Fading the crowd with the COT report: the full test of Jason Shapiro's free-video rule

What we granted, and what we actually tested

Granted, not measured: Jason Shapiro is a long-tenured, respected practitioner: Jack Schwager profiled him as the contrarian in Unknown Market Wizards (2020). By his own account he has traded futures for more than thirty years and has used this report as his main data set for more than twenty (§6). He publishes his method free on his channel, Crowded Market Report. His own doctrine is more cautious than the internet’s version of it: he says “nobody sees the future”, that the report “sets you up in good risk reward trades that's it”, and he encourages people not to copy his trades. His record claim, no down year in twenty years, is his claim; we cannot check it, and nothing here needs it to be false.

Tested: The crowding half. Each week of the positioning data, January 2000 to June 2026, the large speculators’ net position in five futures markets (the S&P 500 E-mini, gold, WTI crude oil, the 10-year Treasury note and the euro), ranked against up to 156 weeks of that market’s own positions, ending with that week (fewer in each market’s first three years: the ranking starts once 20 weeks exist). The bottom 10% counted as crowded short and the top 10% as crowded long. After a crowded-short week the test bought at the first closing price on or after the positioning date plus three days, which our engine took to be the Friday the report came out: a Friday close for 1,954 of the 2,072 crowded weeks, and the next close after a Friday with no price for 72; for 15 crowded weeks whose positioning date was a Monday, the Thursday close, the day before that week’s Friday; for 20 crowded weeks reported before a market’s daily prices begin, the first price there is; and for 11 others a Monday, Tuesday or Wednesday close, up to 11 days after that date (§3). After a crowded-long week it bet on a fall. The result is the price change over the next four weeks, the plan’s primary hold, for each market on its own and for the five pooled. The trigger half, without the crowding check. On the first Friday of each month (our rule for the jobs report), a day whose low went below the previous day’s closing price and which then closed above both that price and its own opening price. Entry at that day’s closing price; the result over the next 21 trading days, counted in units of the distance from the entry down to that day’s low. No stop was executed; that low is the ruler. Both halves together. A crowded-short week in the S&P plus a reversal day, which the plan registered as its own check.

§1 · The two plans, written before the runs

Both plans are dated 2026-06-19, and each states that it was locked before the first run. They lived in a private working folder with no version control, so what backs that word is the plans’ own text and the files’ timestamps, not a commit. The last-write times (UTC) are consistent with the order the plans claim: the trigger plan at 18:42:34, its engine at 18:45:03 and its result at 18:46:35; the crowd plan at 19:24:05, written after the trigger result was in (the plan says so itself), then the crowd test’s first run, which wrote its positioning file at 19:26:20. A last-write time is not proof of order. The crowd test’s engine was last written at 19:31:36, after its first run, so it was edited at least once after running, and what changed cannot be recovered. On 2026-09-24 the eight June files were copied, byte for byte, into our research repository, and every later step below is dated there. Those commit times are our own attestation of the timeline, not a public record a reader can inspect.

The falsifiable question

Two, one per half, each with its bars written first. The crowd: over the next four weeks, do prices rise more after the big speculators are crowded short than after they are crowded long, with the gap’s interval clear of zero and the result beating random weeks? The trigger, without the crowding check: over the next 21 trading days, does a jobs-report reversal day drift up with its interval clear of zero, and beat random days?

• A yes looks like: each plan’s PASS: all three bars met. The result clear of zero, above the 90th percentile of the random controls, and still positive after costs and the plan’s own exit; for the crowd, also at least half the markets pointing the same way.
• A no looks like: each plan’s FAIL: the first bar missed.
• Ambiguous residue: each plan’s WEAK: the first bar met and the second missed, which reads as what the market does anyway, not the signal. For the trigger, its plan also said in advance that a fail could not count against his method: it would only say where any edge has to live, in the crowding check or in his judgment.

