Patterns are real; edges aren't

The setups taught on Turkish trading channels (fair value gaps, liquidity sweeps, order blocks, market-structure shifts, gap fills) were coded from the teachers' own definitions and tested on real prices with costs, random-direction controls, multiple-testing correction and out-of-sample periods. Across 3,552 + 1,536 + 3,828 variants, none survived. The patterns themselves are real: they also appear on shuffled charts and on random walks.

None of this says the methods "don't work". It says that, defined the way they are taught and tested this way, they did not separate from chance after costs. What a teacher does at their own discretion was not tested.

Seeing patterns

Take ninety hours of a real gold chart and shuffle the order of the candles. Each candle keeps its shape; only the order changes, so no pattern that depends on order can carry information. The shuffled chart still shows a resistance line with four touches, a support with three, and a bullish fair value gap, all found by code, not drawn by hand.

Counted on 15-minute gold candles from 2026 (17,314 bars), there were 204.5 fair value gaps per 1,000 bars on the real chart, 241.2 on the same candles shuffled, and 314.3 on a random walk of the same length. The patterns form where no "smart money" exists. That the real chart has fewer means the market is not a pure random walk; it is not evidence that the patterns predict anything.

The definitions are loose enough to matter. On one real chart, the minimum gap sizes the sources allow find 3,540, 1,834 or 875 fair value gaps: a factor of four, from the same teaching.

Seeing a pattern is not the same as the pattern working. That needs a test.

The test

Five ICT/SMC rule families were coded into 3,552 variants and run on one-minute gold, EUR/USD and Nasdaq prices from 2019 to 2025. In-sample (January 2019 to March 2023), after the spread, none of the 3,552 passed a multiple-testing correction: zero under Bonferroni, zero under Benjamini–Hochberg, a White Reality Check p-value of 0.961 and a Hansen SPA p-value of 0.281.

To be fair: before costs there is an in-sample result that data mining does not easily explain (Reality Check p = 0.010). It did not hold. The best in-sample cell, net +0.099R per trade (t = 3.61), was −0.038R per trade (t = −1.16) out of sample, over 2,749 trades.

A control that took the same entries but chose direction by coin flip could not be told apart from the rules' own direction (Reality Check p = 0.256). The rules' choice of direction added nothing measurable.

A high win rate did not rescue them either. Only 34.3% of cells won more often than the cost-inclusive break-even rate.

A whole curriculum

One trading "bootcamp" series teaches a full curriculum. Its untested families were coded into 1,536 more cells, with 1,925,300 in-sample trades. Zero survived the correction, before costs as well as after. Counting the three families of that curriculum already covered above, 2,640 curriculum variants were tested; zero survived.

A further 3,828 cells built from rules taught on other channels (opening-range breakout, inverse fair value gap validity, premium/discount filters, candle-range theory) also had zero survivors after costs. The best in-sample cell was −0.042R per trade out of sample.

Do fair value gaps "hold"?

A gap "holds" if price touches its near edge and then moves away before moving through. On every timeframe from one minute to four hours, the share that held was below one half: 0.4655 at one minute (1,497,187 gaps) up to 0.4977 at four hours. A random level drawn at the same distance held 0.4756 to 0.4902 of the time. The difference stayed within about one percentage point. The share rises with timeframe, but it rises the same way for the random level.

"Gaps fill 60–70% of the time"

S&P 500 opening gaps touched the previous close in the same session 58.4% of the time (3,164 gaps). A random walk with no drift and each day's own volatility fills 56.1%. The fill rate is mostly a base rate. It also depends on the definition: 270 reasonable definitions give anything from 0% to 96.5%. None of 810 gap-fade definitions survived multiple-testing correction. The excess over the random walk is real but small, and costs absorb it.

Order flow

Order-flow imbalance explains price moves in the same moment. Its information about the future lasts seconds: on Bitcoin futures (a proxy market, 92 days), the correlation with the next 5 seconds was +0.041 and with the next 60 seconds +0.002.

What the literature says

The result is not unusual. On over a century of Dow Jones data, one-way costs of 5–35 basis points erase the out-of-sample performance of simple technical rules (Bajgrowicz & Scaillet, 2012). Of 95 modern studies of technical analysis, 56 were positive, 20 negative and 19 mixed, and most of the positive ones used methods open to data mining and cost problems (Park & Irwin, 2007). When many ideas have been tried, a t-statistic above 3 rather than 2 has been proposed as the bar for a new finding (Harvey, Liu & Zhu, 2016).

Method: one-minute prices 2019–2025 (gold, EUR/USD, Nasdaq, S&P 500), Bitcoin futures order flow, deterministic simulations with fixed seeds; rule definitions written from the teaching material. Code and outputs are in a private research repository and available on request.