Do ICT session strategies really have 70% win rates?

Not in our testing. We coded two popular ICT-style session strategies — a New York 8:30 range-reversal on NASDAQ futures and a London-bias order-block setup on GBPUSD — and ran 523 trades with conservative fills. Win rates came out between 18% and 29%, below the mathematical breakeven for their fixed reward-to-risk targets. The 70% figures typically come from manual replay testing, where discretionary choices are made with knowledge of what happened next.

Why do manual backtests show higher win rates than coded ones?

Manual replay testing leaves room for unconscious selection: which order block to prefer, which fair value gap counts as significant, when a candle close is "clean enough." Each small judgment can be biased by having seen similar patterns resolve. Coded rules remove that freedom, and conservative fill assumptions (stop fills first when both stop and target are touched in one bar) remove optimistic execution. The gap between the two numbers is a measure of how much of the claimed edge lives in unfalsifiable discretion.

Does this mean all ICT concepts fail backtesting?

No. In a companion study we tested the ICT fair value gap retrace setup on 53 years of daily Dollar Index data and found a modest but persistent edge: a 38% win rate with roughly two-to-one winners and a +0.29R expectancy, positive in every decade. The daily, rules-heavy framework held up; the fast intraday session setups with headline win rates did not. The difference is which claims survive mechanical translation.

What would make these session strategies testable more fairly?

Longer samples (our NASDAQ test covered about six months, and the strategy was profitable in three of them), explicit confluence rules coded rather than described, and realistic transaction costs. It is possible a regime filter could rescue the New York setup — it lost consistently from September to December 2025 and won from January to March 2026 — but a filter identified after seeing the results would need out-of-sample confirmation before it counts as anything. This article is research, not investment advice.

ResearchJuly 3, 2026· 9 min read

We Coded Two “70% Win Rate” ICT Session Strategies. Neither Survived.

Session-range strategies built on ICT concepts are everywhere: mark a killzone range, wait for the raid, enter the displacement, collect a fixed 2R. The pitch usually comes with a win rate around 70%. We turned two of the most popular variants into code — no discretion, conservative fills — and ran 523 trades on NASDAQ futures and GBPUSD. This is the follow-up to our 53-year fair value gap study, and the results point the other way.

The short version

New York 8:30 range reversal (NASDAQ futures): 18–30% win rate across every documented variant, against a 33.3% breakeven at 2:1. All negative to flat before costs.
London-bias order block (GBPUSD): 463 trades over 2.5 years — 27.4% win rate against a 28.6% breakeven at 2.5:1. −0.04R before spread. Advertised: ~70%.
The gap between advertised and coded numbers is consistent with manual-replay bias: the discretionary layer (“which order block counts”) can’t be coded — and appears to be where the claimed edge lives.
Contrast: the daily fair value gap setup we tested previously held a modest, persistent +0.29R edge across five decades. Slow and rules-heavy survived; fast and headline-grabbing did not.

The two strategies, as mechanical rules

Both setups follow the same ICT template: a time-boxed session range, a liquidity raid through one side, a structure shift with displacement, an entry on the retrace, a fixed reward-to-risk target. They differ only in session and instrument. We tested the versions of these setups that circulate widely in ICT-derived teaching, translated into unambiguous rules:

A — New York 8:30 range reversal (2-minute chart, index futures)

Mark the high/low of the 00:00–08:30 New York window.
The first break of either side after 08:30 is the raid; hunt the reversal.
Enter when a 2-minute candle closes through the most recent confirmed 5-bar fractal swing in the reversal direction, with the displacement leaving a fair value gap — limit order at the gap edge.
Stop above the gap’s first candle (or the raid extreme — both versions are taught). Fixed 2:1 target.
Tuesday–Thursday only; optionally skip days with a degenerate pre-8:30 range.