Our stated prior: one line, in the trigger plan

The trigger plan’s section on a possible second phase says in passing that a fail was “likely for the standalone trigger” (line 143), meaning the trigger tested without the crowding check; the plan was last written before its run. The crowd plan states no expected outcome. The trigger plan also says in advance how each outcome is to be read; the crowd plan says a fail closes the test of the version anyone can build from his free material.

The design as locked

Added, or first run, after the plans

Everything in this table came after the June runs. Three rows were asked for in the plans but not run, or not reported, in June: the other five registered holds, the short mirror’s controls, and the average without its best trade for the trigger tested without the crowding check. The rest were added afterwards.

Promised in the plans and not delivered as written

Deliberately not tested, named in the plans

1. The crowding check inside the trigger test: the trigger was tested without it on purpose. Named in advance: trigger plan lines 13–16.
2. The inflation report and the Federal Reserve decision as news days: deferred because we had no free, exact history of their dates. Named in advance: trigger plan lines 37–42.
3. His discretionary veto, a reading of which setups to skip: designed as a second phase and never signed off. Named in advance: trigger plan line 121; crowd plan line 20.

The bars each plan set

Both first bars missed, so each plan’s own word for its result is fail: for the trigger (tested without the crowding check), a fail its plan said could not count against his method; for the crowd, a fail as run and with the crude splice taken out. Neither third bar was run as written. On 2026-09-24 the channel ruled the crowd’s result corrected from weak to fail. What each plan says, in its own lines:

Corrections to this lab, in order

1. 2026-06-19, the first write-up. The crowd test was recorded as weak, with a small tilt his way clear of zero that could not beat random weeks. The trigger’s comparison with random days of any kind was made in a unit of risk that runs wider on a reversal day, that comparison is not printed here (§4).
2. 2026-09-24, moving the lab into our research repository. Reading the code against the plan, and the data it read, found that the crowd engine added the two sides where the plan said subtract, that the 10-year note’s positioning data stops in February 2022, and that crude oil’s series switches contracts at the same point.
3. 2026-09-24, the correction. A blind reviewer reached the same four points unprompted; the channel then ruled the crowd’s result a fail, and the corrected engine ran at 18:11 UTC, a few minutes after the ruling. The June engine and result file are kept unchanged. A settle-time check then confirmed the early-price flaw the reviewer had raised (§3).
4. 2026-09-24, the later runs. The other five registered holds (21:18 UTC), the calendar check (21:20 UTC) and the trigger on the real release dates, still without the crowding check, its plan committed first (21:32 UTC, run 21:35 UTC). None changed a registered result.

§2 · The headline arithmetic

§3 · Was it the path? The stress tests

Several things could make this result an artifact of how we ran it. We checked what we could and name what we could not. Every interval on this page for the crowd treats each crowded week as independent, but crowded weeks come in streaks whose forward windows overlap, so those intervals are too narrow; that makes passing easier, not harder. The trigger, tested without the crowding check, fires at most once a month, and the lab did not measure whether its intervals are too narrow.

Edge or weather: the trigger’s one year, without the crowding check

Tested without the crowding check, the trigger’s 52 signals added up to +45.1 units of risk, and +31.6 of that came from 2016 alone, 70%. Without its best year, the trigger without the crowding check averages +0.287 per dollar of risk (47 signals) instead of +0.868. The plan asked for a second concentration check, the average without the single best trade: the engine prints it, the June write-up left it out, and the lab’s 2026-09-24 re-run of the unchanged engine prints +0.542 per dollar of risk for the trigger without the crowding check, reported on this page for the first time. Split in two, the trigger without the crowding check averaged -0.138 in 2005–2015 (27 signals) and +1.955 in 2016–2026 (25). Its two registered variants and its short mirror, all tested without the crowding check, are in the trigger table in §7; each one’s interval includes zero.

