Which markets suit the bot
How results change with market conditions
The bot only BUYS, so it lives off periods when gold is clearly rising. This section sums up what past data shows — and says plainly what nobody has measured yet, so you don't have to guess.
19 years: which were up, which were down
Yearly profit/loss risking $50 per trade, re-run on Dukascopy price data 2005–2023 with the exact validated logic (Wilder ATR). Exploratory backtest, not validation. Red = losing year.
Classic Pullback C_V1
Dukascopy 2005–2023 · Losing years: 2012, 2013, 2015, 2018, 2021 (5/19)
Exness 2016–2023 · Losing years: 2018, 2021 (2/8)
Near-Touch Pullback C_V1_P1b
Dukascopy 2005–2023 · Losing years: 2012, 2013, 2018, 2021 (4/19)
Exness 2016–2023 · Losing years: 2018, 2021 (2/8)
3-Bar Pullback C_V1_P2
Dukascopy 2005–2023 · Losing years: 2012, 2013, 2014, 2015, 2018, 2021 (6/19)
Exness 2016–2023 · Losing years: 2018, 2021 (2/8)
Every Signal D_V1
Dukascopy 2005–2023 · Losing years: 2012, 2013, 2018, 2021, 2022 (5/19)
Exness 2016–2023 · Losing years: 2018, 2021, 2022 (3/8)
London Open 7AM L07S
Dukascopy 2005–2023 · Losing years: 2012, 2013, 2015, 2018, 2021, 2022 (6/19)
Every Signal Max 3 D_V1_CAP3
Dukascopy 2005–2023 · Losing years: 2012, 2013, 2015, 2018, 2021 (5/19)
Dukascopy figures run on continuous data. A formal test once showed a lower mean R (+0.13R vs +0.24R, 2016–2023), but that gap came from missing days in that test's dataset; on continuous data, year-by-year results match Exness data. This is still a historical test and does not guarantee future results.
The chart stops at 2023. From 2024 on is the holdout: exploratory numbers for that period are not shown, because they are exactly what “saw” the holdout in advance. The only official 2024+ result is on the Holdout page; the MT5 table under Results still shows 2024–2026 as the EA's own backtest.
The bot does better when
Gold is in an uptrend
The bot only buys. In years when gold rose steadily it tended to profit; in years of long declines it tended to lose.
Show the numbers
- Dukascopy 2005–2023 (24/7 version): Near-Touch lost in 4/19 years; Classic Pullback and Every Signal in 5/19; 3-Bar Pullback and London Open 7AM in 6/19. All 5 lost in 2012, 2013, 2018 and 2021.
- 2012 and 2013 were back-to-back losing years for all 5 strategies. Over 2012–2014, simply buying and holding gold lost about 24% (Dukascopy data).
Source: trading-research-lab/.ai/evidence/champion-rerun-2026-09-17/web-numbers-2026-09-19/result-24h.md (Dukascopy by year, 24/7 version)(not published) · web-numbers-2026-09-19/result.md (London Open 7AM)(not published)
The bot trades every hour of the week
The released version has no session filter: every valid signal is taken, including the Asian night session. The figures on this site are for that version.
Show the numbers
- The 0:00–6:00 UTC block (Exness 2016–2023) was profitable: Classic Pullback 288 trades +0.14R; Near-Touch 411 trades +0.25R; 3-Bar 386 trades +0.38R; Every Signal 503 trades +0.22R.
- The 20:00–23:00 UTC block was slightly negative for the Pullback group. Versus the 7:00–19:00 filtered version: about 27% more trades, while average profit per trade barely moves (+0.262R vs +0.271R).
Source: trading-research-lab/.ai/evidence/champion-rerun-2026-09-17/web-numbers-2026-09-19/result-24h.md (00–06 UTC block, 24/7 vs 07–19 comparison)(not published)
Consider switching off, or not starting, when
2018 was the worst year: all 6 strategies lost
On MT5, all 6 strategies lost money in 2018, four of them about $4,000 at $50 risk per trade. 2019 was slightly negative; 2021 and 2022 also lost. One year like that is enough to wipe out the room on a 50K prop account.
Show the numbers
- 2018: Classic Pullback −$4,013 (profit factor 0.21; 53-trade losing streak) · 3-Bar Pullback −$4,064 · Near-Touch −$3,810 · Every Signal −$3,653 · Every Signal Max 3 −$2,756 · London Open 7AM −$1,215.
- Multi-year drawdowns are larger still, because losing years stack up.
Source: data/backtests.json (MT5 Strategy Tester, 24/7 version)(not published)
You can't stomach a long losing streak
Most losing trades are small and close fast; the profit comes from a handful of trades that run very far. If you plan to switch off after a few losing weeks, you may switch off right before the big winner — and keep only the losses.
