test results on past data, not a promise of future results. Paper trading only, no real money.

See the full score log

// What it is

Four parts, one model

Each part has one job. They are released together, under one version number, so you always know which Kestrel made a call.

  1. 01

    The market model

    Our own models, learning from years of exchange prices. They judge when a move is worth its fees and stay out when it is not, trading only when the expected move is at least 1.5 times the full cost.

  2. 02

    The news reader

    It reads exchange announcements, regulators and the news as they publish, and scores the mood and how urgent it is. Urgent bad news, like an outage or a delisting, can pause a bot.

  3. 03

    The explainer

    A language model of our own writes each decision up in plain words: what it saw, what it did, and why. When a trade closes, it writes what happened.

  4. 04

    The second reader

    A larger AI model reads the same market less often: the charts, the order book, funding and the news. A trade needs both readers to agree, and in the Kestrel Eye you can ask both.

// Live build

Watch it learn

This is the real build, streamed a few seconds behind the work. Training starts again from scratch on its own server soon. Where it is, how far it has got, and what it is learning from.

Candidates tested togetherPaper only. No real money is traded.
-12,289.46 USD
Net after costs, 100 USD
Won
+5,838.98 USD
Lost
-18,128.26 USD
Results
77
0 passed
Trades
79,302

Passed means it made money after real costs on unseen days and stayed within the drawdown cap. Beating holding is shown, not required.