The other registered holds, run afterwards

The plan listed holds of 1, 2, 4, 8, 13 and 26 weeks with four weeks primary, and at first only four weeks was run. We ran the other five on 2026-09-24, after the four-week result and the correction were known, with the same corrected engine and only the hold changed. The first bar missed in 10 of 10 cells. As run, the pooled gap is below zero at every hold; with the crude splice out it is above zero at four of the five other holds, with every one of those intervals spanning zero. One interval sits clear of zero, the other way from fading the crowd: 13 weeks as run, -0.92 (-1.76 to -0.09). The second bar is met in one cell, 26 weeks with the splice out (98th percentile), where the first bar still misses. At long holds the forward windows of neighbouring weeks overlap even more, so these intervals are narrower still than they should be.

The nine pre-declared settings

The plan fixed three lookbacks and three crowding tails in advance. At four weeks the pooled gap is above zero in 1 of 9 as run (+0.01 points at 156 weeks and a 20% tail) and in 1 of 9 with the splice out (+0.27, the same setting). The other eight are at or below zero in both. These are point estimates with no intervals.

Two flaws in the positioning data

The 10-year note’s positioning series in our saved copy of the report ends on 2022-02-01, so that market covers about 22 years, not 26. Crude oil’s series is the New York WTI contract until 2022-02-01 and then, from 2022-02-08, a WTI contract on a different exchange: our code matched markets by name, the regulator renamed its markets that week, and the second contract was joined onto the first. Ranked against the old contract’s much larger positions, the new one often reads as crowded short: 64 of crude’s 137 crowded-short weeks fall after the switch, 48 of them in the three years when the ranking mixed the two contracts. We report both flaws and repair neither; the crude one we can step around, so every crowd result is shown as run and with crude’s weeks after the switch taken out.

Where the prices begin, and the ranking’s first three years

Yahoo’s daily prices for gold, crude oil, the 10-year note and the euro begin between 2000-08-23 and 2000-09-21, about six and a half to eight and a half months after the positioning data, which starts in January 2000 (February for crude oil); the S&P’s prices start earlier. For the 20 crowded weeks reported before a market’s first price (4 crowded short and 16 crowded long, in gold, the 10-year note and the euro, reported from May to September 2000), our engine used the first price it had as the entry, 6 to 103 days after the report came out, and at the four-week hold 15 of those 20 trades show a change of exactly zero. Two series also have a gap of more than five days: the 10-year note’s prices skip from 2008-06-19 to 2008-07-01, so one crowded week reported in June 2008 was entered 11 days late, and the S&P’s skip from 2001-09-10 to 2001-09-17, which moved one crowded week’s entry to that Monday; at the 1-, 8-, 13- and 26-week holds the same gaps also move one to three exits. The ranking also starts once a market has 20 weeks of positions, so 243 crowded weeks, all in the markets’ first three years (93 crowded short, 150 crowded long), were ranked against fewer than 156 weeks. We report both and repair neither: every crowd figure on this page includes those weeks.

Measured afterwards, on 2026-09-26, on the same frozen inputs: without the 20 early weeks, the pooled four-week figures as run are unchanged at two decimals (+0.23% and +0.62%, a gap of -0.39), and with the crude splice out the second becomes +0.46% instead of +0.45%, the gap still just below zero. No single market’s gap moves by more than 0.28 points at any registered hold, and none changes direction. Ranking only once 156 weeks exist moves the pooled four-week gap to -0.34 as run (95% interval -0.83 to +0.14) and to +0.10 with the crude splice out (-0.33 to +0.53), and no market changes direction at four weeks. At the four-week primary hold the first and second bars still miss, as run and with the splice out, under either change, so the plan’s word is still fail.

Market by market: split, not faint

The S&P and the 10-year note leaned his way; gold and the euro leaned the other way; crude flips with the switch. We read these after the pooled result, from intervals too narrow for weeks that come in streaks, and one market in five is exactly the kind of result a plan exists to stop us picking. No single market here is a finding, in either direction. The S&P is the first market we would test again, on data it has never seen (§8).

