Demeter

Its backtest

What the strategy did on the history it was tested on, from 3 May 2017 to 18 September 2026, as BotFactoryFX measured it when it went into quarantine, with the costs it pays there. A simulated account of 10,000 in the market’s currency, with a fixed size; every figure in % of that account.

Trades

93 in 9.4 years, 0.8 a month

Won

76%

Profit factor

2.37

Result

+133.5%

Worst drawdown

−30.3%

Average per trade

+1.44%

Average win · average loss

+3.25% · −4.43%

Best trade · worst trade

+18.15% · −21.51%

Longest losing streak

4 trades in a row

Years in loss

2 of 10

The result, trade after trade

Its drawdowns

How far below its previous high it was at each closed trade: what you would have had to sit through.

Year by year

The first and the last year are only part of a year.

How it was found and judged

BotFactoryFX found it in the 24/7 factory on 31 October 2026, the best of 2,582 candidate rule sets that search tried on GDX, gold miners ETF, daily candles.

Judge’s score

84/100 · Solid

On a similar market

GDXJ, the junior gold miners ETF: 85/100 · Solid (correlation 0.96)

Judged on

history the search never saw, from 19 October 2023

History it was found on

from 1 March 2017 to 30 September 2026, 2,388 candles

Checks measured

13 of 13

The judge is BotFactoryFX’s Reality Check: 13 checks on the strategy, nine of them on history the search never saw. Our bar to put one in quarantine is Solid, 75 or more, on its own market and again on a similar one, judged from the same date.

The judge’s report, check by check

The search that found it on 30 October 2026 kept its score but not this report, so on 1 November 2026 we judged it again exactly as then: same history, costs and cut-off date. It gave the same score: 84/100, Solid. Holds up, including on other instruments and other timeframes. No check failed outright.

• Out of samplePassed−5 %

  Does it still work on history it never saw?

  Profit factor 3.23 in sample, 3.07 out of sample.

• Period consistencyPassed5 / 6

  Does the edge hold across time, or was it one lucky stretch?

  5 of 6 windows closed in profit. The best window alone is 30% of the gains.

• Walk-forwardPassed2.51

  On history it never saw, would it survive being re-tuned as you went along?

  Re-tuned on past data only, the forward trades come out at profit factor 2.51 over 27 trades, profitable in 3 of 4 windows; the tuning picked different settings in 4 of them.

• Other instrumentsPassed84 %

  Does the edge exist anywhere else, or only here?

  The same rules stay profitable on GDXJUSUSD, and not on the other 1. An edge that survives a change of instrument is harder to explain away as a fit to this history. There it was measured only from 2023-10-19, the stretch this one never saw: before that is the history it was built on. Its distances in pips were scaled to each instrument's typical bar: GDXJUSUSD ×1.62, XAUUSD ×2.28.

• Other timeframesPassed100 %

  Does the edge survive on another clock, or only on these bars?

  Rebuilt on other clocks it keeps 100 % of its edge — H4 1.38 · W1 12.71, against 3.10 on the original. The edge is not an artefact of where the bars were cut.

• Better than chanceWarningP95

  On history it never saw, does it beat that same history with the direction of each bar drawn at random?

  Versions of that same unseen history with the direction of each bar drawn at random beat it 7 times out of 139. Profit factor 3.07 against a median of 1.18: better than chance, but not by a margin worth much.

• Parameter sensitivityPassed96 %

  Does it still work, on unseen history, if the numbers were slightly different?

  Nearby settings keep 96 % of the edge — a plateau, not a lucky point.

• DrawdownPassed13.7 %

  How deep underwater did you have to sit? A warning only: it does not count toward the score.

  Peak-to-trough 13.7 %, including the swing of open positions.

• Monte CarloPassedPF 1.30

  On history it never saw, how bad does it look if the trades had fallen differently — or without its best ones?

  Resampling the 29 trades it took on unseen history 1000 times, the profit factor of the worst 5 % is 1.30 one by one, 1.34 in blocks of 3 consecutive trades, and 2.39 skipping one in ten at random; the worst of the three, 1.30, is the score. Without its 1 best trade it is 2.60.

• Sample sizePassedt = 2.3

  On history it never saw, is the average trade distinguishable from zero?

  Over 29 trades on unseen history the average result stands 2.3 standard errors from zero — short of the 3 this asks for, so the edge could still be noise. The three best account for 25 % of gross profit.

• Cost tolerancePassed> 30 pips

  How much broker cost can it absorb before the edge dies?

  The edge is far larger than any plausible execution cost.

• PerturbationWarning59 %

  On history it never saw, does it hold up if that stretch had begun elsewhere and its prices had wobbled?

  Started at 40 random points of the unseen stretch with prices nudged by up to 15% of ATR, it keeps 59 % of its edge — profit factor 3.07 untouched against 2.23 in the middle of the runs, and 1.65 in the worst tenth.

• Bar resolutionPassedunambiguous

  Does it hold up when measured on finer bars?

  No exit depended on the order in which the high and the low happened inside a bar, so nothing here rests on an assumption about it.

What the report does not cover

• Everything above was measured on 2,388 bars — 9 years 7 months, 2017-03-01 to 2026-09-30. Nothing here was tested outside that window.
• That window does not include the Brexit referendum, the Swiss franc unpegging in January 2015 and the 2008 collapse, so no check above could tell you what this strategy does in one.
• Deflated Sharpe: counting the 2,582 rule sets the search tried, there is a 63 % probability that this strategy's Sharpe on the history it was chosen on is more than luck — 0.519 per trade over 62 trades, against 0.447 expected from the best of that many rule sets with no edge. The spread of the tried rule sets' Sharpes is taken as that of strategies with no edge, because the search does not record it yet. It looks only at the stretch where the strategy was chosen: a pattern that was real there and then stopped, or a flaw in the data, scores as high as an edge.
• This was the best of 2,582 different rule sets the search tried. What a search that size scores on data with no edge in it was measured on hourly currency pairs, not on this instrument and timeframe, so there is no reference figure for it here: the more candidates a search tries, the more likely its best is best by luck.

Live, in quarantine

Since 31 October 2026, on prices it had never seen, with closed trades only.

Quarantine verdict: Too early to tell

At its backtest’s pace, by now it would have closed about 0.2 trades and made +0.3%.

Its closed trades

← Back to every strategy in The Factory