# Lending-rate spikes do not pay a short after costs

> A short of the coin for 24 hours after a spike in its spot-margin lending rate earned +62.2 bp before costs and +3.4 bp after them (interval -33.7 to +37.0 bp), against a required +30 bp.

E871 · Verdict: KILL · Cause: Below costs · Market: Crypto & DeFi · Mechanism: Forced flow · Horizon: 4h-24h · Tested: October 2026

Source: https://alphanotfound.eu/experiments/E871-lending-rate-spikes-do-not-pay-a-short

Some holders cannot or will not short a coin directly, so they borrow it on spot margin and sell. People argue that a sudden jump in the lending rate exposes this demand and that the coin then underperforms. We found 1,383 such jumps in 123 coins and shorted each coin's perpetual for 24 hours, hedged against Bitcoin. Before costs the short made +62.2 bp per trade. Costs (gross minus net, funding included, about 59 bp) left +3.4 bp, with an interval from -33.7 to +37.0 bp that includes zero. The test required +30 bp.

## Key numbers

- **Spike events:** 1,383 (123 coins, 2024-06 to 2026-10)
- **Net mean per trade:** +3.4 bp (95% interval -33.7 to +37.0 bp; +30 bp required)
- **Gross vs costs:** +62.2 vs ~59 bp (gross minus net, funding included)
- **Worst single trade:** -6,425 bp (net, after costs)

## The idea being tested

**Belief:** A sudden spike in a coin's spot-margin lending rate reveals new short demand, so the coin underperforms over the next day.

**Why people believe it:** People argue that holders who cannot or will not short the perpetual (locked or vesting tokens, airdrop recipients, coins whose funding is capped) borrow the coin on spot margin and sell it. The lending rate clears every hour, while the passive supply of lendable coins adjusts slowly, so a spike should show sudden demand to sell or hedge before the price has moved.

**Why it would persist:** Most affected coins are small and expensive to trade, and the lending rate is not shown in ordinary price feeds, so few traders would watch it.

## How it was tested

We downloaded the hourly lending rate of each coin and flagged the hours in which it jumped far above its own 30-day norm and to at least 20% a year. Two hours after each flagged hour (to be sure the rate was already known) we simulated a short of that coin's perpetual for 24 hours, hedged against Bitcoin, and charged realistic costs. We also ran the same short one week earlier and in ordinary hours of the same coin, to see what shorting these coins earns in general.

**Pre-registered:** yes

**Data:** Hourly market lending rate (annualised) of spot-margin lending, by coin; no amounts supplied or borrowed (a large crypto derivatives venue, public API (anonymous); 1 hour; 2024-01..2026-10); Perpetual futures 1-hour closes and funding rates, used as a proxy for the lending venue's own perpetual (largest USD-margined crypto perpetual venue, public market-data archive; 1 hour; 2024-01..2026-10)

**Universe:** Coins that have a lending-rate history and a USDT-margined perpetual on both venues (Bitcoin excluded from the events; used as the hedge). 123 coins produced at least one event.

**Signal:** Event hour t when z = (ln rate - 30-day median of ln rate) / (1.4826 x 30-day MAD of ln rate) is at least 3, the rate is at least 2 times its own 30-day median, and the rate is at least 20% annualised. One event per coin per 72 hours.

**Execution:** Short the coin's perpetual at the 1-hour close two hours after the event hour (the rate stamp may be the hour start, so the lag removes look-ahead). Hold 24 hours; a 4-hour hold and the long side are reported. Hedge with Bitcoin, beta from the prior 30 days of hourly returns, clipped to 0 to 2.

**Costs:** Per-coin cost model: 5 bp fee plus a modelled half-spread and impact at a $5,000 trade, charged on both sides, plus 5.5 bp per side on the Bitcoin hedge, scaled by beta. Funding received by the short is added. Prices and funding come from the other venue and proxy the lending venue's perpetual.

**Controls:** The same coin and clock hour 7 days earlier; matched non-spike hours (same coin, |z| below 1, same hour of day, 3 per event); long direction to show the sign; a descriptive daily cross-sectional rank correlation of z with the next 24-hour return.

**Sample split:** Halves at the median event date.

**Pass criterion:** At least 200 events from at least 40 coins; mean net at least +30 bp; coin-clustered 95% interval lower bound above 0; median above 0; both halves above 0; controls below half the event mean.

**Kill criterion:** Mean net below +10 bp, or the interval includes 0.

**Power:** Assumptions made before the test: hedged 24-hour standard deviation about 700 bp, about 600 effective events, giving a minimum detectable effect of about 80 bp, so a modest true effect would not have been detectable. The realised single-trade standard deviation was about 600 bp.

## Result

**Why it was rejected:** Mean net of +3.4 bp is below the +10 bp kill line and the coin-clustered interval (-33.7 to +37.0 bp) includes 0.

