BFG · Variance Lab · Analysis Brief
BFG Variance Lab · analysis brief · prepared 2026-10-01
Full Backstory and Analysis Brief
Everything ChatGPT needs to run a full analysis: the project, the control baseline, all four Raptor variants, every number from the October 1 test session, the gates, the insight, the limits, and what happens next.
Copy for ChatGPT. This page is written as one clean document so it can be copied and pasted straight into a ChatGPT conversation. Plain text, plain tables, no formatting that breaks on paste. Ask ChatGPT for a full analysis of whether P2E is real edge, what to test next, and what would break it.
1. BFG Project Overview
BFG is an independent MetaTrader 5 trading system, rebuilt from scratch: a pattern scanner plus an Expert Advisor. It trades on a demo account only. There is no live trading and no real money involved, and live stays off until paper-trading gates pass.
P2C is the sole control (magic number 20260922). Every variant is measured against it. P2B was retired on 2026-09-30 and is never used as a baseline.
The method: single-variable discipline. Change one thing, run the test, believe the result. P2D changes only the signal flip. P2E changes only the action flip. P2F changes only the stop distance. P2G changes only the TP1 distance. Guessing around breaks things; one variable at a time gives fast, decisive answers.
2. P2C Control Baseline
The strategy
D1 bias traded on H1. D1 close vs 50-SMA sets the daily bias. Entry is a 1-2-3 reversal pullback. ATR stop. 75% of the position closes at 1.5R, the runner goes to 3R. Breakeven plus 0.1R after the partial. 1% risk per trade. Maximum 1 setup per direction, 8 total. Setups expire after 12 trigger bars. No new setups after a 3% down day. Flat at 21:00 server time.
Backtest results, EURUSD H1, every tick real ticks
| Run | Trades | Win rate | Profit factor | Net | Expectancy | Max drawdown |
|---|
| 2026-10-01 rerun | 138 | 33.33% | 0.79 | -16.82% | negative | 26.85% |
| 2026-09-30 baseline | 100 | 31% | 0.78 | -13.82% | -0.139R | 27.59% |
Note: the two runs disagree on trade count and net (138 vs 100 trades). The cause of the discrepancy is unexplained. It is not assigned a cause.
The baseline verdict stands either way: as specified, the P2C strategy loses money. Fix the logic, not the test.
3. The Four Raptors
All four variants were built from the same P2C source and tested on 2026-10-01 on Michael's machine, EURUSD H1, every tick real ticks. Only P2G differs in test window (see below).
Raptor 1, P2D (magic 20260930): flips the signal
Change: the bias signal is inverted. Everything else identical.
| Trades | Win rate | Profit factor | Net | Drawdown |
|---|
| 130 | 35.38% | 0.53 | -29.88% | 34.07% |
Extinct Failed all gates. Hypothesis eliminated: flipping the bias does not fix it.
Raptor 2, P2E (magic 20260931): flips the action
Change: P2C BUY becomes P2E SELL. P2C SELL becomes P2E BUY. No signal becomes no signal. The entry timing is untouched; only the resulting direction reverses.
| Run | Trades | Win rate | Profit factor | Net | Drawdown | Sharpe |
|---|
| First, 2025.01.01 to 2026.09.22 | 78 | 32.05% | 1.15 | +7.35% | 10.71% | 2.92 |
| Extended, 2023.10.01 to 2026.09.22 | 142 | 29.58% | 1.10 | +$922.19 (+9.22%) | 18.87% | n/a |
Extended-run detail: 18,460 bars, 99% history quality, $10,000 starting balance. Gross profit $10,399.87. Gross loss -$9,477.68. Expected payoff +$6.49 per trade. 42 wins, 100 losses. Balance drawdown maximal $2,103.68 (18.87%). Equity drawdown maximal $2,123.04 (19.04%).
The first run passed 4 of 5 gates but failed the mandatory 100-trade minimum with only 78 trades. The history was extended backwards to 2023.10.01 to grow the sample, and the extended run passed all five gates: 5 of 5.
Raptor All 5 gates pass
Raptor 3, P2F (magic 20260932): wider ATR stop
Change: ATR stop multiplier from 0.5x to 1.0x via the new InpStopAtrMult input. Everything else identical.
| Trades | Win rate | Profit factor | Net | Drawdown |
|---|
| 135 | 37.78% | 0.74 | -18.32% | 23.62% |
Extinct Failed all gates. Hypothesis eliminated: stop distance does not fix it.
