Charlie Quant LabFunding Rates Predict Their Own Funeral
Perpetual-futures funding is the bill crowded positioning pays to stay open. This page measures how crowded Bitcoin and Ethereum perps are right now, what prices actually did after past crowding at each level, how many days an extreme reading survives, and whether funding and open interest feed each other. Raw snapshots are archived nightly from 2026-09-08.
Updated 08 September 2026 · 20:37 UTCA year of the bill
Daily mean funding for the two deepest perpetual books on one large venue, annualized for readability. Spikes are crowded positioning paying rent. The reflexive claim being tested: the rent itself helps end the crowding.
What prices did after each crowding level
Every scored day (2025-09-04 → 2026-09-08, 280 days) was bucketed by its funding z-score, then matched with the Bitcoin price 3 and 7 days later. Median moves shown; medians resist the few violent days that dominate averages.
| Funding z-score bucket | Days | Median move, next 3d | Median move, next 7d |
|---|---|---|---|
| z ≤ −1 | 42 | +0.6% | +0.0% |
| −1 to 0 | 78 | +0.3% | +1.1% |
| 0 to +1 | 101 | -0.1% | -0.2% |
| +1 to +2 | 49 | -0.6% | -0.0% |
| z ≥ +2 | 7 | -1.0% | -7.5% |
How fast extreme funding dies
Extreme episodes are runs where |z| reached 2 or more. The decay path averages every episode's funding after its peak day, as a share of that peak. The dashed line is half — the half-life.
Does funding drag open interest with it?
If funding is a crowding tax, expensive funding should walk with swelling open interest — and expensive funding today should say something about open interest tomorrow. Both correlations use daily funding against daily open-interest change on the same venue (31 days of open-interest history; the venue publishes a shorter OI record than its funding record).
| Episode peak | Peak funding (annualized) | Half-life |
|---|---|---|
| 2026-07-04longs paying, 1d at extreme | 11% | 4d |
| 2026-05-27longs paying, 3d at extreme | 11% | 3d |
| 2026-02-06shorts paying, 3d at extreme | 10% | 2d |
| 2026-04-19shorts paying, 1d at extreme | 10% | 1d |
| 2026-06-02longs paying, 1d at extreme | 9% | 2d |
| 2026-02-10shorts paying, 1d at extreme | 8% | 1d |
| 2026-03-11shorts paying, 2d at extreme | 8% | 2d |
| 2026-05-24longs paying, 1d at extreme | 8% | 11d |
Tonight's crowding
Bitcoin perpetual funding sits at z = +0.7 — no extreme crowding. Across 280 scored days on this venue, readings this crowded on the long side (z ≥ +2, 7 days) preceded a median -7.5% move over the next 7 days, while negative-funding days (z ≤ −1) preceded a median +0.6% over 3 days. Extreme readings took a median of 2 days to lose half their size. The bill for staying crowded is currently paid by longs.
Method, scope and what would make this wrong
What is measured. Every funding print for BTCUSDT and ETHUSDT perpetuals on one large derivatives venue over the past year, aggregated to a daily mean rate. Annualized figures multiply that daily mean per-8h rate by 3 funding prints per day and 365 days per year (×1,095). The z-score compares today's rate with the trailing 90 days of daily means. Forward moves use the venue's own daily closes for the same contract, so the funding and the price come from one tape.
Honest edges. This is one venue's positioning, not the market's — funding elsewhere can differ for hours or days. The open-interest history the venue publishes is shorter than the funding record, so the co-movement numbers rest on fewer observations than the return buckets. Forward-return buckets are counts of what happened, not a distribution of what will. "Cost if today's rate lasted a year" is a display convention; funding reprices every 8 hours and never actually lasts a year.
Archive. Each export saves the raw snapshot — current marks, the last 30 funding prints and the last 30 open-interest readings per symbol — into a dated nightly archive that began 2026-09-08. As the archive grows, future editions will test tonight's readings against their own realized outcomes.
Missing stays missing. Prints absent from the venue's record are absent here. If a feed fails at export time the affected figure is marked not available below, never interpolated.