The early-price flaw

The report counts positions as of Tuesday (in our saved copy, 104 of its 6,667 weekly rows carry a Monday, Wednesday or Friday date instead) and comes out on Friday at 3:30 pm New York time. The crowd test entered at the daily closing price dated that Friday as our engine reckons it, the positioning date plus three days, or at the next price after it (for 15 crowded weeks whose positioning date was a Monday, at the Thursday close, the day before that Friday; before a market’s daily prices begin, at the first one there is; see above). For each market we asked at what time of day that closing price is fixed, by matching it against the same day’s five-minute prices over about the last 59 days of the feed. The method recovers the known 4:00 pm close of the S&P cash index. In three of the five markets (gold, crude oil and the 10-year note) the price was fixed before the report was public, and most likely in the euro too, where the method is weakest; only the S&P leg traded after it (one of its 417 crowded weeks is among the 15 entered at a Thursday close).

This reads today’s version of the price feed, not what it served for every year back to 2000, and it does not check weeks when the report came out late, or whether the reports behind the 15 Thursday entries were out by that Thursday’s close: none of our files records the positioning report’s release dates. Measured afterwards, on 2026-09-27, on the same frozen inputs: without those 15 weeks the pooled four-week gap is -0.38 as run and +0.00 with the crude splice out, and keeping only the 1,954 weeks entered at a Friday close it is -0.43 and -0.03; no market changes direction at four weeks, and at every registered hold the first bar still misses, as run and with the splice out, so the plan’s word is still fail. We did not measure what the time-of-day head start did to the result. If fading the crowd worked, a head start could only have helped the crowding signal, and the crowding signal still failed.

The jobs-report dates

Our trigger took the jobs report to be the first Friday of each month. Against the Bureau of Labor Statistics archive, 34 of the rule’s 242 dates were not release days, and it missed 49 of the 257 actual releases. As an extra run outside the written plan, with its own plan committed first, we re-ran the trigger without the crowding check on the real release dates: 49 signals, +0.679 per dollar of risk (-0.945 to +2.301). On the real dates, and still without the crowding check, it missed its first bar, as did the two other arms that run’s own plan asked for, V3 and the short mirror; the run also computed V2, which missed its first bar too and is not in the trigger table. The registered result is the first-Friday one, and this run cannot change it.

§4 · The controls

Random jobs-report days (the trigger, tested without the crowding check). The plan compares the trigger with jobs-report days picked at random, measured with the same ruler. Against random sets of 52 drawn from the 227 eligible jobs-report days, whose average was +0.686 per dollar of risk, the trigger without the crowding check landed at the 54th percentile. Its plan also asked for random days of any kind; in its unit of risk that comparison is tilted against the trigger, because a reversal day’s low sits further below its close (the unit is 0.84% of the price on the days the trigger fired without the crowding check, against 0.51% on any day). So we do not print that percentile. The check in plain percent, run afterwards and not in the plan, is in §2: the trigger without the crowding check sat behind most random days.

Random weeks (the crowd). For the second bar the engine picks weeks at random in the same market, as many as the crowded weeks, treats them as crowded, and works out the same gap, 2,000 times. The ledger gives each cell’s percentile against those random gaps, and the after-crowded-short result against random weeks. The bar was the 90th percentile.

Shuffled timing (the crowd). The engine also shuffles which weeks count as crowded, 2,000 times, keeping each market’s own weeks. The ledger gives that percentile too. It is the null that measured our error:

The receipt for the error. With the crowded weeks shuffled, the number the June engine computed, the two sides added together, comes to twice the market’s own drift, within a hundredth of a point in 5 of 5 markets. That is what the June statistic measured when the report carried no information at all. The pooled June figure even sat below its own shuffled value. We print neither June figure.

A blind reviewer. Before we relied on the correction, a separate AI reviewer was given the crowd plan, the June engine, the June result and the positioning data, with no conclusion supplied. It reached the same four points on its own: the two sides were added; with no information that sum is twice the drift, not zero; on the plan’s gap the pooled result misses the first bar, so the plan’s word is fail; and the 10-year series is cut short while crude’s is spliced. It also raised the early price, which §3 then checked.