B — London-bias order block (5-minute chart, GBPUSD)

Mark the high/low of the 03:00–03:55 New York window.
The first 5-minute body close outside the range (before 08:00) sets the session bias.
Enter on the retrace to the opposite-close candle immediately preceding the breakout (the order block) — limit at its near edge, stop beyond its far edge plus a buffer.
Fills accepted 04:00–08:00 New York only. Fixed 2.5:1 target.

Execution assumptions, both strategies: one trade per day, DST-aware session boundaries, positions closed at 16:00 New York, and the conservative rule that when a bar touches both stop and target, the stop fills first. That last assumption biases win rates down— keep it in mind; it does not explain a 40-point gap.

Results A: New York 8:30 range (NASDAQ futures)

Data: NASDAQ 100 futures, 1-minute bars resampled to 2-minute, September 2025 – March 2026. Four variants, covering both documented stop placements and the range-quality filter:

Variant	Trades	Win rate	Expectancy
Filter on, stop at FVG candle	44	29.5%	−0.08R
Filter on, stop at raid extreme	44	18.2%	−0.05R
Filter off, stop at FVG candle	60	26.7%	−0.17R
Filter off, stop at raid extreme	60	21.7%	−0.00R

Breakeven at fixed 2:1 requires 33.3%. Every variant lands below it, before transaction costs. One honest wrinkle: the setup lost consistently September through December 2025, then won January through March 2026 (+7.6R with the filter on). That could be a volatility-regime effect — or noise in a 44-trade sample. A filter invented after seeing the results doesn’t count until it survives out-of-sample data.

Results B: London bias (GBPUSD, 2.5 years)

Data: GBPUSD 1-minute bars, January 2024 – June 2026 — a much larger sample: 463 triggered trades across 30 months.

Metric	Value
Trades	463
Win rate	27.4%
Breakeven required at 2.5:1	28.6%
Expectancy (before spread)	−0.04R
Best / worst month	+10.0R / −12.0R
Typically advertised win rate	~70%

A coin with a 27.4% payout probability needing 28.6% is not an edge; it’s a slow leak that spread costs widen. The monthly distribution is the tell: swings from +10R to −12R with no stable pattern. Whatever produces the advertised 70% is not in the mechanical rules.

Where do the advertised win rates go?

The pattern across this study and our fair value gap backtest is consistent. Claimed numbers come from manual replay testing, where a human steps through historical charts and decides, case by case, which order block is “strong,” which gap is “significant,” which close is “clean.” Every one of those judgments is an opportunity for hindsight to leak in. Code has no such freedom, and conservative fills close the last loophole. The difference between 70% and 27% is a measurement of that leak, not necessarily of anyone’s honesty.

The discretion is the product.When a strategy’s profitability depends on judgment that can’t be written down, it can’t be verified, taught, or automated — only performed.
Session setups are cost-sensitive. One trade a day at 2R targets means spreads and slippage eat a meaningful fraction of each win. Daily-timeframe setups amortize costs far better.
Not everything failed. The slow, heavily-specified daily FVG framework carried a modest +0.29R edge across five decades. The pattern worth internalizing: the more precisely a rule is specified by its author, the better its chance of surviving translation to code.

Methodology & caveats

One trade per day per strategy; DST-aware New York session logic; positions force-closed at 16:00 New York.
Conservative intrabar assumption (stop first) biases win rates down — but cannot account for gaps of this size.
No transaction costs modeled; including them makes every number worse.
The NASDAQ sample is short (~6 months). The GBPUSD sample (30 months, 463 trades) is the more damning of the two.
We tested the mechanical skeleton. Advocates will fairly object that confluence and discretion matter — our point is that whatever cannot be specified cannot be tested, and untestable claims deserve no benefit of the doubt at 70%.

This article is independent research and education, not investment advice or a recommendation to trade any instrument or strategy. Historical results do not predict future performance. The session concepts discussed derive from publicly taught ICT methodology; the code, testing methodology, and conclusions are our own.

We test claims for a living

Seasonal Edge stress-tests recurring market patterns across 4,500+ securities with coded rules, out-of-sample checks, and honest win rates — the same standard we applied here.

Explore Seasonal Edge →