MT5 V5 — XAU-centric prop-firm trend books

Two research regimes. Since 2026-09-04 this project optimises net Sharpe with drawdown unconstrained (regime U), not prop-firm pass rate (regime P). Which findings belong to which regime — and where to pick up if we go back to a prop firm — is in RESEARCH-REGIMES.md.

Systematic trend-following books built to pass (and hold) prop-firm challenges — FTMO, FundingPips, Maven — trading gold plus a few uncorrelated diversifiers via MT5 on a headless Linux/Wine VPS. The question that drives everything: what book passes a 2-step challenge at the highest probability and fastest median time, net of real broker costs?

Two documents matter most. V5_FINDINGS.md is the settled research ledger — read its MANDATORY CONTROLS before trusting any number in this repo. BlogPOST.md is the plain-English write-up of what was found and what it cost to find out.

The old EURUSD/USDJPY candle-champion path (v5_validate_champion.py, candle features) is superseded. The project pivoted to XAU trend books.

---

Current state — 2026-09-03

Nothing is trading right now. Every trading timer on the VPS is disabled. This is a deliberate, known state, not a fault:

Account	Bridge	Book	Status
FTMO-Demo 100K (1514222432)	18814	XAU+BTC+NDX+BRENT @9%	stopped 2026-08-28 by decision. All 8 positions closed for +$3,215 realised, final balance $102,944.17 (+2.9% into a +10% P1). Timers disabled.
FundingPips 10K (20414926)	18812	XAU(micro)+ETH+DJI30	dead — broker returns "Invalid account" on login. Its bots had already been failing before this; its terminal prefix was repurposed for Maven.
HFM cent live (54939391)	18813	dual ls + champ	drained to $0.12 USC (~$0.001). Timer paused. Was a plumbing-validation account, never a returns account.
Maven 100K (10398574, BNPL P1)	18812	champion (long-only)	built, logged in, verified — BLOCKED. Orders rejected retcode 10026 = autotrading disabled by server. Needs enabling in Maven's dashboard; nothing else is outstanding.

Maven, measured live: spread $0.44 (12 tick samples, stable — FTMO-class, not raw/ECN), leverage 1:75, swap −$45.33 long / +$25.25 short per lot per night, stops_level 0. Executor scripts/v5_maven_champ.py (magic 360571, config configs/v5_xau_maven.json, risk_frac 0.5%) with a once-per-day guard so a manual profit-take is never re-opened by the next hourly reconcile.

Known infrastructure faults

• Two of three MT5 terminals cannot serve the Python API. After the 2026-08-28 reboot (kernel 6.8.0-124 → 6.8.0-138), the two build 5836 instances (.mt5b cent, .mt5c FTMO) never reach "MCP started" and every initialize() returns -10005 IPC timeout. The build 6140 instance (.mt5, now Maven) works. Terminal GUIs are fine; only the API channel is affected. Reviving them needs a build upgrade, not a restart.
• mt5-terminal.service's ExecStop uses a bare pkill -f terminal64.exe + wineserver -k, which would kill all instances. Never systemctl stop it — kill by verified /proc/<pid>/environ WINEPREFIX instead.
• Maven's ini auto-login does not survive a terminal restart (the broker's server discovery is slower than the ini path's patience). A restart currently needs a one-time VNC login. Do not restart that terminal casually.

---

The breakthrough: diversification, and only diversification

The champion (src/v5/xau_dual_signals.py::champion_signal) is a long-only blend of EWMAC (16/32/64 vs 64/128/256 days) and 10/20/40-day breakout, concentrated at conc^1.5 with a small resting long tilt, run through a causal vol-target. On XAU H4, net of a live $0.448 spread, it does Sharpe ≈ 1.02–1.10, maxDD ≈ −11% at 1% risk/trade.

Two levers produced that, and nothing else has since:

1. Killing the shorts. Re-confirmed independently in 2026-09 at a completely different horizon: across ten signal families, the short side lost money on all ten, and long-only roughly doubled every long/short Sharpe (e.g. EWMAC 12–60h: both +0.34, long-only +0.86, short-only −0.58).

2. Uncorrelated sleeves. Re-measured cleanly with per-instrument basis-point costs:

	Sharpe	maxDD
Gold alone	+1.098	−19.5%
4-sleeve book (XAU+BTC+NDX+BRENT)	+1.480	−8.0%

Higher Sharpe than any component and less than half the drawdown of the best one. NDX standalone is a mediocre +0.52 and still earns its place — it contributes decorrelation, not return.

