Lineage

Each bot runs a numbered strategy version with written parameters. A new generation is one change with a written reason, tested on data never used to design it. Nothing comes from an optimizer, and failed candidates stay on record.

6 candidates recordedNone promoted yet

Candidates

Tested on 72 datasets (3 windows each), from Kraken public OHLC (REST, no API key). Chance is corrected for the 6 candidates tested together.

Candidate	Change	Datasets ahead	Typical difference	Chance probability	Status
NEKTON-002	Candle width requirement	25 / 72	−0.20 pts	7.7%	Indistinguishable from chance
NEKTON-003	Trailing stop distance	33 / 72	−0.17 pts	100.0%	Indistinguishable from chance
NEKTON-004	Breakout lookback range	44 / 71	+0.26 pts	34.1%	Indistinguishable from chance
BENTHOS-002	Entry z-score threshold	29 / 72	−0.10 pts	75.0%	Indistinguishable from chance
BENTHOS-003	Average length	31 / 72	−0.16 pts	100.0%	Indistinguishable from chance
PLANKTON-002	Exit chance per candle	51 / 72	+0.11 pts	0.3%	Supported, not promoted

Running versions

No parent · Nekton

NEKTON-001

Confirmed breakout

First generation. Enters when the close breaks the 20-candle range on a wide candle (more than 1.5 × ATR) pointing in the breakout direction. Exits via a 2 × ATR trailing stop from the best close.

Lookback range: 20 closed candles
Entry rule: candle range > 1.5 × ATR(14), in breakout direction
Exit: trailing stop 2 × ATR from best close
Max exposure: 50%
Costs: 0.10% fee + 0.05% slippage per side

Return: +0.48%
Drawdown: −3.38%
Fills: 37

Simulated figures on the live 60-minute BTC/USD run. Recorded candidates: NEKTON-002, NEKTON-003, NEKTON-004.

No parent · Benthos

BENTHOS-001

Return to the mean

First generation. Enters when the price z-score against the 50-candle average leaves ±2, exits when the z-score returns to 0. No trend guard: it will fade any move.

Moving average: 50 candles
Entry rule: z-score outside ±2
Exit: z-score back to 0
Max exposure: 30%
Costs: 0.10% fee + 0.05% slippage per side

Return: −1.15%
Drawdown: −2.91%
Fills: 29

Simulated figures on the live 60-minute BTC/USD run. Recorded candidates: BENTHOS-002, BENTHOS-003.

No parent · Plankton

PLANKTON-001

Seeded random

First generation. The control: enters with a 5% chance per candle in a random direction, exits with a 10% chance per candle. Its random sequence is fixed by the seed, not by results.

Seed: 20261004
Entry chance per candle: 5%
Exit chance per candle: 10%
Position size: 2% to 10% of equity
Costs: 0.10% fee + 0.05% slippage per side

Return: −0.73%
Drawdown: −0.73%
Fills: 54

Simulated figures on the live 60-minute BTC/USD run. Recorded candidates: PLANKTON-002.

Candidate · parent NEKTON-001

NEKTON-002

Breakouts on less extreme candles

Indistinguishable from chance

The 1.5 × ATR requirement is rarely met, so many breakouts are skipped. This candidate tests whether breakouts on slightly narrower candles still mean something, without changing the exit rule.

What changed · 1 parameterCandle width requirement1.5 × ATR 1.2 × ATR

Windows ahead: 83 / 183
Datasets ahead: 25 / 72
Untouched: 33 / 216
Difference per dataset (95%): −0.74 … +0.11
Chance probability: 1.3%
After correction: 7.7%
Typical difference: −0.20 pts
Trades compared: 2,392 / 2,761

By market family

Holding up across several families means more than in just one.

crypto/USDahead 11 of 36
crypto/EURahead 8 of 16
crypto/cryptoahead 6 of 16
non-cryptoahead 0 of 4

Per timeframe

A change that only works on one timeframe may not really work.

