Expert Advisor on gold based on return-to-origin logic: whoever chases the impulse late pays whoever originated and defends it. 590 backtest trades interrogated one by one. The discovery: half the system (the longs) contributed 10% of the profit while containing the 5 worst trades in its history. Cutting longs and two dead hours: −47% drawdown without touching a single entry. Open problems declared: absolute-point thresholds on gold that tripled, regime filter in the development queue.
The full case study →+458.88% gain, 72.31% max drawdown. I publish it whole because it's the most important lesson on this site: gain% without the drawdown next to it means nothing. A +458% earned through a −72% is a well-flipped coin — and the 72% was dug by the hand, not the signal. My whole architecture of automatic constraints was born there.
Verify on Myfxbook ↗Two prices for the same asset: the fast Binance futures websocket feed and the slow CFD broker quote. When they diverge, it's not a prediction: it's a fact with an expiry. Versions 4 and 5 failed (−$127.02, 93% of it commissions) and the autopsy is public. v6 is cost-gated: it only enters if the divergence exceeds 2× live-measured costs, exits at convergence, and shuts itself down if realized EV turns negative. In shadow mode until it earns its promotion.
Architecture and lessons →The most valuable quant research I own: the one that failed. Hypothesis: price action patterns have tradable predictive power on 5-minute Bitcoin. Four independent tests, leak-free features, rigorous in-sample/out-of-sample. Verdict: the best gross edge is worth 1.3 bps against 6.5 bps of costs — a 1:30 ratio. Published in full, because every reliable "no" is an account not burned.
The complete research →+641.24%, 81% win rate, nine thousand trades. Looks like the Grail; it's a time bomb I keep public on purpose. Average loss 5× the average win, negative pip expectancy, Z-score −6.12: the statistical signature of grids/martingales. Learn to dismantle my showcase numbers and you're vaccinated forever against everyone else's.
Dismantle it with me in 5 minutes →The AI analysis engine alongside my lab: macro monitoring and structured signals built with the same philosophy as everything you see here — data before opinions, costs inside every calculation, and no promise the numbers can't keep.
macro-sentinel.com ↗The research pipeline is itself a portfolio: one system in supervised production, one in shadow mode with a written promotion criterion, one active high-potential research direction (perpetual funding/basis carry), one structural upgrade queued (ATR-relative thresholds and a regime filter for Kimera) and a documented graveyard — so no future version of me retests dead ideas from scratch.
Whoever chases one system at a time, falling in love with it, is making serial all-ins on their own research.
Whoever survives ten years has a funnel, a tidy graveyard and at least one thing in shadow mode. Always.