I Backtested 5.5 Years of Bitfinex Funding — Real Numbers Across 58 OOS Windows

“What’s the average Bitfinex Funding APY?” — there’s no single answer. Rates change daily, regimes shift yearly, and different strategies produce vastly different results.

I ran a walk-forward backtest using Bitfinex’s 5.5 years of public candle data — to find out what the same strategy would have produced across all that history.

This post is the full results: median APY, p5/p95, regime breakdown. All numbers are net (after Bitfinex’s 15% fee), produced by scripts/realistic_backtest.py against data/history/.

TL;DR

• fUSD balanced median net APY: 15.7% (p5–p95: 6.7%–27.3%, Sharpe 3.9)
• fUST balanced median net APY: 16.0% (p5–p95: 9.9%–22.9%, Sharpe 6.1)
• 58 OOS rolling 90-day windows across calm / bull / bear / storm regimes
• Includes realistic fill model (queue rank · time decay · depth gate · spike filter) + prepayment hazard
• Marketing pages quote a deliberately lower band (fUSD 12-15%, fUST 11-13%) — see “conservative estimate” section below

Why Walk-Forward Instead of Single Backtest

99% of “I earn X% APY” claims use a single backtest:

1. Grab a slice of history (e.g., all of 2024)
2. Run strategy
3. Report final return

The problem: massive overfit. If you tune parameters to perform best in 2024, 2025 will likely crash.

Walk-forward instead:

1. Split data into “train + test” windows (180d train / 90d test)
2. Calibrate parameters on training window
3. Evaluate on the unseen test window
4. Roll the test window forward by 30 days, repeat
5. Final result: 58 independent OOS evaluations

Each OOS result represents the strategy “not knowing the future” — the realistic condition. The median + distribution of 58 results = real expected performance.

One Chart — Result Distribution

58 OOS net-APY values for fUSD vs fUST balanced preset (λ=1.0 calibrated baseline):

Observations:

• Both medians cluster near 16% with overlapping distributions
• fUSD has wider dispersion — more upside tail (>25% APY) and more downside tail (<8% APY)
• fUST is tighter — narrower band, higher Sharpe, fewer extreme months
• The “stablecoin yield = 5%” intuition is far below the median; the “>30% APY” headlines are real but rare (>p95)

Full Preset Comparison

Symbol	Preset	Median APY	p5	p95	Sharpe
fUSD	safe	14.3%	4.2%	21.1%	4.4
fUSD	balanced	15.7%	6.7%	27.3%	3.9
fUSD	aggressive	15.3%	6.8%	26.4%	3.5
fUST	safe	17.8%	6.5%	34.5%	5.7
fUST	balanced	16.0%	9.9%	22.9%	6.1
fUST	aggressive	16.8%	9.6%	24.4%	6.1

(Sharpe = mean / std × √N. Higher Sharpe = same return, less month-to-month noise.)

Two unexpected findings:

1. fUSD aggressive does not beat balanced (15.3% vs 15.7%). The more aggressive floors push too far out of the actual fill zone — you wait for spikes that don’t arrive often enough.
2. fUST safe has higher median than balanced (17.8% vs 16.0%) but with a much wider p5/p95 band. The safer floors fill faster and rarely miss; the trade-off is more variance.

Production currently ships balanced for both symbols.

Prepayment-Hazard Sensitivity Band

The model has one key uncertain parameter — borrower prepayment intensity (λ₀). We grid it across 0.5×, 1×, 2×, 4× the calibrated baseline:

λ scale	fUSD median	fUST median
0.5×	15.7%	16.3%
1.0× (calibrated)	15.7%	16.0%
2.0×	15.4%	16.3%
4.0×	15.7%	16.4%

The strategy is robust to prepayment uncertainty — even at 4× the calibrated rate, median APY moves <50 bps. (See Prepayment hazard — the hidden cost of long periods for why this matters.)

Why Marketing Shows a Lower Range

Our marketing pages quote “fUSD 12-15%”, “fUST 11-13%” — below the backtest medians above.

Why:

1. Backtest ≠ reality: 5.5y history doesn’t guarantee future
2. Survivorship bias remains: queue rank is approximated from funding_stats deltas, not real order-book replay
3. Inverse over-promise: under-promise and over-deliver

If live performance hits the marketing range, we win. If it hits the backtest median, that’s a bonus.

Cross-Regime Performance — Storm Is Worst, Not Best

Walk-forward windows span 4 regimes (BTC 30d vol percentile):

• Calm: vol < 25th percentile
• Bull: vol 25-75pct, 30d return > 0
• Bear: vol 25-75pct, 30d return ≤ 0
• Storm: vol > 75th percentile

Regime	fUSD median APY	fUST median APY	n anchors
Calm	17.3%	17.3%	34/38
Bull	15.6%	15.8%	49/52
Bear	13.4%	18.2%	48/49
Storm	8.0%	9.6%	26/32

This was the most surprising finding — and it inverts a common assumption.

The intuition (“storm = high vol = high funding rates = big yield”) is half-right: spike rates do appear. But spikes are short, and the prepayment hazard model shows borrowers deleverage during the storm, closing positions early. The bot grabs a 25% rate on day 1, the borrower returns on day 4 instead of holding 30 days, and the realised APY collapses.

Calm is where the work happens — steady fills, full hold periods, low cancellations. Stable rates the model can sit on.

Honest Caveats

A few things to be straight about:

1. 5.5y data ≠ eternity: includes 2020-2021 bull market (funding rate avg 30%), inflating the historical mean. If next 5 years look like 2018-2019 (funding 1-3%), live results will severely underperform.
2. Fill model uses queue-rank proxy: queue position is estimated from funding_stats borrowed-amount deltas, not real order-book history (Bitfinex doesn’t publish historical book snapshots). The proxy may over- or under-state competition.
3. Black swans: 5.5y of data has only 1 major event (2022 LUNA collapse) — too few samples to estimate tail risk.
4. Marketing band still beats Storm regime: if you’re starting in a high-vol week, expect yields below the marketing range until the regime calms.

Honest Caveats

A few things to be straight about:

1. 5.5y data ≠ eternity: includes 2020-2021 bull market (funding rate avg 30%), inflating the historical mean. If next 5 years look like 2018-2019 (funding 1-3%), live results will severely underperform.
2. Survivorship bias: I use Bitfinex’s filled candle rates, not “rates I could have grabbed”. Real queue competition may push fill rates below the model.
3. Black swans: 5.5y of data has only 1 major event (2022 LUNA collapse) — too few samples to estimate tail risk.

Full methodology → docs/strategy_review.md (when repo public)

Where the Code Is

Full backtest code:

• lending_bot/backtest/fill_model.py — fill model
• lending_bot/backtest/prepayment_model.py — prepayment hazard
• lending_bot/backtest/engine.py — walk-forward orchestration
• scripts/realistic_backtest.py — CLI entry

Will be reproducible after repo open-sources (currently private, 2026 Q3 roadmap).

Related Reading

• Multi-bucket vs single — why diversification wins
• Per-symbol floors — why fUSD ≠ fUST
• Prepayment hazard — the hidden cost of long periods
• Building Yieldsforge — the story
• Why Bitfinex Funding beats DeFi yields — the hub

Try Yieldsforge free for 7 days →

---

Disclosure: I’m the developer of Yieldsforge. Backtest uses Bitfinex public candle data, 2020-11 to 2026-04. All APY numbers net of 15% Bitfinex fee. Not investment advice; past performance does not guarantee future results.