15 min · active in 96%ahead 5 of 18
1 hour · active in 87%ahead 5 of 18
4 hours · active in 76%ahead 7 of 18
1 day · active in 80%ahead 8 of 18

NEKTON-002 was tested against NEKTON-001 on 216 windows from 72 datasets (2,392 parent trades versus 2,761). 33 windows were untouched by this change and were excluded. Windows within a dataset share an instrument and indicator history, so the test runs per dataset: 25 of 72 favour the candidate. A split that lopsided happens by chance 1.3% of the time; after correcting for 6 candidates tested, 7.7%. This result cannot be told apart from chance. The candidate stays on record and is not promoted. The candidate made 15% more trades, so part of the difference may come from trading costs rather than the rule itself.

Candidate · parent NEKTON-001

NEKTON-003

Giving trends more room

Indistinguishable from chance

A 2 × ATR trailing stop is often hit by ordinary noise before a trend develops. This candidate loosens it to 3 × ATR; the risk is giving back more open profit when the trend turns.

What changed · 1 parameterTrailing stop distance2 × ATR 3 × ATR

Windows ahead: 98 / 216
Datasets ahead: 33 / 72
Untouched: 0 / 216
Difference per dataset (95%): −0.43 … +1.01
Chance probability: 55.6%
After correction: 100.0%
Typical difference: −0.17 pts
Trades compared: 2,392 / 1,895

By market family

Holding up across several families means more than in just one.

crypto/USDahead 15 of 36
crypto/EURahead 4 of 16
crypto/cryptoahead 11 of 16
non-cryptoahead 3 of 4

Per timeframe

A change that only works on one timeframe may not really work.

15 min · active in 100%ahead 11 of 18
1 hour · active in 100%ahead 5 of 18
4 hours · active in 100%ahead 6 of 18
1 day · active in 100%ahead 11 of 18

NEKTON-003 was tested against NEKTON-001 on 216 windows from 72 datasets (2,392 parent trades versus 1,895). 0 windows were untouched by this change and were excluded. Windows within a dataset share an instrument and indicator history, so the test runs per dataset: 33 of 72 favour the candidate. A split that lopsided happens by chance 55.6% of the time; after correcting for 6 candidates tested, 100.0%. This result cannot be told apart from chance. The candidate stays on record and is not promoted. The candidate made 21% fewer trades, so part of the difference may come from trading costs rather than the rule itself.

Candidate · parent NEKTON-001

NEKTON-004

Breakouts from a longer range

Indistinguishable from chance

Breaking a 20-candle high or low happens fairly often. A 30-candle range makes breakouts rarer, in the hope that the ones left mean more.

What changed · 1 parameterBreakout lookback range20 candles 30 candles

Windows ahead: 109 / 184
Datasets ahead: 44 / 71
Untouched: 32 / 216
Difference per dataset (95%): +0.15 … +1.06
Chance probability: 5.7%
After correction: 34.1%
Typical difference: +0.26 pts
Trades compared: 2,392 / 1,966

By market family

Holding up across several families means more than in just one.

crypto/USDahead 19 of 36
crypto/EURahead 8 of 15
crypto/cryptoahead 14 of 16
non-cryptoahead 3 of 4

Per timeframe

A change that only works on one timeframe may not really work.

15 min · active in 76%ahead 7 of 17
1 hour · active in 83%ahead 8 of 18
4 hours · active in 87%ahead 14 of 18
1 day · active in 94%ahead 15 of 18

NEKTON-004 was tested against NEKTON-001 on 216 windows from 72 datasets (2,392 parent trades versus 1,966). 32 windows were untouched by this change and were excluded. Windows within a dataset share an instrument and indicator history, so the test runs per dataset: 44 of 71 favour the candidate. A split that lopsided happens by chance 5.7% of the time; after correcting for 6 candidates tested, 34.1%. This result cannot be told apart from chance. The candidate stays on record and is not promoted. The candidate made 18% fewer trades, so part of the difference may come from trading costs rather than the rule itself.

Candidate · parent BENTHOS-001

BENTHOS-002

Entering on smaller deviations

Indistinguishable from chance

±2σ deviations are rare. This candidate enters earlier at ±1.5σ to see whether more small trades beat a few extreme ones.

What changed · 1 parameterEntry z-score threshold2 σ 1.5 σ

Windows ahead: 83 / 214
Datasets ahead: 29 / 72
Untouched: 2 / 216
Difference per dataset (95%): −0.38 … +0.28
Chance probability: 12.5%
After correction: 75.0%
Typical difference: −0.10 pts
Trades compared: 1,742 / 2,312

By market family

Holding up across several families means more than in just one.

crypto/USDahead 16 of 36
crypto/EURahead 6 of 16
crypto/cryptoahead 7 of 16
non-cryptoahead 0 of 4

Per timeframe

A change that only works on one timeframe may not really work.

