Negative funding is supposed to mark the bottom. We counted.

Every few weeks the same argument restarts. Funding goes deeply negative, someone posts the chart, and the replies split into two camps: “crowded shorts, this is the bottom” and “the market is telling you something, get out of the way.” Both camps cite the same three memories: March 2020, mid-2021, and the FTX week. Three anecdotes, argued forever, resolved never.

So we counted.

What we measured

The claim needs to be exact before it can be tested. Ours was the sharpest version of the pay-to-be-short story: the funding rate at -1% or worse, within 30 minutes of the settlement where shorts actually pay it. Not “funding has been negative lately.” The moment of maximum stated crowding, right before it gets settled.

We ran that condition over every recorded market-minute across 660+ Binance USDT-M crypto and TradFi perpetuals from 2026-05-15 to 2026-07-26: 47,223,625 eligible market-minutes across 677 markets. The condition was true in 2,357 of them, which group into 105 separate episodes. That rarity is the first honest finding: the state people argue about as though it happens weekly is a hundred-episode event in forty-seven million minutes.

What followed

Forward outcomes are measured from each episode, over all of them. Of the 105 episodes, 103 had a complete 24 hours of record after them:

• 82 of 103 (79.6%) fell 2% or more over the next 24 hours. Across every eligible minute where the condition was false, in the same window, that rate is 30.2%.
• Only 17 of 103 (16.5%) rose 2% or more in the same 24 hours, against a background rate of 25.2%.
• Drawdown was near-universal: 103 of 105 touched -2% at some point inside the 24 hours, against a background rate of 62.5%.

Read that from the folklore’s point of view. At the exact moment the “crowded shorts, squeeze incoming” story is loudest, the recorded market kept falling, at more than twice its ordinary rate, and rose less often than usual. Extreme negative funding into settlement did not mark bottoms in this record. It marked more falling.

Before you trade this, read this

This is the part most posts leave out, and it is the part that matters.

This is not the same claim as the one in the headlines. The articles arguing “most negative funding since 2023, history says bottom” are talking about funding staying mildly negative for weeks or months. We measured minutes of extreme funding right before settlement. Related folklore, different animal. One does not refute the other; ours is simply the version precise enough to count.

105 episodes is thin. The pattern is strong in this record, and this record is ten weeks long. A different regime could read differently, and the scan that produced this was one of 26 hypotheses swept the same week, which means some spread between them is expected by chance alone. Our own sweep labels this a heuristic flag, not significance. We are publishing the count, not a verdict.

The two sides of that comparison are not the same kind of thing, and our engine says so. The 79.6% counts 103 deduplicated episodes. The 30.2% counts every eligible market-minute where the condition was false, and those minutes overlap heavily: neighbouring minutes share almost all of their forward path, so they are nowhere near independent samples. The comparison is the honest one available, and it is still a comparison between an episode and a minute. Treat the gap as large and real in direction, not as a precise effect size.

Falling markets produce this state. Funding gets this negative because price has been dumping. “It kept falling” is partly the momentum of the same move that created the reading. The count is real; the mechanism is not settled by it.

These are counts of what happened in the scanned markets and dates, not a promise it happens again. Nothing here is trading advice.

Check us

The condition, the window, the episode count and the outcome distributions are pinned in the linked report below under q#945e12a2. The background rates come from a cohort comparison of the same document against every eligible minute where the condition was false, at the same dataset revision, under q#e21e17e8. Both keys re-run to the same bytes on any account, on any day. Anyone who thinks we are wrong can check both.

One thing we did while writing this, which is the point of building it this way: an earlier draft of this post carried 81 of 102 and a 30.5% background rate, taken from the sweep that first flagged the pattern. Re-running against the published report moved it to 82 of 103 and 30.2%, because the record grew overnight and one more episode completed its 24 hours. The receipt corrected the draft. That is the standard we think market claims should meet, and it is the entire reason this blog exists: the argument is here, the receipt is one click away.

If you hold the opposite view, we mean it: run the search, slice it differently, and publish your ID. A disagreement between two re-runnable results is progress. A disagreement between two screenshots is Tuesday.