Does the ICT reversal model actually have an edge?

Smart Money Concepts and ICT are everywhere in retail trading. The "reversal model" — sweep a major level, wait for displacement, a market structure shift, and a fair value gap, then enter at the gap — is one of its most-taught setups. We wanted to know something simple: tested mechanically, over a long history, does it actually make money?

So we coded it as an explicit, rule-based strategy and ran it three ways: against a 16-year NinjaTrader 8 backtest log, against a year of 1-minute NQ bars with a full parameter search, and against 800 years of synthetic markets calibrated to real NQ volatility. Here's what we found.

1. The strategy we tested

Only the reversal model was coded. Every discretionary term was turned into an explicit, adjustable rule so the test is reproducible rather than a matter of interpretation:

• External levels: high/low of day (from midnight NY) and previous day high/low. A level must be at least 10 bars old to count.
• Sweep: a bar trades through the level by at least 1 tick and closes back inside. Each level can be swept once.
• Displacement: bar range ≥ 1.5 × ATR(14), body ≥ 60% of range, opposing wick ≤ 25%.
• Market structure shift (MSS): a close beyond the nearest pivot before the sweep extreme, within 15 bars — and the MSS bar must itself be a displacement bar.
• Fair value gap (FVG): a three-candle gap of at least 4 ticks within 3 bars of the MSS.
• Entry / stop / target: limit at the near FVG edge; stop at the sweep extreme + 2 ticks; fixed 3R target; session 09:30–14:00 NY, flat by 15:55.

The continuation model and the ES confluence filter were left out of this test. The strategy was written for NinjaTrader 8 and ported to Python for the grid, Monte Carlo and regime work.

2. The data

One honesty note up front: the bar file has no overnight data, and the NT8 database prices differ in scale from the bar file, so the Python port could not be reconciled trade-for-trade against the log. R-multiples are unaffected by that scale difference, but point-based risk filters are. This, and the other caveats, are why we call this exploratory.

3. The full model over 16 years

Of 7,781 setups, only 485 became trades — 91% died before an order was placed. The ones that made it through:

The payoff ratio was actually healthy — average win +2.62R against average loss −0.93R, a 2.83 ratio. The problem was the win rate: 19.6% against a break-even of 26.1%. The target was hit on just 15.7% of trades (a binomial test against 25% gives p ≈ 4×10−7). The implied Kelly fraction is −8.8% — negative, meaning the optimal bet size is zero.

No slice rescued it. By level: previous-day-high was least bad (−0.04R), previous-day-low worst (−0.27R). Shorts (−0.09R) beat longs (−0.23R). Only 4 of 17 calendar years were positive, none by more than 0.3R per trade. Where setups died: 64% never got an MSS in time, 21% broke without displacement, 3% lacked an FVG.

4. Parameter search & overfitting

We ran the full sequence across 1,728 parameter combinations — swing strength, displacement threshold, sweep-to-MSS window, FVG rules, maximum risk, and more. The trade-weighted average of every run was −0.33R. Only 3.9% of runs with 10+ trades were positive, and the single best was +0.13R on just 15 trades.

The tell-tale overfitting signal: the top 20 runs on the first six months averaged +0.50R there, then −0.43R on the last six months. Settings that looked great in-sample had no predictive power out-of-sample.

Figure 1 — upload fig1-parameter-grid.png and set its URL here

5. Which component actually adds the edge?

We rebuilt the model one filter at a time to see where the signal was — and where it drained away:

This is the heart of it. A bare sweep of an external level was positive (+0.22R) and beat both a random-entry control and a control using sweeps of internal levels — so the external level does seem to matter. But each ICT filter you stack on top removes trades and adds nothing. Adding MSS and FVG stripped out 97% of the trades and pushed the result negative.

Even the sweep-only result isn't something to trade yet: it's +0.37R in the first half of the year and +0.06R in the second, and at a 1R target it's essentially zero — the edge comes entirely from a few winners running to 2–3R. With realistic costs it falls to about +0.10–0.15R.

6. Monte Carlo & market regimes

Resampling the strategy's own trades 10,000 times, a year of +120R never once appeared for the full model — at ~29 trades a year, even a 100% win rate caps out near 87R. The full model finished positive in only 1.6% of 16-year simulations.

Figure 2 — upload fig2-equity-fans.png and set its URL here

Then we generated eight regimes of synthetic NQ-like prices — random walk, bull and bear trends, range/mean-reverting, high-vol chop, crash, momentum breakout, low-vol grind — and ran 100 simulated years of each. The sweep-only rule only made money in one regime:

Figure 3 — upload fig3-regime-fans.png and set its URL here

The takeaway: a tight-stop, 3R-target rule bleeds about 0.10R per trade on a pure random walk, and only turns a profit when the market reverts to a level. Synthetic markets have no "real" ICT structure, so this shows what conditions the rule needs — not whether real markets actually provide them.

7. The prop-firm reality check

We also ran the model through a realistic 150K-style evaluation: $9,000 profit target, $4,500 end-of-day drawdown, with a funded year modeled after passing.

Figure 4 — upload fig4-prop-firm.png and set its URL here

Note the baseline: a coin flip passes about a third of these evaluations, because the target (45R) is roughly twice the drawdown limit (22.5R). So "passes sometimes" means nothing on its own — and the full ICT model passes far less than a coin flip.

8. What this does — and doesn't — establish

What the evidence supports: the full reversal sequence had negative expectancy over 16 years, was negative across the entire parameter grid and in the out-of-sample split, and was indistinguishable from noise (or negative) in synthetic markets. The MSS and FVG filters added nothing beyond a bare sweep.

What it does not establish: that no version of the idea can work. These results depend on one set of rule definitions, bar-based fills (stop assumed first when stop and target hit in the same bar), no commissions in the main figures, and the data limitations noted in Section 2. The sweep-only positive rests on a single, mostly-rising year and decays through it — it needs multi-year testing before anyone relies on it.

Tests still to run

• Full 24-hour NQ history to reproduce the log with true midnight-based levels, plus walk-forward splits.
• A forward-return event study after sweeps vs. random bars, independent of stops and targets.
• Matched-pairs comparison of each filter on identical sweep events.
• Placebo levels, shifted entry times, and replication on ES, MNQ, RTY and YM.
• White's Reality Check / deflated Sharpe to correct for multiple testing, and tick-level replay with a full cost model.