Show the numbers
- Removing exactly the 10 best trades out of 909–1,665: average profit drops 42–60% on Exness 2016–2023 (Classic Pullback: +0.262R down to +0.104R). Over 19 years of Dukascopy data the loss is 20–28%.
- Winners are held a median of 85–121 hours (3.5–5 days); losers are cut after 7–11 hours. So each winner carries 3–5 nights of swap and usually sits through a weekend.
- The backtest history includes losing streaks of 29–81 trades and periods of 1.9–6.1 years below the equity peak. For Every Signal on Exness, one 1,235-day drawdown had still NOT recovered by the end of the data.
Source: trading-research-lab/.ai/evidence/champion-rerun-2026-09-17/web-numbers-24h-2026-09-20/result.md (sections 2–4, 24/7 version)(not published)
A prop firm account is close to its loss limit
The bot can hold several trades at once. If the account has little room left before the firm's cut-off, one reversal can cost you the account. Switch off or lower the risk per trade.
Show the numbers
- Largest MT5 drawdown at $50/trade, continuous run 2018 → Aug 2026 (after swap and commission): Every Signal $7,585, Near-Touch $4,235, 3-Bar Pullback $4,186, Classic Pullback $4,082, Every Signal Max 3 $3,123, London Open 7AM $1,952. So 5 of 6 exceed or come close to the $4,000 room of a 50K prop account; only London Open 7AM stays under $2,000. Closed trades only; with floating losses it is larger.
Source: data/backtests.json (MT5 Strategy Tester, continuous run 2018-01-01 → 2026-08-31, 24/7 version)(not published)
Overnight fees (swap) are high
Winning trades are usually held overnight, and every night costs swap. A broker with high swap on gold longs eats into the profit.
Show the numbers
- Estimated from MT5 Vantage 2025: each day a trade is held costs about 0.04R in swap.
Source: trading-research-lab/.ai/evidence/champion-rerun-2026-09-17/atr-and-recent/result.md (swap estimate section)(not published)
What there's NO basis for yet
Auto-pausing the bot when gold turns into a long downtrend
We tried making the bot pause when gold's long-term trend turns down. On data not used for design (2005–2014), this classification predicted no better than chance. So there is no reliable "switch off when…" rule yet.
Show the numbers
- 20-day forward return difference between days labelled 'up' and 'down': every 95% confidence interval contains 0, for all 3 labelling rules tried.
- Re-run on the current EA version (Classic Pullback 24/7, Wilder ATR, Exness 2016–2023, 2026-09-22): the conclusion CHANGED from the old version. A daily trend filter no longer raises the average profit per trade: the best variant adds only +0.02R (within noise), the other two are lower. All it still does is lower drawdown ($2,957 down to $1,374–1,816), at the cost of fewer trades and less total profit ($10,900 down to $5,300–9,300). That is a risk trade-off, not a new edge.
- Same run: the 95% confidence interval of the baseline mean is +0.03 to +0.45R, no longer containing 0. But this is exploratory; the formal test (RL-13, 0.05R threshold, corrected for 4 strategies) still FAILED, and dropping the 10 best trades puts 0 back inside the interval.
Big news (Non-Farm Payrolls, CPI)
Measured on 201 NFP and CPI releases (2016–2023), using the released EA version (24/7, Wilder ATR). Only 31–53 trades per strategy were opened within ±2 hours of a release, every 95% confidence interval contains 0, and shifting the calendar by 1–2 weeks (fake releases) gives numbers in the same range. So there is no basis for switching the bot on or off around news. This is exploratory, not a formal test.
Show the numbers
- Trades opened within ±2 hours of a release (mean R, before swap): Classic Pullback +0.73R (31 trades) · Near-Touch Pullback +0.52R (34) · 3-Bar Pullback +0.50R (35) · Every Signal +0.07R (53) · London Open 7AM none (it enters at 07:00 UTC, releases come at 12:30/13:30 UTC).
- Every Signal did worse around news than otherwise (+0.07R vs +0.24R), but with only 53 trades and a very wide interval (−0.51 to +0.65) — not enough to conclude anything.
- Trades already open at release time look far better (+1.27 to +1.89R), but fake releases score even higher (+1.79 to +3.19R): long-held trades tend to be winners; the news is not the cause.
- A side hint from the old version (holding through real news was worse than every fake date, 5/5 strategies) now holds for only 4/5 — weaker, still just a hypothesis.
- The old version (session filter, EMA-style ATR) reached the same conclusion; it stays in Documents for comparison.