Both halves together. Of the 52 days the trigger fired without the crowding check, 3 fell in a crowded-short week for the S&P, 8 in a crowded-long week and 41 in neither. Three signals in 21 years support no inference, so we print the counts, not the averages the lab computed; the difference from the crowd alone that the plan asked for was never computed.

§5 · The case against our own test

The strongest case against our test, in four points. These are our words, not his, and every one of them is fair. We sort them by what has a run behind it.

1. Breadth. By his own account he looks at thirty-seven markets and trades only the few showing the most crowded positioning, and at any one time that is a small handful. We ran five markets one at a time and took every crowded week. Our June result file itself calls our design the weaker one. Never run. It is a new test, with its own plan written first, and clean prices for thirty-seven futures markets are a cost we have not worked out.
2. One kind of news. In the example he likes to give, the bad news was an inflation number; our trigger used only the jobs report, and a date rule that picked the wrong day 34 times in 242. Partly run: only the jobs report was ever run, and no other kind of news was tested. The date rule was checked, and the trigger, re-run without the crowding check on the real jobs-report dates, still missed its first bar. Both halves together fired three times in 21 years, too few to read.
3. What the report is for. By his own account the report tells you where the odds sit and sets up the risk-reward, and the news is how he picks the turn (§6). Every bar we ran asked about the average price move after a crowded week; none asked about the risk side, such as how deep the losses ran or whether a stop survived. Never run. We have measured nothing on the risk side and cannot tell you he is right or wrong about it.
4. We traded early. In three of the five markets our entry price was fixed before the report was public, and most likely in a fourth, the euro (§3). Confirmed by a check for three, most likely for the euro; effect not measured. If fading the crowd worked, the head start could only have helped it, and the crowding signal still failed.

Why trust the fix, when we got the first version wrong? Because the fix is checkable without trusting us: the shuffled-timing receipt in §4 shows what the June number measured, a blind reviewer reached the same four points unprompted, the June engine and its result file sit unchanged beside the corrected one, and the ledger below is the file the figures come from.

§6 · Their own terms

His words below are quoted exactly as the caption transcripts of his two videos carry them. We stored both and checked every quoted line against them word for word, spelling included (the captions write the report’s name in lower case). Nothing else on this page is in quote marks as his words; every other line about what he says is our paraphrase, credited to the video it comes from. The video we quote most is unlisted on his channel and appears in his public teaching playlist. We call him he and his, as the episode page of his interview on The AlphaMind Podcast (#72) does.

Reported, not quoted. By his own account he has traded futures for more than thirty years (The Only Mathematical Way to Beat Wall Street, 0:00), has been a hedge fund manager for more than thirty (I'm 59, I've Been Trading Since 25. Watch this., 0:14), and has used this report as his main data set for more than twenty (Why Trading with COT Data is the Key to Success: The Jason Shapiro Strategy, 1:28). He says he looks at thirty-seven markets as potential trades and trades only the ones showing the most crowded positioning, a small handful at any one time, maybe one, two or three (The Only Mathematical Way to Beat Wall Street, 0:30 to 0:52). The example of a news failure he likes to give is an inflation number in 2022 that came in much higher than expected, a market that went down and then closed up, which he calls his trigger to get long (I'm 59, I've Been Trading Since 25. Watch this., 3:31 to 4:05).

What we did with his words. The trigger is his news failure turned into a rule you can read off a chart; the crowd test is his reading of the report turned into a ranking. Neither is his whole method, and his own framing, that the report sets the odds and the news picks the turn, is the reason §5 counts the risk side as untested.

§7 · The complete ledger — all 72 rows

The trigger rows, without the crowding check

Each arm of the trigger, tested without the crowding check: all four on the registered first-Friday calendar, and on the real release dates (an extra run outside the plan) the three arms that run’s own plan asked for. That run also computed V2 on the real dates, again without the crowding check; it missed its first bar too and is not tabled. The first row is the registered primary result.

