Regime luck: +$19,987 on twelve months, −$7,690 on the full history

After direction died twice — at the order book (E-02) and at the indicator level (E-04) — we changed the question. Volatility, unlike direction, demonstrably clusters. So: predict expansion, not direction, and trade breakouts sized for large targets, where the $4.30 toll shrinks to a rounding error. Twelve months of 5-minute data, four out-of-sample walk-forward folds, single-position accounting, real and vol-scaled costs, entries gated by our vol-expansion classifier. Targets: 8 ATR take-profit against 3 ATR stop.

Result: +$19,987 out-of-sample over 754 trades, 44% win rate at ~2.67:1. The first positive number this program ever produced under the honest harness.

We attacked it with the full battery, and it kept standing:

• Direction split: shorts out-earned longs — +$11,608 short vs +$8,379 long — in a period where gold rose relentlessly. Whatever this was, it was not bull-market beta in a costume.
• Concentration: the top five trades were just 8% of gross wins. No single lucky spike carrying the curve.
• Drawdown: max drawdown $3,413 against the ~$20k profit — a return-to-drawdown ratio near 5.9.
• Shape: right-skewed P&L, median trade −$63, mean +$26 — exactly the profile a breakout strategy is supposed to have.

This was, by a wide margin, the most convincing result we ever produced. Which is precisely why the next paragraph exists.

Our operating rule — never trust a positive backtest — had a concrete consequence written into the project before this result appeared: no real money until the candidate survives a mandatory gauntlet, and Gate 1 is more history. Twelve months of 2025–26 gold was one of the most anomalous stretches in the metal's modern record. So we fetched 3.5 years — 241,822 bars, spanning gold's march from $1,805 to $5,596 — and re-ran the identical harness across eight regime-tagged folds: bull, bear, and range.

Aggregate out-of-sample result over 3.5 years: −$7,690.

The same strategy, the same parameters, the same cost model that printed +$19,987 on twelve months lost money on the full history. Only three of eight folds finished positive, all of them small — and all three range-regime folds lost, because a breakout system in a rangebound market is a machine for buying tops and selling bottoms. The twelve-month win had one true author: the 2025–26 gold parabola. The strategy was regime beta wearing a strategy costume.

Here is the uncomfortable epistemology. Every skeptic test in section 02 — direction splits, concentration, drawdown shape — is a within-sample test. When the entire sample is one regime, every one of those tests is asking “is the profit well-distributed across this regime?” and the answer can be a perfectly honest yes. None of them can ask the only question that mattered: is this regime the only one where it works? The sole cure is data from other regimes — which is exactly what the twelve-month backtests plastered across this industry never show you, and why our methodology now treats the regime check as non-negotiable.

One thing in this experiment was real, and the full-history re-run proved it rather than killing it: the volatility-expansion classifier. Gated (−$7,690) beat ungated (−$14,491) — the gate roughly halved the damage, and it did so in every fold, in every regime. The model genuinely knows when large moves are coming. It simply cannot tell you their direction, and a breakout rule is not a good enough translator.

Verdict: killed by the regime check. Predictable volatility is not profit — but it is a real, validated measurement, and it survives today as the live regime probability on our homepage. The strategy died; the instrument lived.