Source: https://forexiapro.com/research/fading-the-crowd-cot-level-hunt/

# Fade the Crowd With the COT Report on a Level Hunt: Promising, Not Proven

Dylan Forexia11 November 20268 min read

Fade the crowd COT report: this is the version of the idea that held up best in our tests, and it is not proven. Trades taken with the crowd lost in 12 of 16 years; fading it on a level hunt made +0.08R a trade over 1,047 trades, promising and not proven. The t-statistic is 1.8, short of the usual bar of 2, and the rule was chosen after seeing most of the data.

R is the amount risked on one trade. Below, each square is one calendar year of the same hunts, handled three ways. The top row is why the idea is worth watching. The middle row is what the same hunts did with no report at all.

[![Fade the crowd COT report test: three rows of yearly squares for the same level hunts, fading the crowd 3 down of 16, no filter 12 down, with the crowd 12 down](/seo/fading-the-crowd-cot-level-hunt/fade-the-crowd-cot-report-three-ways-by-year.webp)](/seo/fading-the-crowd-cot-level-hunt/fade-the-crowd-cot-report-three-ways-by-year.webp)

The same hunts handled three ways, eight markets, March 2011 to 2 October 2026. Down means more than 1R down in the year, up more than 1R up. Fading the crowd is promising, not proven.

## What is the fade the crowd COT report test?

It asks whether a [stop hunt](/glossary/the-stop-hunt/) that takes a prior level pays when large speculators are crowded on the side being hunted. [Fade the Crowd](/glossary/fade-the-crowd/) is our word for the stance. The hunt is what most traders call a liquidity sweep, and a level hunt is one that also takes the previous day's, week's or month's high or low.

In plain terms the rule was this. A program flagged the hunt as it happened, using only candles already printed and only a [COT report](/cot-report/) already published. It took the reversal only when the crowd was crowded against the trade, and held through a wide stop for days. The median hold was 13 days. We report the test here and do not publish its settings.

The data and the costs are the same as in [our test of the COT report as a whole](/research/does-the-cot-report-work/): seven dollar majors and gold, Dukascopy 5-minute bid and ask candles, March 2011 to 2 October 2026, the real spread on every bar, 0.7 to 0.8 pips a round trip for commission and slippage ($0.15 on gold), and financing of 1% a year while held.

## How was the rule chosen?

It was picked after looking, which is the first thing to hold against it. The order matters, so here it is.

1. Fading the crowd's side, by itself, did nothing for a trade closed the same day. The effect builds over days, so the trade became a hold of weeks.
2. Seventeen variants were run on 2015 to 2026. The plain version, every hunt against a crowded side, came out at break-even: -0.001R a trade in 2015-20, -0.043R in 2021-24 and about zero in 2025-26.
3. Two variants stayed above zero in all three periods. One, on M or W hunts only, barely did. The other is the rule here: the same trade, on hunts that also take a prior level. It joined two effects that had each shown up separately in earlier work, and it is the one taken forward.
4. Only then was it run once on 2011 to 2014, data downloaded for the purpose, with the pass mark written down first: money overall, and in at least three of the four years. It passed.

That last step is the only clean test the rule has had. A correction belongs right beside it: eight of the 217 trades in that period fell in January and February 2012, in one dollar-wide reversal, and made about two thirds of the period's result. It passed the mark. It was a much thinner pass than the table in our notes first suggested.

## What are the caveats on the result?

The sentence at the top is only true with these beside it.

| What to hold against it | Figure |
| --- | --- |
| t-statistic, against the usual bar of 2 | 1.8 |
| Sharpe ratio | about 0.4 |
| Worst drawdown | 27R |
| Years more than 1R down | 3 of 16 (2013, 2020, 2024) |
| Trades a year, across eight markets | about 71 |
| Most positions open at once | 8 |
| Financing in the test | a flat 1% a year, for swaps that really depend on direction and broker |

Financing is the largest thing the test could not measure. A trade held for two weeks pays or earns real swaps, and a flat charge stands in for all of them. In 2011 and 2012 the index also had less than three years of history behind it, so those early readings are not the same measure as the later ones.

A rule that holds up to eight positions, in markets that move together, is not 1,047 separate results. Every version of this rule has the same bad years and the same good ones. The whole family is one bet: it fades moves, so it loses in years when the dollar trends and wins in years when the dollar turns.

## Where is the fade the crowd result thin?

It is thin by market. We split the 1,047 trades for this page, and four of the eight markets were at or below zero over the whole record.

| Market | Trades | Per trade |
| --- | --- | --- |
| Gold | 153 | -0.068R |
| USD/JPY | 137 | -0.046R |
| EUR/USD | 110 | -0.012R |
| AUD/USD | 131 | -0.006R |

So the whole result comes from the other four: USD/CHF, NZD/USD, GBP/USD and USD/CAD. With around 100 to 150 trades a market we would not read that as a list of pairs to keep. Picking markets by their record did not carry forward elsewhere in the study, and doing it here would be the same hindsight.

Gold deserves its own warning. On the plainer version with no level taken, it lost -0.570R a trade on 35 trades in 2025-26, the worst row in that table. [Our page on gold](/research/gold-strategies-tested/) sets that in context, in words.

## Does the result survive a stability test?

Only as a family, not as a setting. To see whether the rule was an isolated lucky spot, we ran 18,000 combinations of its settings over the whole record, each with its own Monte Carlo. This grid is a stability check on this one rule and nothing else.

