Every hypothesis we killed
Most research that gets published is the survivors. This page is the rest: every idea we took seriously enough to test, that the data then executed. Each grave lists the claim, what the data said, and the date of death. We consider this the most valuable page on the site — every grave here is a deposit someone doesn't have to lose learning the same thing.
The wins were real (59–60% win rate, +4.4R average winner). The losses were a fiction: 91% of losing trades were single-tick spikes gapping through the stop, filling at the far extreme — average real loss −7.4R against the imagined −1R. Honest fills inverted +2.16R/+2.23R into −0.47R / −0.25R.
Tide is buried with respect. It was structurally reasoned, honestly built — and it died so the ruler could be calibrated: its corpse exposed the fill-model lie, which is now permanently wired into our simulator and into every test on this site. It still runs in shadow to this day, not because we hope it revives, but because a known-dead strategy is a perfect calibration instrument for the scoring machinery.
Spike arrival is memoryless: geometric intervals (CV ≈ 0.99/0.98), flat per-tick hazard (0.93–1.10× the mean regardless of elapsed time), zero interval autocorrelation. The mean is ~502 ticks; the median is ~353; a quarter of gaps exceed 700. There is no safe window. There is only the same coin, every tick.
The −6% move in our 30-day window was 0.57% of the instrument's total motion — within ~2σ of a fair heavy-tailed random walk. Boom was fair to within 0.018%. The grind funds the spike; neither instrument has a harvestable drift.
Faded 2σ bars at 15-minute, 1-hour, and daily scales over a full year: t-statistics +0.10, −0.40, −0.67. Noise at any spread. A point estimate that "clears costs" at |t| ≪ 2 is a coin flip wearing a backtest.
Contemporaneous correlation 0.75–0.78 — they move together. Cross-lag correlations ≈ ±0.02, symmetric in both directions: by the time you can see gold move, silver has already moved. Nothing leads; nothing lags; nothing pays.
Weekly trend-efficiency autocorrelation ≈ 0 at every lag tested. Gating the profitable gold breakout on the prior week's regime collapsed +41R to +6R, and the strictest variant went negative (−9R). Gold's trend state lives on a 3–10 day horizon and is gone by the weekly one.
Every parameter transplanted unchanged, predictions pre-registered. The filter did nothing (+0.063R with or without); the trend-tercile signature inverted (+68% / −14% / +46% of P&L across chop/mid/trend — the wrong shape); and silver's spread-to-range economics (costs 15–25% of risk per trade) meant the geometry could barely express an edge at all. The gold candidate went on notice the day this published.
On the clean-cost instrument, the pre-registered 5-day filter scored −0.143R against −0.027R unfiltered — it selected worse days. The mechanism's tercile shape survived (monotonic, profitable only in trending weeks); the filter's claim to forecast those weeks did not. This is now the precisely-located doubt our live Trial scores separately.
Eight graves. One survivor — currently standing trial under rules it cannot bend. If it joins this page, you'll read its grave written with the same respect as Tide's.