Does ICT or Smart Money Concepts Work? What the Data Says

ICT (Inner Circle Trader) and Smart Money Concepts (SMC) have no demonstrated edge as trading rules: we found no peer-reviewed test of them, and the most systematic test we found, an unreviewed 2026 preprint, reports that liquidity sweeps signal continuation, the opposite of the textbook reversal. Both names describe one vocabulary for reading price charts (order blocks, fair value gaps, liquidity sweeps, structure breaks, kill zones) that largely renames older supply-and-demand, gap and session ideas, and two of its pieces rest on documented market mechanisms.

Key takeaway

ICT and Smart Money Concepts have no peer-reviewed evidence of an edge, and the popular claim that fair value gaps usually fill ignores the base rate: in a simulated random walk, 83.4% of bullish fair value gaps were revisited within 20 bars, against 83.5% of random levels at the same distance. Stop-loss clusters just beyond round numbers and the busy London and New York hours are real, but an ICT rule earns a place in a trading plan only after it beats a random baseline, after costs, on data it was not built on.

What are ICT and Smart Money Concepts?

ICT and SMC are a set of named chart patterns and timing windows that claim to show where large institutions (the smart money of the name) place their orders; SMC is the broader label for largely the same vocabulary. No official rulebook fixes the terms, so the table gives each as commonly taught, beside the older idea it overlaps.

Concept	As commonly taught	Older idea it overlaps	Documented mechanism
Order block	The last opposite-colored candle before a strong move that breaks structure, treated as a zone where price reacts on return	Supply and demand zones	None found as defined
Fair value gap (FVG)	A three-candle imbalance: candle 1's high below candle 3's low (bullish), or candle 1's low above candle 3's high (bearish)	Price gaps and gap-fill trading	None found; in a random walk, fill rates match random levels
Liquidity sweep or raid	Price trades briefly beyond an obvious high, low or equal highs, triggers resting stops, then reverses	Stop runs and failed breakouts	Partly: stop-loss orders cluster just beyond round numbers
Break of structure (BOS) and market structure shift (MSS)	BOS: price breaks the prior swing in the trend's direction. MSS, or change of character (CHoCH): the first break against the trend	Higher highs and higher lows	None found as defined
Premium and discount	Price above (premium) or below (discount) the 50% midpoint of the range; buy in discount, sell in premium	The 50 percent retracement rule	None found
Optimal trade entry (OTE)	A pullback of roughly 62% to 79% of the latest swing, with 70.5% often called the sweet spot	The 61.8% Fibonacci retracement	None found
Kill zones	Session windows, commonly about 2:00 to 5:00 a.m. (London) and 7:00 to 10:00 a.m. (New York), New York time	Trading the session opens	Partly: open currency orders peak during the overlap of the London afternoon and New York morning

The ideas behind premium/discount and OTE predate ICT: a Federal Reserve Bank of New York study, Osler (2000), lists the 50 percent retracement rule among practitioner conventions and notes that "Fibonacci series, which are widely used, suggest that 38.2 percent and 61.8 percent retracements of recent rises or declines are common."

A bullish fair value gap is a price band that traded only during candle 2, and the claim is that price returns to it. In the random-walk simulation below, the average gap was 0.61 bar standard deviations wide, with its top 0.83 standard deviations below candle 3's close.

A fair value gap is fully defined by three candles, which makes it testable. In a pure random walk such gaps appear on almost a third of bars, so their existence alone says nothing about direction.

Is there any research on smart money concepts?

We found no peer-reviewed test of ICT or SMC concepts as defined. Crossref searches on 27 September 2026 combining smart money concepts, Inner Circle Trader, fair value gap, liquidity sweep, order block, break of structure and change of character with trading returned no journal article that tests them; the only direct tests were two SSRN preprints (papers posted before peer review).

The more systematic preprint, Mahadzva (2026), coded six ICT/SMC rules mechanically and ran them through walk-forward, bootstrap, permutation and holdout tests on intraday EURUSD, GBPUSD, AUDUSD and USDJPY. Its abstract states: "We find no evidence that ICT/SMC concepts work as literally taught." Three separately built liquidity-sweep rules agreed, in 65 of 66 combined fold selections, that a swept level predicts continuation, not the textbook reversal. A change of character confirmed by a fair value gap retracement did better faded than traded: the flagship result (EURUSD one-hour, faded) reached p = 0.0192 and returned +19.3% on 60 holdout trades with a profit factor (gross profit divided by gross loss) of 1.50. The abstract also reports one strategy whose per-trade edge was smaller than the fixed commission its tight stops forced, until a minimum stop distance was imposed.