15 min · active in 100%ahead 6 of 18
1 hour · active in 100%ahead 7 of 18
4 hours · active in 98%ahead 8 of 18
1 day · active in 98%ahead 8 of 18

BENTHOS-002 was tested against BENTHOS-001 on 216 windows from 72 datasets (1,742 parent trades versus 2,312). 2 windows were untouched by this change and were excluded. Windows within a dataset share an instrument and indicator history, so the test runs per dataset: 29 of 72 favour the candidate. A split that lopsided happens by chance 12.5% of the time; after correcting for 6 candidates tested, 75.0%. This result cannot be told apart from chance. The candidate stays on record and is not promoted. The candidate made 33% more trades, so part of the difference may come from trading costs rather than the rule itself.

Candidate · parent BENTHOS-001

BENTHOS-003

A faster average

Indistinguishable from chance

A 50-candle average lags price, so 'returning to the mean' can take a long time. A 30-candle average stays closer to the current price.

What changed · 1 parameterAverage length50 candles 30 candles

Windows ahead: 97 / 216
Datasets ahead: 31 / 72
Untouched: 0 / 216
Difference per dataset (95%): −0.18 … +1.31
Chance probability: 28.9%
After correction: 100.0%
Typical difference: −0.16 pts
Trades compared: 1,742 / 2,626

By market family

Holding up across several families means more than in just one.

crypto/USDahead 15 of 36
crypto/EURahead 6 of 16
crypto/cryptoahead 8 of 16
non-cryptoahead 2 of 4

Per timeframe

A change that only works on one timeframe may not really work.

15 min · active in 100%ahead 5 of 18
1 hour · active in 100%ahead 3 of 18
4 hours · active in 100%ahead 8 of 18
1 day · active in 100%ahead 15 of 18

BENTHOS-003 was tested against BENTHOS-001 on 216 windows from 72 datasets (1,742 parent trades versus 2,626). 0 windows were untouched by this change and were excluded. Windows within a dataset share an instrument and indicator history, so the test runs per dataset: 31 of 72 favour the candidate. A split that lopsided happens by chance 28.9% of the time; after correcting for 6 candidates tested, 100.0%. This result cannot be told apart from chance. The candidate stays on record and is not promoted. The candidate made 51% more trades, so part of the difference may come from trading costs rather than the rule itself.

Candidate · parent PLANKTON-001

PLANKTON-002

Holding random positions longer

Supported, not promoted

A control for the control: does holding time alone, with no rule at all, change the results? If so, differences in the other bots should be read more carefully.

What changed · 1 parameterExit chance per candle0.10 0.05

Windows ahead: 147 / 216
Datasets ahead: 51 / 72
Untouched: 0 / 216
Difference per dataset (95%): −0.44 … +0.08
Chance probability: 0.1%
After correction: 0.3%
Typical difference: +0.11 pts
Trades compared: 3,888 / 2,808

By market family

Holding up across several families means more than in just one.

crypto/USDahead 25 of 36
crypto/EURahead 12 of 16
crypto/cryptoahead 11 of 16
non-cryptoahead 3 of 4

Per timeframe

A change that only works on one timeframe may not really work.

15 min · active in 100%ahead 17 of 18
1 hour · active in 100%ahead 16 of 18
4 hours · active in 100%ahead 16 of 18
1 day · active in 100%ahead 2 of 18

PLANKTON-002 was tested against PLANKTON-001 on 216 windows from 72 datasets (3,888 parent trades versus 2,808). 0 windows were untouched by this change and were excluded. Windows within a dataset share an instrument and indicator history, so the test runs per dataset: 51 of 72 favour the candidate. A split that lopsided happens by chance 0.1% of the time; after correcting for 6 candidates tested, 0.3%. That is evidence worth recording, but not a verdict: this corpus is one period of history and mostly crypto. Nothing is promoted automatically. The candidate made 28% fewer trades, so part of the difference may come from trading costs rather than the rule itself.