AbyssWatch: every call, including the ones that failed

Where it stands

Not enough to say yet.

Reading what the bot has called so far.

How it decides

Three things have to happen, in order.

The bot never places an order. It watches, counts, and hands over the exact numbers if you want to use them.

Price comes back to a level the bot has marked

A marked level is a price where buyers or sellers showed up more than once before. Tested against levels placed at random, price reacts at these a little more often, but not by enough to pay after fees.

Two fifteen-minute candles close on the right side of it

A wick through the level does not count. Only closes count, and that is what keeps the bot out of the flush that stops everyone else out.

The trade is still worth taking after fees

If the target is near enough that Kraken's fees eat the edge, the bot says so and tells you to skip it.

Where it runs

The bot reads Kraken.

Every level on this page is drawn from Kraken's price feed, every ratio is worked out on Kraken's fee schedule, and the scout only sees pairs Kraken lists. The same coin charts much the same elsewhere, but the numbers here are Kraken's numbers. If you want to act on them as printed, that is the account they were built for.

Open a Kraken account

Spot and margin, the same fee schedule the numbers on this page are calculated against.

This is a referral link. If you sign up through it, Kraken pays me a share of your trading fees for 90 days. Your fees are the same either way, and I would tell you to use Kraken regardless, because it is the only exchange the bot reads.

Open an account

Live

Watching now

Levels the bot is waiting on, newest first. Nothing here is a trade yet. Where two rows share a level, the same setup is running with two different targets to find out which one is better.

Unprompted

Levels the bot found on its own.

Every half hour it screens all 600-odd Kraken dollar pairs, throws out everything too thin or too wide to trade, runs the level finder on the deepest books that are left, and names at most three. Nobody chooses the coins. Every one is written down the moment it is named, before anyone knows what happens next. Every fifteen minutes each one is replayed against the candles under the same rules a watched setup gets, taken or not, and the result is the last column. Most say still open: a call gets 48 hours before it is called anything. There is no hit rate here and will not be until twenty have resolved, because a rate off a handful of results is a number you can make say anything.

First day

What a new Kraken listing actually does.

When a pair opens for trading the bot starts a clock and records the price at 0, 15 minutes, 1 hour, 4 hours and 24 hours. Nobody publishes this, so there is no received wisdom about it worth trusting. The dataset is small and says so.

The record

Every call, settled

Each setup that finished, newest first. Cancelled means the level broke before the bot would commit to it.

Tested

What twelve years of history says

Every rule the bot uses can be run over Kraken's full trading history, 2013 to 2025, with the bot's own code and no changes. Each rule below has a name, a verdict, and the number of trades behind it. A loss is printed as a loss. A row says so when it used only part of that history. Everything from January 1, 2026 onward is held back and never looked at, so a rule that looks good on the earlier years still has to pass a test it has not seen. These numbers are an upper bound on live: no spread, entry at the closing price, no cap on how many setups run at once. Live would be lower.

confirm-2close-v1 the rule running now

A level, two fifteen-minute closes above it, enter at the second close, a fixed stop and target, forty-eight hours to resolve.

Loses. About a quarter of what it risks, per trade, over 9,410 tests.

rest-at-level-v1

Same levels, a resting limit at the level with no confirmation, same stop and target. Every touch fills, including the ones that then fail.

Loses more. About four tenths of what it risks, per trade, over 30,011 trades. Waiting for confirmation was the better of the two.

rest-after-refusal-v1

When the running rule refuses an entry because price ran past the level, rest a limit back at the level and wait, as the bot used to suggest.

Loses. The limit fills 87% of the time and loses about three tenths of what it risks per fill, over 4,627 fills. The bot no longer suggests it.

confirm-2close-v1, 48h window

The running rule gives a confirmed setup 48 hours to resolve. What happens to the ones that do not?

Measured, not changed. Of 878 setups that ran out the clock without touching the stop, the ones that resolved later won more than they lost. Folding that back in leaves the rule still losing, by a little less. The window stays at 48 until a change is written down and tested.

trail-under-level-v1

The running rule, with the stop moved to just under the level the moment it confirms. What the owner did by hand on one trade, tested on all of them.

Loses more. It gives up 0.407 of what it risks per trade over 11,812 tests, against 0.251 for the running rule. In 36.5% of its tests the stop fires within an hour of confirming: price dipping back through the level is the normal thing, and a stop that close stands right in its way.

atr-stop-v1

The running rule, with the stop set at one and a half times the coin's recent hourly range instead of by structure.

