Stop guessing. Start building.

Of 7 entry families we backtested across 4 markets, exactly one beat a coin flip in a majority of its variants everywhere: Order blocks. 81.5% on SPY, 66.7% DIA, 59.3% QQQ, 55.6% IWM. That's ICT's most defensible concept — a real, modest, cross-market signal. It still never beat buy-and-hold. Anywhere. 648 backtests: https://statoasis.com/overfit/research/ict-backtest-what-survives

I backtested ICT / Smart Money Concepts. Mechanically. 648 backtests across 4 markets. Order blocks, fair value gaps, liquidity sweeps, Optimal Trade Entry — versus 3 boring entries, a coin flip, and buy-and-hold. What survived: 🧵

17,424 backtests, and here is what they cannot tell you. Every figure in my Crabel rebuild is frictionless. No commission, no slippage, no spread. The winning settings average a fraction of a per cent per trade. That is where your costs decide it. https://statoasis.com/overfit/research/unveiling-toby-crabel-s-up-down-thrust-trading-patterns

A pivot is a claim about the future, and that is what makes pivot backtests so easy to fake. A pivot high is only the highest bar around once the bars after it have printed. Publish it one bar early and your backtest turns profitable for free, along with every other pivot strategy ever written. So before any result in my Toby Crabel thrust rebuild, here is the control that makes the rest of it worth reading. The study publishes a pivot only after its right-hand bars exist, two bars late in the headline setting and never earlier, and every fill is at the next bar's open. Then it proves the rule by truncation: the signal at every bar is recomputed with the entire future deleted, and it has to come out identical. Any bar whose verdict changed would fail the build. The scale on top of that: Eight markets, two of them chosen specifically because the pattern has no business working there. ES, NQ, SPY, QQQ, DIA and IWM, plus natural gas and Japanese yen futures. 193 instrument-years, with SPY back to 1993 at 8,398 daily bars. 2,024 thrust events, 1,182 up and 842 down. 17,424 backtested variants: 3,024 in the results grid, 14,400 in the parameter sweep. A 50-trade reliability floor, which 936 of the 3,024 grid variants clear. The rest are flagged, never dropped. Frictionless throughout, $35,000 fixed capital, no compounding, one position at a time. Adding the two markets that should not work cost nothing, and they are what turned "this is a pattern" into "this is stock-index exposure". Methodology and full results: https://statoasis.com/overfit/research/unveiling-toby-crabel-s-up-down-thrust-trading-patterns

"Fair value gaps get filled — that's the edge." I measured it. Both halves. Bullish FVG retrace events on SPY since 1993: 1,122. Gaps do get revisited, constantly. That half is true. The average 5-day return after price fills the gap: -0.005% versus the market's ordinary drift. T-stat: -0.08. Statistically zero. The market wanders into the gap, then does exactly what it always does. The fill is real. The tradable meaning attached to it is not — at least not on daily bars, where anyone can check. This is the pattern with most trading folklore: the observation survives scrutiny, the story attached to the observation does not. "Gaps fill" was never the claim that made money. "Gaps fill because institutions leave footprints, and you can trade the footprint" was — and that half did not survive a t-test. Full forward-edge tables for all four ICT concepts: https://statoasis.com/overfit/research/ict-backtest-what-survives

Here is how to tell whether an indicator result is real, using one chart and no statistics. Run the same test on several markets. Then look at the spread. I did this with the OBV MACD. Six markets, and for each one the share of matched backtest pairs where the OBV version beat the plain MACD version it was paired against: QQQ 65% NQ 59% DIA 58% SPY 34% ES 34% IWM 16% QQQ and IWM are both US equity index ETFs. Same country, same exchanges, same macro, same decades. And the verdict almost completely reverses between them. That is what a result looks like when it belongs to a history rather than to an indicator. Test first on QQQ and you publish a confident article. Test first on IWM and you conclude the thing is broken. Both ran a real backtest. Both read noise. The same chart shows the alternative. The confirmed version, where OBV vetoes a MACD crossover rather than replacing it, runs from 45.7% on its worst market to 88.3% on its best and clears the coin flip on five of six. Half the spread, and it never collapses. Nothing in that comparison required a p-value. It required running the test more than once. Three practical versions of the same rule: Run the mirror. If your long filter works, test the short one. A filter that only helps on the side of a thirty-year uptrend has not been shown to filter anything. Run the second market. Not a second date range on the same market. A different instrument. Report the spread, not the best cell. A single best row exists for every claim anyone wants to make. Six markets, 9,216 backtests, and the spread reported for every finding: https://statoasis.com/overfit/research/obv-macd-vs-traditional-macd-which-one-wins