The second preprint, Bindra (2026), reports a 78.0% win rate and +265.0R (R is one unit of initial risk) on 473 backtested fair value gap trades in gold and three currency pairs during 2025; its abstract does not mention an out-of-sample test, a random baseline or transaction costs, the three checks applied below.

The best template for testing a level-based claim predates ICT. Osler (2000) compared support and resistance levels published daily by six firms active in the currency market, from January 1996 to March 1998, with 10,000 sets of arbitrary levels in the same one-minute data: "Exchange rates bounced off arbitrary support and resistance levels 56.2 percent of the time on average," against 60.8% for the published levels, an edge of 4.6 percentage points. A pure random walk would bounce about 50% of the time, so even arbitrary levels beat theory by about 6 points in real prices, which is why a baseline must come from the same data. Osler left profitability as a subject for future research. The wider record on chart methods is in whether technical analysis works.

This is what a proper test of a level-based idea looks like: an arbitrary-level baseline from the same data. Published levels beat it by 4.6 points, a real but modest effect, and far from the certainty ICT content implies.

Are liquidity sweeps real, or just stop hunts?

The mechanism behind a liquidity sweep is real: stop-loss orders cluster just beyond obvious price levels, so price that trades through such a level sets off a burst of market orders. Osler (2001 working paper; Journal of Finance, 2003) studied 9,667 stop-loss and take-profit orders at a large currency dealing bank from September 1999 to April 2000 and found that "stop-loss buy orders tend to be clustered just above round numbers." Among executed stop-loss buy orders, 14.4% had requested prices whose last two digits were 01 to 10 (just above a round 00) against 7.4% at 90 to 99 (just below); executed stop-loss sell orders showed the mirror image, 10.0% just below against 5.1% just above.

Osler measured round numbers, not swing highs; the swing-point version rests on practitioner habit, which the same paper records: "market practitioners recommend placing stop-loss orders just beyond support and resistance levels." Whether anyone runs those stops on purpose is covered in is stop hunting real.

After the sweep, the evidence points away from the textbook. Osler's follow-up study (Journal of International Money and Finance, 2005) found that exchange rate trends accelerate when rates reach levels where stop-loss orders cluster, and Mahadzva's continuation result points the same way, so the ICT reversal trade rests on the part without evidence.

Stop clusters beyond round numbers are documented, so the "liquidity" in ICT language has a real basis. The part that fails the data is the usual claim that price reverses after the sweep; the evidence leans toward continuation.

Why do ICT kill zones line up with real liquidity?

Kill zones sit on the hours when the currency market is busiest, a documented fact about liquidity rather than evidence of a directional edge. The Bank for International Settlements (2025) puts global FX turnover at $9.6 trillion per day in April 2025, with the United Kingdom at about 38% and the United States at about 19% by sales desk location; those shares describe where trades are booked, not the hour.

Evidence on timing is mixed. In Osler's bank data, open conditional orders peaked during the overlap of the London afternoon and New York morning, and Osler (2000) notes that major US macro announcements generally came at 8:30 a.m. New York time. On the interbank electronic broking system, Ito and Hashimoto (2006) found that "The U-shape of intra-day activities (deals and price changes) and return volatility is confirmed for Tokyo and London participants, but not for New York participants," with more activity going along with higher volatility and narrower spreads.

Busy windows offer tighter spreads and larger moves in both directions, not a bias toward one. The same logic built older session-open methods such as the opening range breakout, and futures traders can check their own contract's active hours in the best time of day to trade futures.

Do fair value gaps really get filled most of the time?

Fair value gaps do get filled most of the time, and so does almost any nearby price level, so a fill rate alone proves nothing. Fill-rate claims in the 70% to 90% range circulate in trading content, but we found none traced to a primary source or reported beside the fill rate of comparable random levels.