Raptor 4, P2G (magic 20260933): earlier first take-profit
Change: TP1 from 1.5R to 1.0R via InpTp1RMultiple. Everything else identical.
| Trades | Net | Profit factor | Expected payoff |
|---|
| 138 | -$1,682.06 | 0.79 | -$12.19 |
Results were identical to the P2C control, which strongly indicates the intended TP1 change did not alter the tested behavior. It has not been conclusively proven why. Do not treat "the partial never triggered" as established fact.
Tabled Invalid as a TP1 hypothesis test. Needs diagnosis and repair, not deletion.
4. The Five Gates
All five must hold before any variant advances to frozen-parameter out-of-sample testing and forward/demo testing:
- Net profit greater than 0
- Profit factor greater than 1.00
- Expectancy greater than 0R
- Total R greater than 0
- At least 100 trades
Only P2E has ever passed all five. Never use real money.
5. The Key Insight
P2E flips only the action, not the signal. The working hypothesis: the signal timing may contain information, but the original direction was wrong.
Do not overstate this. One test on one pair and one timeframe cannot establish causation or a durable edge. The 142-trade sample is the minimum viable proof, not a conclusion.
The methodological lesson held up all night: single-variable tests eliminated the bias-flip and stop-distance hypotheses in one run each, while hours were lost guessing around the broken P2G implementation.
6. Limitations, Honest
Not ready for real capital. P2E is a backtest result, not a product. Anyone who says it is ready for investor money today is wrong.
- The return is modest. +9.22% over roughly three years is about 3% a year. A savings account competes with that.
- The win rate is low. 29.58% means roughly 7 of every 10 trades lose. Long losing streaks are guaranteed.
- The drawdown is real. 18.87% maximum drawdown on a 1% risk setting. At higher risk, that scales up.
- Backtest only. No frozen-parameter out-of-sample run, no forward demo, no live paper trading yet.
- One market, one timeframe. EURUSD H1 only. Nothing is known about other pairs, other sessions, or other periods.
- Costs are modeled. Real fills, slippage, and spread widening in live conditions are not proven.
7. Next Steps
- Frozen-parameter out-of-sample test for P2E. New data the variant has never seen, no parameter changes allowed.
- Forward/demo testing. Run it forward on demo for months, not days.
- Diagnose and repair P2G separately. Its TP1 test was invalid; the mechanism needs real technical diagnosis from the source code.
- Build disciplined variations off P2E. One variable at a time, same as tonight.
8. Copy Block: The Numbers in One Place
For pasting straight into ChatGPT:
P2E (Raptor 2) Backtest Results
Period: 2023.10.01 to 2026.09.22
Symbol: EURUSD, Timeframe: H1
Model: Every tick based on real ticks
Bars: 18460, History Quality: 99%
Initial Deposit: $10,000.00
Total Net Profit: $922.19 (+9.22%)
Gross Profit: $10,399.87
Gross Loss: -$9,477.68
Profit Factor: 1.10
Expected Payoff: $6.49 per trade
Total Trades: 142
Total Deals: 366
Profit Trades: 42 (29.58%)
Loss Trades: 100 (70.42%)
Balance Drawdown Maximal: $2,103.68 (18.87%)
Equity Drawdown Maximal: $2,123.04 (19.04%)
Strategy: Flips P2C action (BUY becomes SELL, SELL becomes BUY).
Same entry logic, opposite direction. 1% risk per trade.
Gates: Net>0 PASS, PF>1.00 PASS, Expectancy>0 PASS,
Total R>0 PASS, 100+ trades PASS (142). 5 of 5.
Control for comparison (P2C, 2026-10-01 rerun):
138 trades, 33.33% wins, PF 0.79, net -16.82%, max DD 26.85%.
Dead variants: P2D (Raptor 1) 130 trades, PF 0.53, net -29.88%.
P2F (Raptor 3) 135 trades, PF 0.74, net -18.32%.
Tabled: P2G (Raptor 4) 138 trades, PF 0.79, identical to P2C;
its TP1 change did not alter tested behavior.