Two findings that changed how we size

risk_frac is a pure risk dial with a violent asymmetry. Sharpe is invariant at +1.02 across 0.25%→4%; return scales linearly, breach probability does not:

risk/trade	Sharpe	CAGR	maxDD	P(−10% in 12mo, trailing)
0.50%	+1.015	+4.4%	−5.8%	0.0%
1.00%	+1.024	+9.0%	−11.2%	1.9%
2.00%	+1.024	+17.8%	−21.5%	45.9%

1%→2% doubles CAGR and multiplies 10%-drawdown risk by ~24×. Use 0.5% under a 10% wall; never exceed 1.5%. Also: static drawdown (from starting balance) is far more forgiving than trailing — 4.4% vs 17.9% breach at 1.5% risk. Sizing off the forgiving one by accident is how books get too big.

Sharpe does not predict prop-firm pass rate — the dial-scaled left tail does. The two deployed books are statistically equivalent (matched-vol paired t = −0.40) and 0.748 correlated, with ETH≈BTC and DJI30≈NDX as near-substitutes. What separates pass rates is whether the vol dial pushes the worst day through the daily-loss limit:

book	dial	worst scaled day ×1.5	vs 5% daily	fail_day	pass
XAU+BTC+NDX	7%	−3.96%	ok	0.0%	98.7%
XAU+BTC+NDX	9%	−5.09%	BREACH	13.9%	83.3%
+BRENT	9%	−4.80%	ok	0.0%	96.8%

BRENT is tail insurance, not return (on returns it is insignificant, t=+0.47; it takes daily breaches 13.9%→0.0%). This independently reconstructed the deployed design decision.

---

What did NOT work

The negative results are the bulk of the value here. Everything below was built, measured honestly, and rejected — each is recorded in V5_FINDINGS.md with a "do not re-run" marker.

Five consecutive overlay families on the champion, ~40+ configurations, all failed:

attempt	outcome
ADX-percentile regime gate	Sharpe 1.084→1.007, drawdown worse, paired t −2.88, pass 95%→80%
scale-in / averaging down (54 variants)	leverage in disguise — plain champion at higher risk_frac beat it on Sharpe, drawdown, Calmar and worst trade
five untried lifts (range-vol denominator, ensemble agreement, trend age, parameter cloud, skew)	0/29 passed; the one that looked real was a post-2022 artifact and the live bot's ATR beat it anyway
6-component stack (vol regime, trend strength, HTF confirm, hours, cost filter)	0/29 passed; 2 of 6 were already implemented, 1 was collinear with the champion at +0.85, 1 a category error, 1 had no bad session to remove
high-vol taper	best single result found (gold 1.098→1.311, both halves) — then failed pre-registered replication: 1 of 4 sleeves, and it hurt the portfolio (1.480→1.310)

Other closed lines: 12 ICT/SMC concepts + 4 confluence combos (zero survivors); turning-point bottom-detector as a scalp (all 48 cells negative); 4–15h long/short book (shorts lose on 10/10 signals, cost wall); intraday/multi-timeframe (H4 beats M30/H1 even at $0.12 raw spread — the intraday signal is genuinely worse, not merely cost-limited); LLM-as-analyst (no edge); martingale (ruins 71–97% of runs); meta-labeling (failed twice).

Maven's 0% daily-loss mandate is infeasible. Backtested honestly: with tight stops 41.7% of days realise a loss; loosening stops trades frequency for severity (worst day −5.0%). No parameterisation gets near "never lose on a day" — that constraint is incompatible with taking directional risk at all. The account is therefore run with manual exit discipline plus a 0.5% risk dial, not an automated zero-loss guarantee.

A commercial EA ("Happy Gold – OxS") was reverse-engineered and proved un-replicable. The log is genuine (balance reconciles to the cent). Its direction rule was fully recovered (momentum, t≈30) but carries no standalone edge (+$0.22 gross vs a $4.40 spread). The edge is in entry timing, which is provably absent from 15-minute bars (every feature t≤2.6 vs matched controls) and was not reproduced by tick-level burst triggers on 2.95M real ticks. Four candidate mechanisms eliminated; see strategy/REVERSE_ENGINEERING_REPORT.md.