Loses, a little less. It gives up 0.219 of what it risks per trade over 9,371 tests, against 0.251 for the running rule. That gap is 2.01 standard errors, just past the two it takes to count. It still failed the bar set before the test, which asks for a rule that makes money by a margin clear of chance, so it does not get the held-back years.

btc-vs-20d-close-v1 a measurement, not a rule

Does it matter whether Bitcoin is above or below its own 20-day average when a level is named? Every pick now records that number, and the running rule was split on it over the same 9,410 tests.

Measured, not changed. Picks named with Bitcoin above its average did better than picks named below it, by a margin just large enough to count and small enough that it may not hold up. Both sides still lose. The number is shown on each pick above and decides nothing.

shadow-level-null-v1 a measurement, not a rule

Do the levels the bot picks do any better than levels picked at random? Every level the bot named was set beside a random level on the same coin at the same moment, with the same stop and target distances and the same fees, and both went through the running rule.

No difference in profit. The bot's levels give up 0.251 of what they risk per trade over 9,410 tests, the random ones 0.259 over 25,272. That gap is 0.62 standard errors, well inside chance. Price did hold two closes at the bot's levels a little more often, 34.67% of the time against 31.68%, and that gap is 12.00 standard errors, so it is real. Put plainly: the bot finds levels where price really does react a little more, and that reaction does not turn into profit after fees.

btc-above-sma-v1 not a rule the bot trades

Hold Bitcoin only while its daily close is above its average over the last 20, 65, 150 or 200 days, and hold cash otherwise. Tested over 2015 to 2025 with 0.35% paid on every switch, against 20,000 random timings per average that keep the same days in the market and the same number of switches.

No average passed the bar set before the test. To pass, the share of random timings doing as well as the rule had to be below 0.00135. It was 0.02205 for 20 days, 0.01205 for 65, 0.01105 for 150 and 0.09955 for 200. The 65 day and 150 day rules came closest. A live record of all four was started below. None of them is a working rule.

flash-drop-reversal-v1 (eight versions)

Buy a liquid coin after it drops 4, 6, 8 or 12 percent in one hour, sell after 4 or 24 hours. Eight versions, tested on Kraken history from 2016 to May 2022, with fees and slippage counted at 1.10 percent a round trip.

None passed. Before fees, every version made money, and more the deeper the drop and the longer the hold. After fees, and after raising the bar because about 50 tests had been run in all at that time, none cleared it. The gains were concentrated in 2021, and every version lost money from January to May 2022. The test only includes coins that were still trading after May 2022, which flatters it. The two 12 percent versions had too few events to count. Nothing was run on later years.

donchian-breakout-long-v1 (six versions)

Buy a coin when it closes above its highest price of the last 20, 55 or 100 days, and sell after it closes below its lowest price of the last 5 to 50 days. Six versions, tested on Kraken daily history from 2016 to May 2022, with fees and slippage counted at 1.10 percent a round trip.

None passed. Before fees, every version made large gains on average, but a few big winners drove the averages and most trades lost. The best month, January 2021, was the best for every version, and every version lost money on trades entered in early 2022. There were only 125 to 589 trades per version, too few under the test's own minimum of 1,000. The test only includes coins that were still trading after May 2022, which flatters it. Nothing was run on later years.

The one-hour surge test (six versions)

When a coin rises at least 2 percent in one hour, and at least 3, 4 or 5 times its own normal hourly move, buy the next candle and sell 4 or 24 hours later. Six versions in all. Kraken history from 2013 to May 2022 only. Later years stayed sealed.

None passed. The four hour versions made about nothing before fees, between minus 0.11 and plus 0.04 percent a trade, and fees and slippage cost 1.10 percent. The 24 hour versions made between 1.2 and 2.2 percent before fees, but five trading days carried all of the profit. Without those five days, every 24 hour version lost money after costs. None of the six beat the bar we set before looking. The coins in this test are the ones that were still trading in 2022, which flatters any buy rule. A pass would have been weaker than it looked. A fail is a fail.

The market flush test (four versions)

When at least 30 percent of the liquid coins on Kraken fall 3 or 5 percent in the same hour, buy each of those coins at the next candle and sell 4 or 24 hours later. Four versions. Kraken history from 2018 to May 2022 only, because before 2018 there were not enough coins trading at once to call anything a market-wide move. Later years stayed sealed.

