Every strategy we killed — and the bug that killed it.

The entry trigger admitted closes below the band while the stop sat just above it, so the backtest booked instant “wins” at stop prices the market never traded through. On 79% of 4h trades. Tighter stops produced bigger phantom profits, which looked like an “optimal stop” discovery.

The universe held today’s 322 S&P survivors back to 1997, so the engine was quietly buying future index winners before their inclusion dates. The edge was pre-inclusion buying, not stock selection.

Regime inputs used each month’s end value to trade from its start; crash months got retroactively classified as profitable shorts. Plus seven liquidation events at 6.9× leverage, entirely hidden by monthly-close sampling.

The monthly signal was computed at the month’s close and booked against that same month’s return — no lag. With a fast EMA equal to the current price, it was literally “did this month go up?” applied to that month. The celebrated “EMA(1/3) is universal” result was the bug’s signature.

The daily risk block set today’s exposure from today’s closes, then booked today’s close-to-close return on those weights — so a VIX spike cut exposure retroactively on the very day it would have cost money. Not implementable live.

Per-trade returns from a 48-hour hold were annualized at the hourly rate, √(365×24). A pure units artifact — plus a lookahead coin-filter and a 2bp cost fantasy.

Month-end equity sampling hid the intra-month drawdown at 3× leverage — the peak and the crash both happened inside a single month and the monthly series never saw them.

Full-day volume used at the open as an entry filter (data not yet known). And at 100× leverage, fees alone are 10% of margin per round-trip — negative expectancy no position-sizing can fix.

The average loss is the stop distance. The backtest assumed a 0.010% stop always filled at the stop price. It cannot: a stop that tight is inside the spread. Live paper on 2026-06-05 stopped out all five positions intraday. Same signature as the phantom-fill family — the profit was in the fill assumption, not the signal.