---

Methodology — the actual product

Forty-plus failures are only informative if the measurement is trustworthy. V5_FINDINGS.md holds the full MANDATORY CONTROLS list; the ones that changed answers:

• Floor every cost at live broker rates. The CSV spread column understates the venue by up to 10×. Measure the real spread by sampling ticks.
• Benchmark against the incumbent, not zero. Most overlays are positive-Sharpe and still worthless.
• Gate de-risking overlays at MATCHED VOL. Anything that multiplies exposure by ≤1 will show a negative raw paired-return test even when Sharpe and drawdown clearly improve. One variant read Sharpe 1.155 vs 1.084 and DD −13.4% vs −19.8% with a raw t of −2.52.
• Split-sample everything, including things that pass. A pooled t of +2.25 with 7-of-9 years positive still turned out to be a post-2022 artifact.
• DSR is near-vacuous for overlays. When every candidate is a variation of an already-good base they all inherit its edge; we logged DSR 0.998 alongside PBO 0.845. For overlays the gate is the paired test against the base.
• Probe for lookahead mechanically (src/evaluation/lookahead_probe.py) — truncation invariance on synthetic data, run before trusting a backtest.
• Pre-register the decisive test. The taper replication would have produced a positive story if a per-asset schedule had been cherry-picked after the fact.
• Watch the units. Dollar constants don't transfer across instruments ($0.10 slippage is 0.002bp on BTC and 12bp on Brent); bar-count lookbacks don't transfer across timeframes (D=6 silently makes the champion a 96–1536 day system on daily bars). Both produced spectacular nonsense before being caught.

---

Layout

Engines — scripts/v5_basket_challenge.py (champion recipe + portfolio vol-target), scripts/v5_basket_challenge_exec.py (generic guard-first executor, --config <firm>.json), src/v5/xau_trend.py (discrete trade-ticket engine) via scripts/v5_xau_dual.py / scripts/v5_maven_champ.py.

Configs — v5_ftmo_challenge.json (4-asset @9%), v5_fp_flex_10k.json (10K book), v5_xau_maven.json (Maven, 0.5% risk), v5_xau_dual.json (cent dual), v5_fundingpips_funded.json (planned weekend-flat @6%), v5_xau_champ_fundingpips.json (planned champ-only half-risk).

Research — V5_FINDINGS.md (ledger), BlogPOST.md (narrative), strategy/REVERSE_ENGINEERING_REPORT.md (EA forensics), FUNDED-STAGE-PLAN.md, SPRINT-OPTIONS.md, CHALLENGEBOT.MD. Session memory in ~/.claude/projects/.../memory/ (MEMORY.md index) is the fast-recall layer for what is deployed, disproven, and why.

---

Immediate next actions

1. Enable algorithmic trading for the Maven account in Maven's dashboard — the only thing blocking a live champion there. Everything else is built and dry-run verified.
2. Decide the FundingPips 10K account's fate — the login is invalid; it needs renewing or writing off.
3. Fund or retire the HFM cent account — at $0.001 it cannot size a single trade.
4. Upgrade the two build-5836 terminals if their accounts are ever needed again; a restart alone will not fix the IPC fault.
5. If a Friday-flatten is wanted for FundingPips funded stage, that executor feature is still unbuilt (fundingpips-b-account-plan in memory).

Ops gotchas

• np.float64 must never cross the rpyc bridge — it ships to the Wine interpreter as literal np.float64(...) and is eval'd without numpy → NameError. Cast to plain float at the boundary (bit the cent bot's trailing stop and sizing twice).
• Vendor trade logs may not share your feed's clock. The reverse-engineered EA log ran +2h ahead; before correcting it, 43% of its trades looked "physically impossible." Always solve the offset by maximising in-bar fill consistency first.
• Broker tick retention is ~1 month — verified: ticks exist at 1 day/1 week/1 month and are zero at 3, 6, 9 and 11 months. Any tick-level study must be planned prospectively.
• Local dev MT5 on Wayland can die at launch (nodrv_CreateWindow); fix is wineserver -k then relaunch with the mutter Xwayland cookie, not ~/.Xauthority.
• Any systemd unit launching Wine needs KillMode=none, or teardown kills the shared wineserver and takes the persistent terminal with it.
• Prop firms can restrict a symbol mid-challenge without notice (FundingPips set XAUUSDmicro CLOSE-ONLY once). The executor checks trade_mode before sending.
• This repo can have concurrent Claude sessions committing and switching branches; check git reflog before assuming a missing file is data loss.