Does high funding predict crypto crashes? 8 coins, 2020–2026

In 2023 three economists at the Bank for International Settlements described crypto carry — the premium of crypto futures over spot — as large and volatile: up to about 60% a year, pushed up by small investors who want leveraged upside in a boom, and kept there because too little arbitrage capital is willing to take the other side. Their abstract adds that high carry predicts future price crashes.¹

On a Binance perpetual, carry is the funding rate: when it is positive, longs pay shorts every 8 hours. That makes the claim testable directly — and it suggests three things a trader might do with it: expect a crash, short it, or step aside. This study checks all three on eight coins, and then looks at what the carry itself cost someone who simply held a perpetual long.

Data and method

• Market: Binance USDT-M perpetuals BTC, ETH, SOL, XRP, DOGE, BNB, ADA and LINK, 4h candles, 2020-01-01 → 2026-10-01. The funding rate is the last settled 8-hour rate.
• "High carry": funding above 0.05% per 8 hours (about 55% a year), or a funding z-score above 2 — today's rate measured against its own last 90 days (ZSCORE_540@FUNDING on 4h bars). Both were fixed before the first run.
• Crashes: from the platform's candles and funding history: for every 4h bar, did the low of the next 30 days come 20% or more below that bar's close, and what was the 30-day return? Bars with high carry against all other bars.
• Backtests: 1x leverage, cross margin, 95% of the account, fills at the next open, 0.05% fee and 0.02% slippage per side, funding charged and paid. Every perpetual rule waits for 90 days of funding history, so each starts on the same bar as its benchmarks (BTC and ETH on 2020-03-31; the others between April and December 2020, after their perpetual listed).
• How much was tried: three rules for a long (two thresholds and the z-score, plus a basis z-score on BTC and ETH), three shorts, two benchmarks, each on 8 coins over the whole period and both halves of it. Nothing was tuned.

Crashes came more often — after bigger gains

The paper's claim shows up in the data: after funding above 0.05%, a fall of 20% or more within 30 days was more common than on other bars on 7 of 8 coins — on BTC 32.9% against 15.8%. Only SOL had fewer. With the z-score instead of the fixed threshold the effect was weaker: more crashes on 4 of 8, all of them altcoins.

But the same bars were followed by better 30-day returns, not worse: the median 30-day return after funding above 0.05% was higher than usual on 7 of 8 coins (XRP was the exception). High carry came with wider swings in both directions — the crash, when it came, tended to come after the price had gone further up first. A rule that bets on the crash has to survive that first leg.

Bars overlap: a hot-funding spell lasts many 4h bars, so 550 BTC bars are 79 separate episodes — and 60–77% of each coin's hot bars fall in 2021, so this is largely a description of how one boom ended.

Shorting high carry did not pay

A 7-day short whenever the funding z-score went above 2 lost money on all 8 coins, in both halves of the period, and was liquidated twice on XRP. It collected funding while it waited, but far less than it lost on price. The same short on a basis z-score above 2 with rising open interest (BTC and ETH only, dry runs) lost on both. Adding a second condition — open interest up more than 10% over the week, a crowded long — left one coin in profit (BNB, 14 trades). This agrees with the earlier test of funding as a signal, where a short after funding above 0.05% lost on 18 of 20 coins.

Backtest

Short 7 days when the 90-day funding z-score is above 2

BTCUSDT · 4h · Futures · 2020-01-01 → 2026-10-01

Return

-69.52%

Max drawdown

78.21%

Sharpe

-0.67

Win rate

41.0%

Trades

39

Open the report →

Stepping aside while carry is high

The softer use of the claim is to keep a long and step aside while carry is high. The fair benchmark is the same perpetual long without the filter, closed and re-opened at the start of every month so its size follows the account — why that matters is the next section.

Stepping aside while the z-score was above 2 had a higher Sharpe ratio than the benchmark on 5 of 8 coins and a smaller drawdown on 6 — but in 2023–2026 alone the Sharpe ratio was higher on only 3 of 8 (dry runs). The 0.05% filter beat the benchmark's Sharpe on 3 of 8, and so did the 0.03% filter (dry runs); a basis z-score filter on BTC and ETH beat it on one of the two. Drawdowns stayed above 63% everywhere: stepping aside did not avoid the deep falls. On BTC both filters ended ahead of the benchmark, on SOL both well behind (+3,973.38% and +662.92% against +7,421.50%).

Backtest

Perpetual long 1x, flat while funding is above 0.05%

BTCUSDT · 4h · Futures · 2020-01-01 → 2026-10-01

Return

+952.30%

Max drawdown

75.88%

Sharpe

0.95

Win rate

65.0%

Trades

80

Open the report →

What the carry itself cost

Here the paper's first finding — that carry is large — shows its full weight. A 1x perpetual long that was never closed was liquidated on 7 of 8 coins. BTC's was opened on 2020-03-31 at 6,390.80 and paid 27,123.50 USDT of funding on that one trade — almost three times the 10,000 USDT account — before it was liquidated on 2022-11-08 at 17,978.88, well above the price it was bought at. The position's size was fixed while the price, and with it the funding bill, grew many times over; the account's cash went negative and the price only had to fall back for the margin to run out. Only BNB, whose long was paid funding on balance, survived.

Backtest

Perpetual long 1x, never closed (funding paid)

BTCUSDT · 4h · Futures · 2020-01-01 → 2026-10-01

Return

-97.09%

Max drawdown

99.91%

Sharpe

0.56

Win rate

50.0%

Trades

2

Open the report →

Re-sizing the long every month avoids the liquidation, but not the bill: it still paid 30,120 USDT of funding on BTC, and simply holding BTC on spot from the same bar ended at +1,196.03% against +418.38%. Spot beat the perpetual long on all 8 coins.

Backtest

Buy & hold benchmark (spot 4h)

BTCUSDT · 4h · Spot · 2020-03-30 → 2026-10-01

Return

+1,196.03%

Max drawdown

76.98%

Sharpe

0.97

Win rate

100.0%

Trades

1

Open the report →

Backtest

Perpetual long 1x, re-sized monthly (funding paid)

BTCUSDT · 4h · Futures · 2020-01-01 → 2026-10-01

Return

+418.38%

Max drawdown

78.23%

Sharpe

0.73

Win rate

54.4%

Trades

79

Open the report →

Conclusion

• High carry did precede crashes more often on these eight perpetuals — but it also preceded higher 30-day returns, so on its own it was not a timing signal.
• Shorting it lost money on every coin, and stepping aside while it was high helped on some coins and hurt on others.
• The cost that was certain was the carry itself: a 1x perpetual long paid more in funding than the account was worth and was liquidated on 7 of 8 coins; even re-sized monthly it trailed spot buy & hold on all 8.

More on funding: funding as a signal and the hour before a funding settlement. How to read the cards: backtest metrics explained.

Notes

1. Maik Schmeling, Andreas Schrimpf and Karamfil Todorov, "Crypto carry", BIS Working Papers No 1087, April 2023; the abstract was read on 2026-10-23. ↩