The 86% week: the concentration check that killed our best-looking strategy

This was the case for reopening a closed question: every dead strategy had lived on fast timeframes where the cost wall looms enormous. An hourly swing system is a different animal — 3 ATR targets against 1.5 ATR stops, holds up to three days, costs at 2–5% of target instead of 21%. We built a deliberately different feature family from the dead TA set (E-04): sixteen stationary features spanning multi-timeframe trend (EMA gaps and stacking, H4/daily trend computed with no-lookahead resampling), macro momentum, regime measures, and session encodings — on 36 months of clean hourly data, 17,485 bars, zero bad candles.

One detail from the build is worth recording, because it is the honest-labels discipline (E-01) as muscle memory. With asymmetric barriers — 3 ATR up, 1.5 ATR down — a short trade's barrier geometry is not the mirror image of a long's, and reusing a long-only precomputed path mispricies every short. We caught it in review, before the run, and rebuilt execution to walk the full unfiltered price series with direction-correct barriers. Six months earlier, that bug would have shipped — and this article would be about a fake edge instead of a subtle one.

+$8,877 over 173 out-of-sample trades — the first candidate ever to clear our twin bar of net-positive and ≥100 trades. And it came with the credential the gold breakout never had: profit in a bull fold (+$5,999) and in bear folds (+$3,188, +$172). Both directions contributed. Not simple long bias. For about a day, this looked like the one.

Two details didn't sit right. First, the profitable folds were back-to-back in time — together they were 103% of total profit, and they jointly covered a single continuous stretch: gold's late-2025 parabola and its 2026 retracement. One extended episode of extraordinary volatility, split across a fold boundary. Second, in the two calm 2024 folds the model produced zero trades — never confident enough to act. The strategy had only ever really traded one kind of market.

So we logged every individual trade in the profitable folds — all 166 — and asked where the money actually lived. Top five trades: 24.2% of gross wins, four of the five inside a single six-day span. And then the decisive number:

The single week of January 26 – February 1, 2026 produced +$7,889 — 85.9% of the entire two-fold net profit of $9,187.

Remove that one week and $1,298 remains, scattered across 27 other active weeks — earned against a $4,243 maximum drawdown, with a median trade of −$237 (more than half of all trades lose). Statistically indistinguishable from zero. The “edge” was one correctly-timed short into one violent macro week, surrounded by noise that slowly gave the winnings back.

We then ran one more test, designed to help the strategy rather than kill it: adding exogenous macro features — dollar-index returns and 10-year real-yield changes (merged with a one-day shift so no future data could leak). The logic: if the swing pattern were real, causally-relevant information should reinforce it.

It did the opposite. The lucky folds deflated toward zero (+$5,999 → +$2,799; +$3,188 → +$250; +$172 → −$424), a previously-silent fold produced a new loss (−$2,831), and the aggregate landed at −$99.96 over 219 trades. Real information erased the “pattern” instead of sharpening it — which is exactly what happens when the pattern is noise the model had memorized capacity to fit. Two independent methods, one verdict.

Killed by the concentration check — the question aggregate P&L is built to hide: not how much, but from where. An n-of-1 windfall cannot validate a strategy no matter how it's split, sliced, or dressed in a 173-trade sample. Every result on this site now answers the where-question before we believe the how-much — that rule was purchased with this experiment, and it is the single most transferable lesson in this journal for anyone evaluating anybody's track record: find the best week, subtract it, and see what's left.