Before a video goes through the machine, years of our own research have usually been here. Every family below was tested with honest costs, random-entry twins and the wrong-place placebo, across many markets, many times. This is what it said. Claims are tested, never people.
families
17
lab studies
55
survived the screen
2 of 92
One page per idea. The count on the right is this site's library; the studies are our own lab's. Survivors are counted, never named while they are being traded.
The things that stayed true across hundreds of tests. Each one killed more than one family.
The toll (spread, commission, slippage, financing) is the biggest lever we have ever measured: lowering it from 2 to 0.55 basis points changed a strategy's result by more than any rule change. An edge that only exists below real costs is not an edge, and most intraday videos sit in that gap: positive gross, negative net.
Every level, box, gap and block is tested against the same shape shifted by a random amount. If the wrong one performs like the right one, the picture was hindsight. ICT zones, order blocks, fair-value gaps, premium/discount ranges and volume-profile levels all failed this test.
Crosses, flips and zones are found by looking back from a move. The ones in the middle of nothing are as common and never shown. Backtesting counts both.
A rule that works on one market and fails on the next four is a coincidence with a chart. Every survivor must transfer; the opening-range weekday effect replicated on the Nasdaq and failed everywhere else.
The moves that run do not come back to be bought at a ratio. A pullback entry systematically selects the breakouts that failed. 60 of 60 cells lost.
Wicks, engulfing pairs and confirmation candles: 0 for 131,822 cells. A bigger wick means a bigger next move in both directions. Volatility, not a forecast.
Many effects are statistically real and smaller than the spread. A tilt of half a basis point with a t-stat of 20 is a fact about markets and a losing trade. We report both numbers so you can see which one you have.
Stacking filters cannot beat the square root of their summed strength, and every added filter costs a multiple-testing haircut. 888 combinations of over-extension detectors could not clear the toll together any more than alone.
Risking 80 points to make 40 wins most of the time and loses money. Win rate says nothing without the size of wins and losses; the scoreboard on every verdict shows both.
A stop placed at the moving average or the zone edge is inside the market's normal wobble. The 86%-win-rate scalp with its stop at the 34 EMA measured 20%: the stop was hit by noise before the idea had a chance.
Sizing up after wins, down after losses, or by volatility: every variant we tested did worse than the same risk on every trade. The risk dial changes the speed of the outcome, not its sign.
Removing losers from a list of past trades always looks good. A real filter changes which trades happen next and what they cost; only a fresh simulation shows that.
Resting orders fill through the level, with slippage, or not at all. A breakout engine that arms a stop at the high and fills it there overstates the result; ours fills at the worse of the level and the next open.
The first bar's open, high and low hide a whole auction. Trades placed 'at the open' only count when filled live or on tick data; on bars they pick up prices that never traded.
Cross-market lead-lag: 0 for 8, on the source side and the target side, including options-implied inputs. When a second market disagrees with price, it is late, not early.
Expiry pins, settlements, positioning reports, auction results, holiday patterns: priced within minutes of release or before. Public information carries no per-trade edge.
Seventy wins out of a hundred at a true 46% rate happens about once in ten thousand tries. When a video's sample sits that far from the measured rate, the sample was picked. Every verdict runs that arithmetic.
Higher-timeframe values are used only after their bar closes; swing points only once confirmed. Many indicator scripts repaint, and the videos made with them show trades that could not have been taken.
Thirty-six variants need a t-stat above 3.2, not 2. Selection on the first 70% of history, confirmation on the last 30%, or it is curve-fitting with a nicer name.
An idea that made money in 2020-2022 and nothing since is not in a drawdown; it belonged to a regime. The year-by-year section is on every report for this reason.
A verdict is only worth what its engine is worth. These are the errors we found in ours, dated, with what changed.
A maintenance script that re-reads failed runs treated a published video's dead duplicate rows as failures and replaced the good report with a weaker read.
what changedRestored the same day from the script's own backup (same verdict, same 1,098 trades). The script can no longer touch a video that has a finished report.
A session window meant for intraday bars was applied to daily bars; no daily bar ever 'closed inside' it.
what changedSession windows are ignored at the daily timeframe. Two library reports changed.
Strategies taught as 'first this, then wait for that' were being written as a single condition on one bar. Two events almost never share a bar, so 11 of 23 test videos got zero trades and a thin-sample verdict instead of an answer.
what changedEvery extracted entry is now counted on the real data before the report is written; a setup that almost never fires is sent back to the reader with the measurement. Behavioural fidelity on our test set rose from 27.6 to 40.8.
Several broker 1-minute exports carried timestamps shifted by an hour in the weeks when US and European daylight-saving clocks disagree, and flat bars where the market was moving.
what changedEvery feed is now stamped week by week against exchange-timed data before any strategy sees it; failed weeks are corrected or excluded. It is on the Our data page.
A strategy validated over months at +4.4 bp per trade turned out to rest on a backtest engine that walked resting limit orders through the bar and found deeper fills than the market would have given.
what changedThe engine now fills resting orders at the worse of the level and the next open. The strategy was retired; its live record was losing, as the corrected backtest said.
The AI reader is not deterministic: the same video produced 8,153 trades one day and 5,726 the next, with the same verdict.
what changedOne canonical report per video; a second submission opens the one on file. Admins can re-read a video in place, and the report says which model read it.
Frame sampling took the first N scene changes, so a 44-minute video's charts were sampled up to 03:02 and the strategy taught at minute 30 was invisible to the reader.
what changedFrames are sampled across the whole runtime, every read frame is on the report's reading board, and the transcript is synced to it.
Found a video that claims one of these works? Run the autopsy and add it to the record.