To measure that baseline, we simulated 1,000,000 bars of a pure random walk, a price series with no memory and so no edge (each bar built from 20 normally distributed sub-steps, open equal to the prior close, NumPy random seed 42). A fair value gap formed on 29.8% of bars (14.9% bullish, 14.9% bearish). Within 20 bars, price revisited the near edge (the gap top) of 83.4% of the 148,789 bullish gaps with enough bars after them and filled 73.1% completely; random levels at the same distances below price, drawn ten per gap, were revisited 83.5% of the time and filled 73.1%. Revisit rates were 68.5% against 68.6% at 5 bars and 92.5% against 92.6% at 100 bars. A second run with seed 7 gave 83.6% against 83.5% for revisits and 73.3% against 73.2% for full fills; across both runs the gap-minus-random difference in revisit rates never exceeded 0.12 percentage points and changed sign between seeds, which is sampling noise. In a memoryless series, the path after any bar is independent of the pattern that formed there, so the two rates must match on average.

"Fair value gaps get filled 80% of the time" is true and meaningless: random levels at the same distance get touched just as often in a market with no structure at all. A gap claim needs to beat the grey bars, on real data, after costs.

Random prices touch nearby levels often. For Brownian motion (the continuous random walk), the probability that price touches a level d away within n bars is P = 2 × (1 - Φ(d ÷ (σ√n))), where σ is one bar's standard deviation and Φ the standard normal distribution. For a level one bar-sigma away, P is 31.7% within 1 bar, 65.5% within 5, 75.2% within 10, 82.3% within 20, 88.8% within 50 and 92.0% within 100, so any close level, given time, looks like a magnet.

A touch is also not a profit. In a driftless market, a trade with equal stop and target wins 50% of the time before costs wherever it is entered (the gambler's ruin result: the chance of reaching a target a away before a stop b away is b/(a+b)). Buying at market the moment price first traded at or below a gap top within 20 bars, with a stop and a target each one bar-sigma away, won 50.2% of 123,750 simulated trades, against 50.1% for the same trade at random levels. The simulation omits volatility clustering, session effects and drift, so it models the base rate, not the futures market.

How do you test an ICT concept properly?

Test any chart concept, ICT or otherwise, by comparing its outcome rate with a random baseline from the same data, converting the difference into ticks after costs, and repeating the test on data the rule has never seen. Seven steps keep noise from passing as edge:

Write the definition as code. Two people applying it to the same chart must mark the same bars, so a strong move or an obvious high needs a number.
Lock away out-of-sample data first. Set aside the most recent part of your history (for example, the last 30%) and do not look at it until step 7.
Count every event and its outcome. Record all occurrences, not the memorable ones, and the exact result claimed: a touch, a full fill, or a move of a stated number of ticks.
Build the random baseline from the same data. Measure the same outcome at random bars, at the same distance from price and in the same session hours.
Turn it into a trade and charge costs. Fix entry, stop and target, include commission and slippage, and compute the break-even win rate before looking at results.
Count every variant you tried. After tuning many settings, demand a t-statistic above 3, not 2; the reasons are in how futures backtests overfit.
Run once on the locked data, then trade small. Scale only a result that survives the locked data and a stretch of live trading at minimum size.

The table walks one fair value gap rule through those steps. Row A is computed from the simulation; rows B to D are illustrative numbers, not measurements.

Row	Objective definition	Events	Outcome rate	Random baseline	Difference	Net after costs (ticks per trade)
A (computed, random walk)	Bullish FVG; gap top revisited within 20 bars	148,789	83.4%	83.5%	-0.1 pp	Not a trade
B (illustrative, 5-minute futures bars)	Same definition	1,000	86.0%	83.0%	+3.0 pp	Not a trade
C (illustrative, in-sample)	Buy limit at gap top on first touch; 8-tick stop, 8-tick target	600	58.0%	50.0%	+8.0 pp	+0.28
D (illustrative, out-of-sample)	Same rule on later, untouched data	300	52.0%	50.0%	+2.0 pp	-0.68

The random baseline is the same outcome at random bars and matching distances in the same data; for rows C and D it is the 50% a symmetric bracket wins by chance. Net after costs = T × p - S × (1 - p) - C, where T is the target, S the stop, C the round-trip cost and p the win rate; with T = S = 8 ticks and an illustrative C of 1 tick, net = 16p - 9, so break-even is 9/16 = 56.25% and a random entry nets -1.00 tick.

Row B beats its baseline (z = 2.53, the difference in standard errors), but a touch rate says nothing about profit. Row C looks excellent at z = 3.92 and +0.28 ticks per trade; row D, the same rule on later data, is 2.0 percentage points over chance at z = 0.69 and loses 0.68 ticks per trade. Row C even clears the stricter hurdle that Harvey, Liu and Zhu (2016) propose for new findings in a heavily data-mined field, "a t-ratio greater than 3.0," and still fails out of sample, which is how overfitting looks in practice.