Every test, in moneyEvery version, compared

Training logLive
  1. kestrelMoving on. Marking which past moments would have paid after costs, so Kestrel has honest examples to learn from.
  2. kestrelxs volatility-lowvol-ls vol30-hold14 (kucoin perp) on 8h (long) put a hunch to the test: the calmest coins tend to do better than the wildest ones over time. After costs it finished 0.002% a day above break-even and beat simply holding by 0.021% a day. So, no pass. The same rule gets its turn on 12h later in this run.
  3. kestrelKestrel is working through the examples now, building a sense of which signals are worth following and which are just noise.
  4. kestrelDownloads, last 30 minutes. Passes finished: 103. Markets up to date: 26. Not finished: 10, to be asked for again.
  5. kestrelThe idea here is simple: when bitcoin sits above its long average, short the broader market on 12h bars. Across 640 trades, costs came to 0.002% against a gross of 0.03%. 0.029% a day above break-even after costs | beat simply holding | Not a pass. On to 1d next.
  6. kestrelOn to the next part. Marking which past moments would have paid after costs, so Kestrel has honest examples to learn from.
  7. kestrelAll on paper: across 118 runs this hour, only 4 finished above break-even after costs, 36 were eaten by fees alone, and 26 showed no edge at all. The best of the lot was xs volatility-lowvol-ls vol60-hold28 (kucoin futures) on 1m at 0.012% net. Next up is 3m.
  8. kestrelNext up was ts momentum-tsmom-90 tsmom-90d (kraken futures) on 1d (short). The idea: what has been rising tends to keep rising for a while. 0.044% a day above break-even after costs, and it beat simply holding. Noted, with suitable suspicion. That is not a pass under the gate. Later in this run it tries 3d, a slower clock.
  9. kestrelThe premise is that the calmest coins quietly outlast the wildest ones over time. Across 106 trades on 8h bars, gross came to 0.005% against costs of 0.002%. 0.002% a day above break-even after costs | beat simply holding | Not a pass. Moving to 12h next.
  10. kestrelMeanwhile, in the data room, last 30 minutes: download passes finished, 96; markets caught up to their newest candle, 34. Stalled: 5, back in the queue.
  11. kestrelKestrel is working through the examples now, sorting the signals worth following from the ones that are just noise.
  12. kestrelTested xs volatility-lowvol-ls vol60-hold28 (binance perp) on 8h (long), on the idea that the calmest coins tend to do better than the wildest ones over time. Net of every fee: 0.005% a day above break-even, and it beat holding by 0.013% a day. Not a pass. Next for this rule: 12h, one step slower.
  13. kestrelNow labelling the historical moments that would have cleared costs, giving Kestrel honest ground truth to train on.
  14. kestrelHousekeeping for the last 30 minutes. Download passes done: 105. Markets fully current: 29. Timed out: 10, with another try on the next pass.
  15. kestrelxs volatility-lowvol-ls vol30-hold14 (kucoin perp) on 8h (long) put a hunch to the test: the calmest coins tend to do better than the wildest ones over time. After costs it finished 0.002% a day above break-even and beat simply holding by 0.021% a day. So, no pass. The same rule gets its turn on 12h later in this run.
  16. kestrelKestrel is working through the examples now, learning to tell a real signal from background noise.
  17. kestrelSo far, 118 on the board. Above break-even after real fees: 4. Counted in money at 100 USD a trade, the board is 22,501.02 USD down in total. The best was xs volatility-lowvol-ls vol60-hold28 (kucoin futures) on 1m, 0.012% a day above break-even. 36 had an edge before costs and traded it away in fees. Another 26 had no edge, before costs or after. Everything here is paper, tested on history.
  18. kestrelDownloads, last 30 minutes. Passes finished: 88. Markets up to date: 17. Not finished: 21, to be asked for again.
  19. kestrelThe premise here is that calm coins outperform the volatile ones over time, tested long on 8h bars across 452 trades, with gross of 0.006% trimmed by costs of 0.001%. 0.005% a day above break-even after costs | beat simply holding | Not a pass. On to 12h.
  20. kestrelNext stage. Marking which past moments would have paid after costs, so Kestrel has honest examples to learn from.
  21. kestrelThe idea that calm coins outlast the wild ones got 106 trades on 8h bars: gross came in at 0.005%, costs took 0.002%, and a 15.9% drawdown along the way. 0.002% a day above break-even after costs | beat simply holding | Not a pass. Moving to 12h next.
  22. kestrelKestrel is working through the examples now, building a sense of which signals have edge and which are just noise.
  23. kestrelMeanwhile, in the data room, last 30 minutes: download passes finished, 104; markets caught up to their newest candle, 11. Stalled: 8, back in the queue.
  24. kestrelThe board so far: 110 in all, 25 above break-even after costs. Add it all up at 100 USDT a trade and the board is 40,355.66 USDT down. Top of the board: xs momentum-xs-mom-30-28 lb30-hold28 (kraken spot) on 1m, 0.12% a day above break-even. The recurring lesson, in 31 of them: something found, then paid it all to trade. 49 had no edge and were losing before a single fee. Paper only, as always.
  25. kestrelNext up was xs volatility-lowvol-ls vol60-hold28 (binance perp) on 8h (long). The idea: the calmest coins tend to do better than the wildest ones over time. 0.005% a day above break-even after costs, and it beat simply holding. Noted, with suitable suspicion. That is not a pass under the gate. Later in this run it tries 12h, a slower clock.
  26. kestrelThe idea is that quieter coins outlast the wild ones over time; on 8h bars with 106 trades, gross came to 0.005%, costs took 0.002%, and the drawdown reached 15.9%. 0.002% a day above break-even after costs | beat simply holding | Not a pass. Moving to 12h.
  27. kestrelHousekeeping for the last 30 minutes. Download passes done: 109. Markets fully current: 9. Timed out: 10, with another try on the next pass.
  28. kestrelDownloads, last 30 minutes. Passes finished: 110. Markets up to date: 18. Not finished: 9, to be asked for again.
  29. kestrelAll on paper: across 118 markets on 1m bars, only 4 finished above break-even, 36 were eaten by costs, and 26 showed no edge at all. The best of the hour was xs volatility-lowvol-ls vol60-hold28 (kucoin futures) at 0.012% net; the combined paper result across $100 stakes came to $22,501.02 USD. Next up is 3m.
  30. kestrelMeanwhile, in the data room, last 30 minutes: download passes finished, 104; markets caught up to their newest candle, 16. Stalled: 24, back in the queue.
  31. kestrelHousekeeping for the last 30 minutes. Download passes done: 101. Markets fully current: 16. Timed out: 7, with another try on the next pass.
  32. kestrelDownloads, last 30 minutes. Passes finished: 108. Markets up to date: 14. Not finished: 7, to be asked for again.
  33. kestrelMeanwhile, in the data room, last 30 minutes: download passes finished, 80; markets caught up to their newest candle, 7. Stalled: 9, back in the queue.
  34. kestrelHousekeeping for the last 30 minutes. Download passes done: 99. Markets fully current: 4. Timed out: 11, with another try on the next pass.
  35. kestrelDownloads, last 30 minutes. Passes finished: 84. Markets up to date: 2. Not finished: 24, to be asked for again.
  36. kestrelDownloads, last 30 minutes. Passes finished: 42. Stalled: 8, back in the queue.
  37. kestrelMeanwhile, in the data room, last 30 minutes: download passes finished, 68. Timed out: 13, with another try on the next pass.
  38. kestrel29 downloads wrapped up in the past 30 minutes, all on paper data being prepared for Kestrel's runs.
  39. kestrelMeanwhile, in the data room, last 30 minutes: download passes finished, 122; markets caught up to their newest candle, 5. Not finished: 15, to be asked for again.
  40. kestrelHousekeeping for the last 30 minutes. Download passes done: 103. Markets fully current: 17. Stalled: 9, back in the queue.

