Bitfinex funding rates: what 3 years of data actually says
Most claims about Bitfinex lending yield are either a screenshot of one lucky week or a brochure number. We run a lending bot for a living, so we keep three years of hourly funding-rate candles in a database and replay strategies against them daily. This post is the summary we wish had existed when we started: the real distribution, the regimes, the spikes, and the one statistical property that decides how you should lend.
The dataset
Everything below comes from public Bitfinex funding candles — hourly open/high/low/close of the 2-day tenor rate for fUSD (US dollar funding), 2023 through mid-2026, roughly 26,000 hourly observations per currency. Anyone can pull the same series from the public API; our API guide shows how. Rates are quoted daily by Bitfinex; we annualize (×365) and state everything gross, before the exchange’s 15% fee on earned interest.
One framing rule for the whole post: this is a description of the past, not a promise about the future. Funding yield is regime-bound — it moves with leverage demand, and no historical average obligates next month to match it.
The distribution: calm most days, violent a few
APR band where calm-regime USD funding spent most of the period — the unglamorous baseline that compounds.
annualized rates printed repeatedly during demand surges — days, sometimes weeks, of elevated prints in hot regimes.
median duration of a USD rate spike in our data. Most flashes are gone before a human notices them.
gross APR band where our out-of-sample automated USD results cluster across the full period — spikes included, averaged down by calm months.
Regimes, not averages
The single most misleading thing you can do with this market is quote one average. The series is regime-bound: months of quiet 8–12% APR, then a stretch where leverage demand rises — a rally, a volatility event, a stablecoin scare — and the whole distribution shifts up for weeks before decaying back.
That structure has a practical consequence: your yield is mostly decided by how you handle the hot regimes, because the calm ones pay roughly the same no matter what you do. A lender who captures spikes at 25–30% APR for a few weeks a year can move their annual number by multiple points; a lender who always lends at the going rate simply gets the base.
You can see the regime texture month by month in our monthly funding-rate reports — average, high, low, and spike-day counts per currency, regenerated from the same candle data.
Spikes are short. Really short.
The median USD spike in our data lasts about an hour. That number should reorganize how you think about this market: the highest-paying moments are not events you react to, they are events you must already be positioned for. By the time a spike is visible on a chart, the loans it paid for are usually already written.
This is why "check the rate a few times a day and adjust" underperforms: manual lending structurally misses the right tail. The fix is mechanical, not attentive — keep an offer resting above the market at all times, priced where spikes historically reach, and let the spike come to you.
The one property that matters: mean reversion
Funding rates mean-revert, strongly. In our out-of-sample tests, simple mean-reversion models explain roughly R² 0.29–0.46 of rate variation at one-to-seven-day horizons — remarkably high for a financial series. High rates decay toward the long-run level; depressed rates recover. The current rate tells you a lot about next week’s rate.
For a lender, the actionable translation is a relative floor: refuse to lend below a high percentile of the recent rate window. When the rate is unusually high for its regime, lock it in for a longer term (reversion says it will fade); when it is unusually low, stay short and liquid (reversion says it will recover, and you don’t want capital stuck at the bottom).
We tested dozens of external signals — order books, open interest, perpetual funding, liquidations — hunting for something that predicts rates better than the rate’s own history. Almost everything is either redundant with mean reversion or dead out-of-sample. The boring conclusion survived three years of adversarial testing: the rate series itself, plus patience, is the edge.
What this means for how you lend
Four rules fall straight out of the data:
- Always be resting an offer above market. Spikes are too short to chase; a standing limit offer is the only reliable way to be filled at the tail.
- Price your floor relative to the regime, not in absolute terms. A fixed "never below 10%" rule idles you through entire calm years; a percentile floor adapts.
- Extend the term when rates are unusually high, shorten when low. Mean reversion makes tenor choice the second-biggest lever after the floor.
- Judge results in years, not weeks. Most of the annual yield arrives in a few hot stretches; any single month is noise.
Methodology and caveats
Source: public Bitfinex candles (trade:1h and trade:1D, 2-day tenor) for fUSD, 2023–mid-2026; analysis run in the same backtest engine that powers our calculator. Backtests are optimistic by construction — they assume fills at the bar and no market impact — which is why we quote out-of-sample bands rather than best-case numbers, and why every figure here is gross of the 15% fee. Past rates do not predict future rates; treat everything above as measured history.
Questions this data answers
What does Bitfinex USD lending realistically pay?
Over 2023–2026, calm regimes paid roughly 8–12% APR gross and automated out-of-sample results cluster around 8.5–10% gross (≈7.2–8.5% after the 15% fee). Hot regimes pay materially more for weeks at a time. No number here is a guarantee — yield follows leverage demand.
How long do Bitfinex funding-rate spikes last?
The median USD spike in our three-year sample lasts about an hour. A minority run for days during genuine demand regimes. The practical consequence: you catch spikes with standing offers, not by reacting.
Do funding rates mean-revert?
Yes, strongly: out-of-sample R² of roughly 0.29–0.46 at 1–7 day horizons in our tests. Unusually high rates tend to decay and unusually low rates tend to recover, which is what makes percentile floors and rate-conditional term selection work.
Can I reproduce this analysis?
Yes — the candle data is public and free. Pull trade:1D or trade:1h candles for fUSD from the Bitfinex public API (our API guide has working Python), and check the monthly report pages on this site for the per-month numbers we publish from the same series.
The data is the strategy
Every Stratum strategy is a rule set tested against this exact series — percentile floors, rate-conditional tenors, spike capture. Model one on three years of history, or read the monthly numbers yourself.