- **Taken as a whole, the grid is a coin flip.** The median set lost -0.002R a trade.
- **One setting has a clear, steady effect.** The more crowded the speculators had to be, the better the result, step by step. That is the strongest thing in favour of the idea.
- **About half of what the best sets show is selection.** The 50 most profitable sets did nearly twice as well as their immediate neighbours in the grid.

Then the test that matters: choose a set using only the years up to a date, and judge it on the two years after. Five steps, 2017 to 2026.

| How the set was chosen | Out of sample, per trade |
| --- | --- |
| Most robust (best neighbourhood, steady years) | -0.038R |
| Best value of each setting on average | -0.014R |
| The whole grid | -0.010R |

The average of the 20 most robust sets came to about zero, and the robust picks lost in three of the five steps. The most profitable pick did better on average, and swung between large gains and a loss of 74R from one step to the next, on stops so narrow that they would not survive in an account. So choosing a set from the past did not reliably make money in the two years after. What the walk-forward says to expect going forward is somewhere between zero and a fraction of the record.

## Why might fading the crowd work at all?

We read it as stops, and that is a reading and not a result. A crowded side has its stops resting in a predictable place. A hunt into a prior level takes those stops, and when price [closes back](/glossary/the-close-back/) the crowd is left on the wrong side with fewer left to push price further. The [trap candle](/news-trading/trap-candle-forex/) is the same idea on one bar.

From the market maker's seat, as we read it, the crowd's stops are the liquidity. From the observer's, a story this tidy is exactly what should make you suspicious of a t-statistic of 1.8. The test shows an effect. It does not show this cause. No figure in the report is a stop order, and nobody outside can see who took the other side.

It is also a different trade from the one most people picture. [Our intraday setup, run as a rule](/research/mechanical-setup-backtest-2015-2026/), lost money in every period. This one holds for weeks through a wide stop, sits through deep drawdowns, and has several positions open at once. Few traders who like the first would enjoy the second.

## What does this test not prove?

It does not prove an edge. Nothing here has been traded live, and these are the limits.

- **It is not a forward result.** The clean period is four years and a thin pass. The rest of the record is what the rule was chosen on.
- **It does not prove the cost side.** Real swaps could make this better or worse, and more than anything else in the study.
- **It does not prove the markets.** Four of eight did the work, and we have no reason to expect the same four next.
- **It does not prove the setting.** The surface is smooth, which is good, but the walk-forward did not reward choosing a point on it.
- **It does not say a crowded reading is a reason to trade.** The plain version broke even.

The one stronger finding sits at the other end. Taking a setup the same way large speculators are already crowded lost clearly, at a t-statistic of -4.1, and as a rule to avoid trades that is the part we would lean on, with the gold exception [on the main COT page](/research/does-the-cot-report-work/).

More is on [the research page](/research/). Course 1 of ForexiaPRO, Reading the Map, teaches Fade the Crowd as a read of the report, and lesson 1.3.3 goes through this result and how thin it is.

## What changed when this page came back?

This page was taken down on 31 October 2026 and put back on 11 November 2026. Its figures were rerun on 11 November 2026 with scripts/seo-research-figures.mjs, cotswing.mjs and the 18,000-set grid. One figure changed: gold on the plainer version with no level taken in 2025-26 is -0.570R a trade on 35 trades, where the first version had -0.578R. The 1,047 trades, the 1.8 t-statistic, the years and the by-market figures are the same. The 18,000-set grid was rerun and its median set is still about -0.002R a trade. The walk-forward table and the neighbour comparison come from the study's earlier analysis of that grid, which needs a Python package this rerun did not have, so they are as first printed. No finding was reworded to read better than the numbers.

## Frequently asked questions

### Does fading the COT crowd work?

Trades taken with the crowd lost in 12 of 16 years; fading it on a level hunt made +0.08R a trade over 1,047 trades, promising and not proven. The t-statistic is 1.8, short of the usual bar of 2, the rule was chosen after seeing most of the data, and picking a version from the past did not reliably carry forward.

### Should I fade every COT extreme?

Our test does not support that. The plain version, every hunt against a crowded side with no level taken, came out at break-even. The version that held up best also needed the hunt to take a prior level, and even that one was at or below zero on four of the eight markets.

### Is this a trading system I can copy?

No. We report the test and do not publish its settings, and it has not been traded live. It holds trades for about two weeks with a wide stop and several positions open at once, and the financing cost in the test is an assumption. It is a result to keep watching and not a system.

### How is this different from the intraday setup?

It is a different trade. The intraday setup, run as a rule, lost money in every period we tested. This one uses a similar moment to enter but holds for weeks, because the COT effect builds over days and does nothing for a trade closed the same day.

## More from the blog

### [COT Report Release Date: When It Comes Out, With the 2026 Schedule](/cot-report/cot-report-release-schedule/)

The COT report release date is Friday 15:30 New York, positions counted the Tuesday before. The CFTC's 2026 dates, six holiday Mondays and the next four.

### [Does the COT Report Work for Forex? 16 Years Tested, Honestly](/research/does-the-cot-report-work/)

Does the COT report work for forex? On 16 years with costs, joining a crowded side lost, and the version that held best is not proven. Sample and limits.

### [COT Report CAD: The Canadian Dollar Turned Round, and a Crowd Long at Index 83](/cot-report/cot-report-cad/)

The COT report CAD counted 18 August 2026: Canadian dollar futures turned round to USD/CAD, the crowd at index 83, and the low that held for twelve days.