A paper posted on SSRN in September 2026 set out to test the rules behind ICT and smart money concepts on real forex quotes. SSRN is a site where researchers share work before any journal has reviewed it. The abstract says the author finds "no evidence that ICT/SMC concepts work as literally taught". The title, "Smart Money or Costly Folklore?", is a question with a side already chosen.
The result is more mixed than the title, and the paper is more careful than most of the social media posts that will quote it. This post goes through what was tested and what came out. It covers where the author says to be careful, and one separate backtest on a different market that points the same way.
Quick timeline
What is the paper, and who wrote it?
The paper is "Smart Money or Costly Folklore? A Systematic Evaluation of ICT/Smart-Money-Concepts Trading Rules in G10 FX" by Shepherd Mahadzva, who lists no institution and signs as an independent researcher. It is Working Paper 3 of 10 in a self-published project called The Deflated Alpha Series.
The author's note on the first page is blunt about the limits. It says the paper is an independent working paper that has not been through peer review (the process where other experts check a paper before a journal prints it). It says all data come from one retail MT5 broker feed, not from an academic data set. It also says an AI language model helped with drafting and revision under the author's direction. SSRN lists no funding and no conflict of interest.
None of that makes the numbers wrong. It does mean nobody outside the project has checked them yet.
What exactly did it test?
Six rules that the author turned into strict, mechanical instructions a computer can follow. Here is each one in plain words, using the paper's definitions.
- Liquidity sweep reversals. Two swing highs or lows form at about the same level. Price pokes through the level, closes back inside, and the textbook says it will now reverse.
- A cost fix for that strategy. A minimum stop distance, added because tiny stops and a flat commission were eating the edge.
- Channels and trendlines. Included on purpose as a negative control, a test that is expected to find nothing, to show what a meaningless result looks like.
- Order blocks and Change of Character. An order block is the last down candle before price breaks above a recent swing high, with the mirror image for the other side. The study waits for a shift in short-term structure before entering.
- Asian session range sweeps. The overnight range is treated as a shelf of resting orders that London or New York may sweep.
- Change of Character plus a Fair Value Gap retracement. A fair value gap is a three-candle hole: for a bullish one, the low of the third candle sits above the high of the first. The rule buys (or sells) when price returns into that gap after a structure shift.
The data are four pairs: EURUSD, GBPUSD, AUDUSD and USDJPY. Hourly history runs from September 2020 to September 2026 (USDJPY from September 2021), 15-minute history from September 2022, and about ten weeks of one-minute data for one study only.
How was it tested?
Costs are included, which many backtests skip. Every backtest charges half the median historical bid/ask spread, a flat $7 round-turn commission per standard lot, and one point of adverse slippage on entry and exit. Position size risks 1% of equity per trade.
The test is called walk-forward. The computer picks the best settings on an older slice of history, trades the next slice it has never seen, then rolls forward and repeats, six times. The newest 15% of each pair's history is locked away as a holdout and looked at once, at the end.
Three checks sit on top. A bootstrap resamples the trades 5,000 times to see how wide the uncertainty is. A permutation test (5,000 draws) randomises the direction of trades to ask how often luck alone would match the result. A Deflated Sharpe Ratio, a published method, then adjusts for how many strategies the project has tried in total.
The author also left the trade direction free. The computer could choose to trade with the textbook or against it. That is the clever part, because it lets the data say which way is right.
What did it find?
The headline is the direction. In the liquidity sweep rules the computer chose to trade with the move, not against it. That happened in 24 of 24 slice-by-slice picks in the first study and 18 of 18 in the Asian range study. Across the three sweep-style constructions the abstract counts 65 of 66 picks for continuation. In plain words, when price poked through an obvious level and came back, the search never preferred the textbook reversal. The paper reads that as the sweep confirming a real break. It does not present trading the break as a proven money-maker, and the first study also ran into a cost problem, because its own stops were so tight that commission mattered.
The fair value gap rule went the other way. The computer chose to bet against the textbook entry in 23 of 24 picks. A bullish structure shift followed by a pullback into the gap was, on this evidence, better sold than bought.
Two results cleared the author's significance bar, both on the hourly chart with the fade direction:
Expectancy in R means the average win or loss measured in multiples of the amount risked. USDJPY (p about 0.054) and GBPUSD (p about 0.19) did not pass. The 15-minute version of the same rule found nothing.
Two other items from the paper are worth a trader's time. First, a strategy's edge can be smaller than the commission it creates when the stop is tiny and the size is large, which is why the author added a minimum stop distance. Second, a real bug in the order block code, a wrong starting candle, passed every statistical gate and was found only when a human looked at a chart.
Where does the paper say to be careful?
The author lists the cautions, and they matter more than the headline.
- The strict test fails everywhere. After the Deflated Sharpe adjustment, every strategy in the paper scores 0.00. The author calls this an expected result of screening at the scale of the project and says none of the findings should be read as clearing that very high bar.