The in-sample result would convince most people; it is significant and profitable after costs. The rule still dies on data it has not seen, which is the only test that counts and the one most concept content skips.

Testing 20 independent rule variants at the 5% level gives a 64.2% chance that at least one looks significant by luck (1 - 0.95^20), and 50 variants give 92.3%. The sample needed to show an edge is n = z² × p0(1 - p0) ÷ (p - p0)², with p0 the break-even or baseline rate: proving that a true 58% win rate beats the 56.25% break-even at z = 2 takes about 3,215 trades, roughly 12.8 years at one trade per trading day and 252 trading days a year. Mahadzva's flagship p = 0.0192 is under 0.05 but well above the p of about 0.0027 that a t-ratio of 3.0 implies, and the preprint's abstract reports that the result fails a Deflated Sharpe Ratio check (a correction for the number of strategies tried) at every defensible trial count.

Can you pass a prop firm challenge with ICT or SMC?

You can pass an evaluation with ICT or SMC, as with any method, because a short challenge is dominated by luck; keeping a funded account depends on whether your specific rules beat a baseline after costs. A strategy with a true 50% win rate on a symmetric bracket wins 12 or more of 20 trades 25.2% of the time, so a strong 20-trade run says little about skill.

A concept can be useful as a framework for reading the chart and still have no measurable edge as a rule. ICT vocabulary points attention at session opens and at obvious highs and lows where stops rest, both real features of markets; order flow and the DOM show resting liquidity directly, and a volume profile shows where volume actually traded. Turning any framework into a tested edge is the slow part of becoming consistent, a timeline covered in how long it takes to become a profitable trader.

A copier multiplies expectancy, not edge

A trade copier, including Thor (this blog's own product), repeats a master account's trades on other accounts, so it multiplies whatever the master's rules earn and is no fix for an untested strategy. A rule netting -0.68 ticks per trade, copied to five accounts, loses 340 ticks across them over 100 trades instead of 68 on one account.

Go deeper

Frequently asked questions

Is ICT trading a scam?

Not in the sense of hidden rules: the concepts are published openly and can be tested, and what they lack is evidence of an edge. Judge any course or method by whether its rules are precise enough to test against a random baseline.

Are order blocks the same as supply and demand zones?

Largely, yes: an order block is a narrower rule for drawing a demand or supply zone, anchored on the last opposite-colored candle before a strong move. The narrower rule is easier to test, which is its main practical advantage.

Do fair value gaps always get filled?

No. In our random-walk simulation, 12.3% of bullish gaps were still not completely filled after 100 bars, and gaps that open far from price, relative to the time allowed, fill less often.

What win rate does an ICT or SMC setup need to be profitable?

The answer depends on the target, stop and costs, not the concept: break-even win rate = (S + C) / (T + S). With a 16-tick target, an 8-tick stop and 1 tick of round-trip cost (illustrative), break-even is 9/24 = 37.5%, so a quoted win rate means nothing until the target and stop are stated.

Why do so many traders say smart money concepts work?

The patterns appear on every chart, and hindsight always finds examples that worked. Highlighted examples are a biased sample: failed setups rarely get posted, and without a count of every occurrence there is no denominator to judge them by.

How many trades do you need to test an ICT setup?

Hundreds at minimum and thousands to confirm a thin edge. At z = 2, separating a true 58% win rate from a 50% baseline takes about 157 trades, while separating 52% from 50% takes 2,500.

Does classic technical analysis have better evidence than ICT?

Somewhat, because it has been tested more rigorously. Lo, Mamaysky and Wang (2000, Journal of Finance) detected patterns such as head-and-shoulders with an algorithm, compared returns after each pattern with returns at all other times on US stocks from 1962 to 1996, and concluded that several indicators carry incremental information that may have practical value.

Do ICT kill zones apply to futures markets?

The liquidity mechanism transfers, but the clock does not automatically: kill zones were drawn around currency sessions, and a futures contract's busiest hours depend on its own exchange session and scheduled data releases. Check your contract's volume and spread by hour before assuming a window matters.

Is a liquidity sweep the same as a stop hunt?

Both terms describe the same price event, a quick move through a level where stop orders rest. Stop hunt adds a claim that someone caused the move on purpose, which price data alone cannot show.