Housekeeping for the last 30 minutes. Download passes done: 103. Markets fully current: 17. Stalled: 9, back in the queue.

Times in UTC.

// How it is judged

A release has to earn it

No version reaches members because it trained. It reaches them because it passed.

Profit after real costs

Every specialist is tested on history it never trained on. Its net return after real costs must be above zero and its worst dip must stay within a cap; a leverage specialist must also beat holding. None has passed yet.

One exam, once

The latest six months are kept out of all training. Each release sits one exam on them, once, so the result is not tuned to the test.

A founder presses Publish

A version that passes waits with its release notes and evidence. It reaches members only when a founder approves it.

// The Kestrel Eye

See markets the way it does

A command centre for the market: Kestrel in the middle, glowing with the Mood index, with live alerts, the news and both readers around it. Focus on one coin, keep a watchlist, set your own layout or go full screen. Free for everyone while we are in paper trading.

// Build diary

Written up as it happens

At every milestone, the diary says what changed and what it means, in plain words.

// What Kestrel noticed today

A quiet day for the news reader

Today the build of ALPHA v0.1.0 is in the harvest stage. The news reader took in 0 items. The overall mood reading was 0.06, which is a reading of news tone and not a signal. There were 0 storms of urgent news and 0 pauses. No specialists reported and no milestones were reached today.

Milestone ·

A fresh start for ALPHA v0.1.0

None of the three specialists, swing, breakout and trend, passed its gate, so the exam was never sat. Entries on 13 and 14 October said the build had reached the exam stage. That was wrong, and their counts of complete exchanges did not agree. The three models are retired and their derived training data cleared. The next design tests simple, well-known rules on every Bitunix and Coinbase market, then learns which of their trades are worth the fees. Training starts again from scratch on its own server, and each stage will show on the live page. The score log keeps every earlier result. Paper trading only until launch.

Milestone ·

All 7 exchanges are now complete

The exchange data for ALPHA v0.1.0 is complete. All 7 exchanges now have a complete history. The three specialists, swing 1h, breakout 4h and trend 1d, move to the exam stage. Each must still beat simply holding the market, after fees, to pass the gate.

Hear when it ships

Release notes when a version is published, and a short recap of the build each week. Confirm your address and leave any time.