Claude Tested Over 25,000 ICT Strategies (Here's What Works)

ICT Silver Bullet: A 25,000-Configuration Backtest

Brendan, a UCLA math/econ graduate and former Raymond James investment banker, used Claude to run over 25,000 backtests of the ICT Silver Bullet day-trading strategy across 16 years of 1-minute NQ and ES futures data, simulating more than 10 million trades. The goal was to turn the strategy’s famously discretionary rules into explicit, testable configurations and determine which parts of the ICT framework actually hold up.

Methodology and Scope 🧪

• The ICT Silver Bullet setup combines six elements: a liquidity level, a sweep of that level, displacement, a fair value gap (FVG), a retracement entry into the gap, and a target at opposite liquidity.
• The strategy is only traded in three hourly windows: 3–4 a.m. London, 10–11 a.m. NY, and 2–3 p.m. NY.
• Because ICT rules are discretionary, every judgment call was converted into a “dial”: liquidity definitions, gap size, entry location, stop placement, bias, and trading window.
• Multiplying these dials produced 258 million possible strategy versions; the study ran a uniform random sample of 25,000 plus named “textbook” and “as-traded” versions.
• The testing stack included real commissions, slippage on stop-outs, and same-minute stop/target collisions counted as losses. Results were compared against coin flips, corrected for luck across all tests, and survivors were run cold on ES. The final two years of data were locked away until the very end.

Headline Results 📊

• Only 9,481 configurations produced at least 100 trades; 6,190 were profitable after costs; 2,191 cleared a risk-adjusted bar; 1,614 survived luck correction; and 420 remained profitable on ES.
• That means only 1.7% of all ICT Silver Bullet versions survived the full validation stack.
• The 5-minute textbook version made a little money but failed at the risk-adjusted stage. The 1-minute textbook version reached the final ES check and then failed. The loose “as-traded” version, which most ICT traders actually use, survived every stage.

Do the Famous Claims Hold Up? 🤔

• Win rate: The quoted 70–80% win rate was not found. Top versions ranged from 34–48%; surviving configurations centered around 52%. Only five configurations hit 70%, all by taking quick 1R profits roughly 20 times per year.
• Special hours: Running the same logic on every hour of the session showed the Silver Bullet windows were not special. The 10 a.m. window was slightly negative; the London and 2 p.m. windows did fine but sat inside the pack.
• Rule-by-rule scoring: Starting with only FVG entries scored 33. Adding the liquidity sweep jumped it to 69, making it the only rule that added real edge. Displacement filters, restricted hours, and especially the 15-minute bias hurt performance—adding the bias dropped the score to 0.16. The best versions used no bias or only a simple daily one.
• Trading frequency: The 5-minute textbook version fires about 34 times/year; the 1-minute version about 133 times/year; the as-traded version about 1,996 times/year (~68/day).
• Income reality: Textbook versions lost more days, weeks, and months than they won. The as-traded version had 62–68% positive months, but the best configuration once had a 19-day losing streak.

Survivors and Holdout Performance 💰

• 420 of 1,614 luck-corrected survivors stayed profitable on ES with no retuning.
• On the locked-away two-year holdout, the best survivor made $455,000 on a single contract with a Sharpe of 13 and all 24 months positive.
• The loose as-traded version made $251,000, and textbook versions stayed slightly positive.
• Caveats: the holdout period was a strong rising market, making it a ceiling rather than an expectation, and results assume perfect paper fills—in reality traders keep losers and miss winners.

What Actually Works ✅

The surviving version of ICT Silver Bullet is the simplest one:

• Require a liquidity sweep.
• Fade the side that was just swept.
• Enter on the 1-minute FVG left by the reversal.
• Trade often, because edge per trade is small.
• Drop the 15-minute bias.
• Ignore the special hour windows.

As the video notes, this is essentially mean reversion: a sweep and snapback on a 1-minute chart, repackaged with ICT vocabulary.

Replicating the Test 🔧

The system is built in layers: (1) data pulling/cleaning, (2) pattern detectors for gaps/sweeps/structure shifts, (3) a trade engine with real costs, (4) a 25,000-config sweep with validations, and (5) a dashboard. The prompts can be screenshotted and used in Claude.ai or Claude Code. The broader point is that any discretionary strategy can be turned into explicit rules and tested this way.

Key Takeaway

The ICT Silver Bullet’s edge is real but narrow. The liquidity sweep matters; the elaborate rulebook around it mostly detracts. The framework survives only in its stripped-down form: one signal, simple entry, basic risk management, and high frequency.