- AUDUSD looks fragile. Its walk-forward efficiency, the ratio of out-of-sample to in-sample results, is 2.08, above the author's own ceiling of 1.7. The author says it could be real or good luck and does not resolve it.
- Small numbers. The two best results rest on 63 and 48 trades. In the Asian range study, an early version with 33 trades was too small to be conclusive, and a version with 48 trades crossed the significance line.
- Four pairs and four to six years. Wider data could change the verdicts, especially for AUDUSD.
- One broker, one cost model. The commission and spread figures are from one feed. The numbers will shift elsewhere.
- Mechanical, not discretionary. The setups come from a tradition with no formal theory behind them. The paper tests rigid versions, and the author notes that the economic stories offered for the results are plausible readings, not tested hypotheses.
That last point is where a fan of the style would push back. A rule a computer can run is not the same as a trader reading context, and the paper does not claim to test discretion.
What did the separate backtest find?
A different write-up, "I Backtested ICT / Smart Money Concepts, What Survives" on the StatOasis site by Ali Casey, was published on 3 September 2026 and updated on 17 September. It tests on daily bars of four US stock index funds (SPY, QQQ, DIA and IWM) with histories starting between 1993 and 2000. That is a different market and a much slower chart than the paper above.
It turned order blocks, fair value gaps, liquidity sweeps and a Fibonacci "optimal trade entry" into long-only rules. Together with three simple comparison entries that made 54 variants, tried with three exit times on each of the four funds, for 648 backtests. According to the article, only one of 32 statistical tests on forward returns passed a t-statistic of 2. No variant beat simple buy-and-hold on net profit. Order blocks were the most consistent family, beating random entries in most variants. The author says the test is frictionless (no costs) and does not speak to intraday or discretionary trading. The author runs a trading course and a backtest-grading app, and both are mentioned on the page.
The two studies use different markets, timeframes and rules, so their numbers cannot be added together. They agree on one thing: the simple versions do not show a strong edge.
What the other side says
We could not find any public reply to the paper from Michael Huddleston, who is ICT, or from fans of the style. He has not commented on it that we could find. The nearest thing to a statement is a March 2025 reply to a follower, in which he told people to backtest and then forward-test for months before trading live. It was written long before the paper and is not a response to it.
Reactions on X could exist that our searches did not reach. We would not guess what they say.
What traders can take from it
- Test the exact rule you trade, with costs. A good backtest charges slippage and commission on every trade. Many do not.
- Count your trades. Fewer than 30 is thin, and 48 or 63 is only a little better. See sample size.
- Let the data pick the direction. If you only test the textbook version of an idea, you will never learn that the opposite works better.
- Look at the chart, not only the statistics. The bug in the paper's order block code got past every test and was caught by eye.
- Separate "the idea" from "my version of the idea". A journal built from your own broker data records what you actually traded, with costs, so you can check your own numbers instead of borrowing somebody else's. Start with backtesting and a trading journal.
Frequently asked questions
Does the 2026 SSRN paper prove ICT does not work?
No. It is a preprint that has not been peer reviewed. It tests six rigid, mechanical versions of ICT and smart money ideas on four forex pairs, from one broker, and finds no support for the textbook direction. It does not test a trader using judgment.
What did the paper find about fair value gaps and liquidity sweeps?
In its data, trading the textbook reversal after a liquidity sweep lost out to trading with the break. For the fair value gap rule, the best result came from betting against the textbook entry on the hourly chart of EURUSD and AUDUSD. The strictest test in the paper still scored every strategy at zero.
Who is Shepherd Mahadzva?
The author is listed on SSRN as an independent researcher with no institution. The paper says it is self-funded, drawn from one retail broker's data, and drafted with help from an AI language model. We did not find a public profile beyond the SSRN pages.
Can I check the paper myself?
Yes. The abstract page and the PDF are free on SSRN, and the licence allows non-commercial sharing without changes. The paper says its methods follow Paper 1 of the same series, so that paper is the place for the formulas.
Are there other ICT backtests?
There is a separate daily-bar test on US index funds by StatOasis, summarised above. It uses different rules and a different market. Treat any backtest, including these two, as a starting point and check how many trades it used, whether it charged costs and who wrote it.
Sources
- Smart Money or Costly Folklore? A Systematic Evaluation of ICT/Smart-Money-Concepts Trading Rules in G10 FX, SSRN abstract page: abstract, dates, licence, download counts. We also read the PDF behind it.
- I Backtested ICT / Smart Money Concepts, What Survives, StatOasis: the separate daily-bar backtest and its stated limits.
- Post by The Inner Circle Trader on X, 25 March 2025: his reply telling a follower to backtest and forward-test.
This post is based on public sources as of October 30, 2026. Found a mistake? Write to info@setupbook.io.