Method version 1.0, signed off by both founders on 15 October 2026. Any change will be a new version with its own date; old readings are never rewritten.

1. What goes in

• CFTC Commitments of TradersLegacy report, futures only: long and short positions of "non-reportable" traders, everyone below the CFTC's reporting size. Every Friday, for positions as of the Tuesday before. Source: publicreporting.cftc.gov.
• Exchange account ratiosBybit and Binance: share of all accounts holding a long. Collected daily; they join after 180 days of our own collection. None qualify yet. We publish only our derived readings, not their raw figures.
• Our daily pollShown, weight 0 until it has 200 or more answers a day for 30 days.
• Broker client sentimentMyfxbook, IG, OANDA: weight 0 until those firms give permission.

2. The Crowd Gauge

• Long shareLong ÷ (long + short) × 100, for each source and market.
• Release date, not position dateA COT figure counts from its Friday release (position date + 3 days), so history never uses data before the public could see it.
• PercentileWhere today's long share sits among the last 156 weekly readings (1,095 days for daily sources). Ties count half. A source needs 52 weekly or 180 daily readings first. This removes each market's normal lean.
• Market readingThe average of that market's sources.
• Futures and cryptoFutures: E-mini S&P 500, Nasdaq-100, gold, WTI crude, euro and yen. Crypto: bitcoin and ether (solana joins once exchange data qualifies).
• Composite50% futures + 50% crypto, fixed in advance, never tuned to the backtest. If a half is missing, the other counts alone; under half the weight, nothing is published.
• Stale dataA COT figure is used for up to 10 days, a daily figure for up to 2. Then the source drops out.

Crowded short

0–20

Leaning short

20–40

Balanced

40–60

Leaning long

60–80

Crowded long

80–100

3. Crowd Fear & Greed

Crowd only: no prices, volatility or news.

• Half: how longThe Gauge percentile for each source.
• Half: how fastThe percentile of the 4-week change in long share.
• AveragingA market averages its sources. A scope averages its markets: crypto (bitcoin, ether), bitcoin, stock market (S&P 500, Nasdaq-100, Russell 2000).

Extreme fear

0–25

Fear

25–45

Neutral

45–55

Greed

55–75

Extreme greed

75–100

Not the mood-based kind. CNN Business (stocks) and alternative.me (crypto) mix price, volatility and survey inputs. See Crowd Fear & Greed.

4. Backtest: did extreme readings come before unusual returns?

Result, under the rule we fixed before running it: crowded long, an edge (passes the rule fixed in advance); crowded short, no reliable edge. In plain words: after crowded long (80+) readings, prices did better than usual over the next 4 weeks (+1.0 percentage points on average, 2018 onwards), so the crowd was on the right side on average, the opposite of "fade the crowd"; after crowded short (20 or less) readings: no reliable difference from normal. The effect is small next to normal price swings, much of it comes from the most volatile markets (see the table by market and the extra check below), and a pattern in the past may not repeat.

The rule, written before the test. We call it an edge only if, pooled across markets, the 4-week return after extreme readings differs from normal with a t-statistic of at least 2 in 2018 onwards, and in the same direction in 2006–2017.

A negative number after crowded-long readings means prices did worse than usual (what "fade the crowd" expects). t-statistics use Newey-West standard errors because 4- and 12-week windows overlap. Readings that overlap in time are not independent, so the real evidence is thinner than the counts suggest.

Extra check: crypto left out (not pre-registered)

We added this after seeing the result, because bitcoin moves far more than currencies and can dominate a pooled average. It is not part of the rule above. 2018 onwards, 4 weeks, 5,438 readings: after 80+ 0.70 points (t 1.97); after 20 or less 1.04 points (t 1.58). Pooling also treats markets as if they moved independently, which overstates the t-statistics somewhat.

By market, 2018 to latest, 4 weeks

Not tested: Russell 2000 (no free daily price series we can use); Gold (no free daily price series we can use); Silver (no free daily price series we can use); Copper (no free daily price series we can use); US dollar index (no free daily price series we can use); Solana (no free daily price series we can use). Prices: daily series from FRED, Federal Reserve Bank of St. Louis, used only for this test and not republished. Returns run from the first price on or after the day after each release. Full results: crowd-gauge-backtest.json. Test run on data to 9 October 2026.

5. Limits

• "Retail" is a proxyThe CFTC's non-reportable group is every account under its reporting size, which includes some small firms. Exchange account ratios count accounts, not money.
• Weekly, with a lagCOT positions are three days old when published. Until exchange data qualifies, readings change once a week.
• Futures onlySpot forex and CFD retail positioning is held by brokers; we use it only with their permission.
• A percentile is relative90 means "more long than in 90% of the last three years", not "90% of traders are long".

Questions or a mistake? Tell us. Corrections are logged on the corrections page.