**Observations:** 1,383

**Net mean:** +3.4 bp

**95% interval:** -33.7 bp to +37 bp

**Median:** +38.5 bp

**Worst:** -6,425 bp

- Signed so that + means the short made money. Over 24 hours the hedged short earned +62.2 bp per trade gross and +3.4 bp net after costs of about 59 bp (gross minus net, funding included).
- The coin-clustered 95% interval for the net mean is -33.7 to +37.0 bp. The median trade earned +38.5 bp and 54.6% of trades were profitable, but the worst trade lost -6,425 bp; the worst 14 trades (1%) average -3,285 bp. The net interval includes zero, and the few worst trades outweigh the typical one.
- By period the net mean was not stable: -37.5 bp in the first half and +44.2 bp in the second (691 and 692 events), so the halves disagree and neither is evidence; by year -6.1 bp in 2024 (40 events), -22.6 bp in 2025 (326) and +12.1 bp in 2026 (1,017).
- A 4-hour hold earned -27.7 bp net, and the long side over 24 hours -88.3 bp.
- The pre-registered control condition (the 7-day-earlier and 3-per-event matched controls below half the event mean) did not discriminate, because both controls were negative in the same window.
- The daily cross-sectional rank correlation between the rate z-score and the next 24-hour return was -0.006 (t = -0.8).

### Net mean sits between the kill line and the pass line (bp)

Mean net return of the 24-hour hedged short after a lending-rate spike, with its 95% coin-clustered interval. The mean (+3.4 bp) is below the +10 bp kill line and the +30 bp pass line, and the interval runs from -33.7 to +37.0 bp, so zero is inside it.

|   | Mean | 95% low | 95% high |
| --- | --- | --- | --- |
| Short after a spike, net (1,383 events) | 3.4 | -33.7 | 37.0 |

### The net mean changes sign across periods (bp)

Net mean per trade by year and by half of the sample (events in brackets). Halves are -37.5 and +44.2 bp, years -6.1, -22.6 and +12.1 bp: no consistent sign.

|   | Short after a spike, net |
| --- | --- |
| 2024 (40) | -6.1 |
| 2025 (326) | -22.6 |
| 2026 (1,017) | 12.1 |
| First half (691) | -37.5 |
| Second half (692) | 44.2 |

### A positive median, a mean pulled down by the tail (number of trades)

Number of 24-hour short trades by net return in bp. The median trade earned +38.5 bp, but the long left tail (the worst trade lost -6,425 bp) pulls the mean to +3.4 bp.

| Bin | Short after a spike, net (count) |
| --- | --- |
| < -800 | 76 |
| -800..-400 | 116 |
| -400..-200 | 150 |
| -200..-100 | 136 |
| -100..-50 | 63 |
| -50..0 | 87 |
| 0..50 | 82 |
| 50..100 | 78 |
| 100..200 | 150 |
| 200..400 | 211 |
| 400..800 | 176 |
| > 800 | 58 |

## Interpretation

**Why it failed (interpretation, not established):** The gross return of +62.2 bp is almost the same size as the costs of about 59 bp, because most of the coins are small and expensive to trade, so the net is close to zero. The spread of single trades is also large (single-trade standard deviation about 600 bp), so the interval is wide and the sign of the net mean flipped between periods; this sample could not have detected an effect much below 80 bp.

**Lesson:** When gross return and round-trip cost are the same size, the net result is decided by the cost model and is easily moved by one or two periods. An average over events should be read together with its cost, its median against its mean, and its behaviour across periods: here the median was +38.5 bp while the mean was +3.4 bp, and the periods disagreed on sign.

**Scope:** Covers the hourly market lending rate of one large crypto derivatives venue, 123 coins with a perpetual on two venues, one spike definition, a 2-hour lag, a 24-hour hold and a Bitcoin-beta hedge, 2024-06 to 2026-10 for the events. Prices and funding are from a different venue than the lending rates, and the lending data carry no amounts.

## Reproducibility

**Steps:** 1) Download the hourly market lending rate for every coin with a perpetual on both venues, from 2024-01. 2) Compute z = (ln rate - 30-day median) / (1.4826 x 30-day MAD) and flag hours with z at least 3, rate at least 2x its 30-day median and at least 20% annualised; keep one event per coin per 72 hours. 3) Short the perpetual at the 1-hour close 2 hours after the event hour and hold 24 hours, hedged with a 30-day Bitcoin beta clipped to 0 to 2. 4) Subtract 2 x the per-coin cost (5 bp fee plus modelled half-spread and impact at $5,000) and 2 x beta x 5.5 bp, add funding received. 5) Bootstrap the mean by coin; run the 7-days-earlier and matched quiet-hour controls.

**Grade:** COMPLETE

---
AI-generated research, checked but not peer-reviewed. Historical, experimental research, not investment advice. Dataset Terms of Use: https://alphanotfound.eu/terms