§8 · What would change our mind

1. His design. Each week, find the most crowded of about thirty-seven markets and trade only those, across the full calendar of news releases, with a fresh plan written before the run. If it comes back positive, this verdict narrows to the one-market-at-a-time version.
2. The S&P on data it has never seen. It leaned his way (+1.02 points, at the 100th and 98th percentiles of its two nulls) and is the one market whose daily closing price is fixed after the 3:30 pm release (one of its 417 crowded weeks was entered at a Thursday close; §3). But we singled it out after seeing all five, with intervals over 93 streaks of crowded weeks that are too narrow; only a new test on unseen data, planned first, can make it evidence.
3. Our data, repaired. Re-fetch the renamed positioning series, freeze the prices, enter only at prices fixed after 3:30 pm, and re-run the corrected test unchanged.
4. Both halves together, with enough signals to read. The full calendar of releases across many markets, so the combined setup fires often enough; that calendar costs money.

§9 · What survives

The report itself survives. It is free, it is real, and every week it shows you where the big speculators stand. His advice on stops was never on trial here, and nothing we found argues with it.

What did not survive, on the version we could build, is the idea that a crowded reading on its own tells you which way to trade next. The check to carry out of here is the one that measured our own number: when someone shows you a chart of trades taken after crowded readings, ask what the same trades would have made on random weeks. When a number beats zero but cannot beat random weeks, believe the random weeks.

§10 · Tools for your own testing

Everything we publish is free — the reports, the result ledgers, all of it. This test ran on free public data; most of our tests need market data we pay for, and that is the one thing that costs us real money.

Positioning: the Commodity Futures Trading Commission’s free weekly Commitments of Traders report, legacy futures-only annual files for 2000 to 2026, the non-commercial long minus short position for each market, saved once on 2026-06-19 (6,667 weekly rows) and never re-fetched. The 10-year note’s series in that copy ends on 2022-02-01 and crude oil’s is spliced across two contracts (§3).

Prices: Yahoo Finance’s free daily data through the yfinance library, the daily closing price as served: the S&P 500 cash index, gold, WTI crude oil, 10-year note and euro futures in their continuous front-month series, and SPY with its daily open, high, low and close for the trigger. For the settle-time check (§3) we also used the same vendor’s five-minute intraday prices for the five markets over about the last 59 days, downloaded through yfinance on 2026-09-24. We do not control this vendor, we did not check how it joins one futures contract month to the next, and prices count the price change alone, with no dividends. The June runs, and the first runs of 2026-09-24 including the corrected four-week run, downloaded prices afresh; the later runs that day (the other holds and the real-date re-run) froze the inputs they used, and our re-check below reads those frozen files.

Release dates: the Bureau of Labor Statistics archive of Employment Situation releases, saved on 2026-09-24. Cash and risk-free rates: none are used. No other market data went into this report. All of the data behind these results came from three sources: the Commodity Futures Trading Commission, Yahoo Finance and the Bureau of Labor Statistics. They do not pay us, have no say in a verdict, and none of them is the link below. EODHD, the affiliate below, supplied none of it and has no say in a verdict.

Checked by us: re-running the lab’s own crowd engine on the frozen inputs at all six registered holds reproduced 768 of 768 of the fields we compared: the ten figures per cell that the ledger carries, for all 72 cells, and each of the 12 blocks’ three bars and verdict. It also reproduced 18 of 18 settings rows, after a planted mismatch was caught first. It did not compare the other figures the logs print, such as the three more on each cell’s row (216 in all; the receipt table in §4 uses ten of them). That is consistency with the recorded runs at their printed precision, not a check on the vendor’s prices. The trigger, tested without the crowding check, was not re-run for this page; the lab’s own re-run of 2026-09-24, with one more day of prices, reproduced every figure of the June trigger write-up that this page prints, and it is the source of the average without the best trade (§3), which June never reported.

Get every report + the 7 Questions one-pager (free)