None passed. Before fees, all four made money, from 0.3 to 2.9 percent a trade, and more for the deeper drop and the longer hold. After fees and slippage of 1.10 percent, only the 5 percent versions were positive, and neither came close to the bar set before looking. The idea behind this test was that a coin falling with the whole market bounces more than a coin falling alone. We checked that directly, and it is not true: the same drops on the same coins did as well or better when the market was not falling. One day, May 12, 2022, carried nearly all of the profit in one version. The coins in this test are the ones still trading in 2022, which flatters any buy rule.

The daily momentum test (four versions)

When a coin's daily close is at least 10 percent above where it closed 20 days earlier, buy at the next day's open and sell 5 or 20 days later. And the mirror: when it is at least 10 percent below, short it on margin for 5 or 20 days. Four versions. Kraken daily history from 2016 to May 2022 only. Later years stayed sealed.

None passed. Buying after a rise made money before fees, about 2 percent a trade over 5 days and 7 percent over 20, and stayed positive after fees. It still missed the bar by a wide margin, because these trades arrive in bunches on the same days and in the same months, so the real sample is far smaller than the trade count, and every version lost money once its five best months were taken out. Shorting after a fall lost money after fees and the margin cost of holding a short. The coins in this test are the ones still trading in 2022, which flatters the buy side and hurts the short side. One correction is on the record: the formula we wrote down for scoring the shorts was wrong, it overstated one trade on a coin that collapsed, and the backtest record shows both the number as registered and the corrected one. The verdict is the same either way.

Cross-sectional momentum (buy last month's winners)

Every Monday, rank the liquid coins by their return over the past 28, 56 or 91 days, buy the top 5 or 10 percent at the next open, sell seven days later. Six variants, 961 to 2,291 trades each on Binance history from 2017 to May 2022, Kraken costs applied (1.10 percent round trip).

All six failed. Before costs every variant made 0.3 to 1.2 percent a week on average. After costs five of six lost money, and the profit came almost entirely from 2021. Outside that one year every variant lost about 3 percent a week. Failed design test, October 2, 2026. Not traded. 76 candidates tested in all so far.

Results are pinned to the exact code that produced them. The first four rows: the fit run of September 26, 2026 (levelwatch.py 2f1cfc81, grade_scout.py 88be25fc, archive.py 3d95d851, replay.py 1fee3a15). The two stop variants: the runs of September 28, 2026 (levelwatch.py 4aff93fc, grade_scout.py ed909287, archive.py bfe54d47, replay.py 9110c9fc). The Bitcoin 20-day row: the run of September 27, 2026 (levelwatch.py 4d39a823, replay.py 6b89ced8). The random-level test: the runs of September 29, 2026 (replay.py 62326836, the rest as for the stop variants). The Bitcoin average test: the run of September 29, 2026 (trend_replay.py a726ae31, archive.py bfe54d47, replay.py 62326836). The flash-drop test: the run of September 30, 2026, on 2016 to May 2022 only (family_fd.py 679b0b39, archive.py bfe54d47, grade_scout.py ed909287, levelwatch.py c9283939, replay.py 62326836, supa.py 7e3ff3f2). The daily breakout test: the run of September 30, 2026, on 2016 to May 2022 only (family_dc.py 66c03a95, archive.py bfe54d47, grade_scout.py ed909287, levelwatch.py c9283939, replay.py 62326836, supa.py 7e3ff3f2). The one-hour surge test: the run of October 1, 2026, on 2013 to May 2022 only (family_fu.py 6b2cc00f, family_fu_ref.py 80c02ff4). The market flush test: the run of October 1, 2026, on 2018 to May 2022 only (family_mf.py 35720476, family_mf_ref.py a0035cf5). The daily momentum test: the run of October 2, 2026, on 2016 to May 2022 only (family_tm.py 745e7040, family_tm_ref.py 04824a76). Every later run of the bot's own rule reproduced the first run's numbers exactly before measuring anything new. Nothing about a rule changes after its number is seen. A change is written down first, measured, and only then given the held-back years, once.

Live signals

Bitcoin trend signals, recorded as they happen.

Four slow rules ask one question each day: is the daily close of Bitcoin above its average over the last 20, 65, 150 or 200 days? Above means hold, below means cash. We publish each answer as soon as the day closes, before anyone knows what comes next.

In the backtest, none of the four passed the bar we set for beating random timing, and the 65 day and 150 day rules came closest. So this record exists to test all four on data that did not exist when the rules were written. None of them is a working rule yet.

See how these rules were tested.

Your say

What would you want from this?

This is being built in the open and the record decides what it becomes. If there is something you would use, or something you would never trust, say so.

AbyssWatch.app

Every call it makes.
Including the ones that failed.