Strategy Graveyard
✕ INVALIDATEDPostmortem · No. 001

Funding Rate Arbitrage

Long Spot / Short Perpetual, Delta-Neutral

Apex Quant tests cryptocurrency trading hypotheses with strict in-sample / out-of-sample discipline. When a strategy fails out-of-sample, we document it publicly. This is one of those entries.

Fig. 1 — Cumulative edge, IS vs OOSUNI/USDT · representative
highzeroentry threshold (1 bp)IN-SAMPLE · 2022–2024OUT-OF-SAMPLE · 2025–2026edge peaks, then collapsestime →
Figure 1. Cumulative edge accumulates steadily through the in-sample period, then flattens and reverses out-of-sample. The same parameters, applied to unseen data, produce no exploitable signal. Curve is illustrative; underlying metrics below are exact.
IS Sharpe6.94
OOS Sharpe-0.71
VerdictInvalidated
01

The Hypothesis

In Binance perpetual futures, when funding rate is consistently positive, short perpetuals receive payments from longs every 8 hours. A delta-neutral position (long spot + short perpetual of same size) captures this funding payment while remaining indifferent to price direction.

The academic literature (Soska 2021, Doblas 2025) suggested this strategy could generate 15–30% annualized returns with low drawdown — citing “anomalous” persistent positive funding rates in crypto.

We set out to validate this claim with our own data and methodology.

02

What We Built

Data infrastructure

143,535 funding rate settlements across 30 crypto pairs, 4.4 years (2022 → 2026).

Backtesting engine

Full delta-neutral simulator with realistic costs (0.40% round-trip: spot taker + perpetual taker + slippage).

Methodology

Strict in-sample / out-of-sample split with predefined verdict thresholds.

In-Sample (IS)
2022-01-01 → 2024-12-31
3 years · 82,200 settlements
Out-of-Sample (OOS)
2025-01-01 → 2026-05-13
1.4 years · 37,350 settlements
03

What We Found In-Sample

— looked promising

Funding rates are structurally persistent in crypto:

0.687
Mean autocorrelation (8h lag)
81.82%
Sign persistence
24 / 30
Symbols with positive mean funding
UniverseIS CAGRIS SharpeIS Max DD
Top 3 (UNI, LINK, LTC)+4.06%6.71-0.51%
Top 5+4.17%6.94-0.30%
Top 10+4.18%7.27-0.24%
Table 1 — In-sample performance by universe size. Sharpe > 6 across all cuts.

These metrics looked institutionally compelling: Sharpe > 6, drawdown < 1%, edge robust across symbols. If we had stopped here, we could have written marketing copy claiming “Sharpe 6.7 carry strategy on crypto.” Many people do.

04

What We Found Out-of-Sample

— the truth

The same strategy, with the same parameters fixed in-sample, run on 2025–2026 data:

UniverseOOS CAGROOS SharpeOOS Trades
Top 3-0.08%-0.712
Top 5-0.05%-0.712
Top 10-0.07%-1.184
Full Universe (25)-0.06%-1.459
Table 2 — Out-of-sample performance. The edge does not survive.

The edge collapsed.

05

Why The Edge Collapsed

Finding 1 — Funding rates compressed in 2025–2026

The mean funding rate of our top symbols in OOS:

UNI
0.48 bps
vs ~0.78 bps historical
LINK
0.44 bps
LTC
0.42 bps

A ~42% compression versus historical averages, driven by public availability of the research since 2021, more capital chasing the same anomaly, and exchanges adjusting funding mechanisms.

Finding 2 — Our 1 bp entry threshold almost never triggered in OOS

Settlements above 1 bp in our top symbols:

UNI
0.13%
~2 of 1,494
LINK
0.00%
LTC
0.07%

The threshold that worked in 2022–2024 became operationally meaningless in 2025–2026. With only 2–4 trades per universe over 1.4 years, transaction costs (0.40% round-trip) consumed any micro-gains.

06

The Honest Methodological Disclosure

Our IS threshold (1 bp) was selected by looking at the full historical period, including what later became the OOS. This is a documented form of selection leakage.

We explicitly considered re-running the analysis with a threshold chosen using only IS data. We declined — for two reasons.

  1. 1It would have been parameter mining: choosing thresholds to make the strategy “work” violates the methodology we committed to before seeing results.
  2. 2Even if a re-tuned threshold worked OOS, it would be a different strategy than the one we set out to test. Calling it the same hypothesis would be retrospective rationalization.

The hypothesis we tested failed. We accept that.

07

What This Means for Funding Arbitrage

We are not claiming Funding Arbitrage cannot work for anyone. We are claiming:

  1. 1With Binance VIP 0 fees + slippage = 0.40% round-trip, the strategy is not viable on top-30 USDT pairs in the 2025–2026 regime.
  2. 2With a fixed entry threshold of 1 bp, the OOS edge collapses to noise.
  3. 3Public availability of the research has likely eroded the edge since 2021–2025.

Variations that might work (not tested by us): maker-only execution, higher fee tiers, other exchanges, larger universe, or a regime-adaptive threshold. We are not pursuing these. If you do, please publish your findings.

08

Technical Appendix

Methodology

Strict IS/OOS split, universe pre-frozen via IS-only ranking + persistence filter (Top 15 in ≥2 of 3 annual sub-windows). All thresholds and risk warnings defined in advance and machine-applied.

Risk warnings fired

15 (5 strong, 10 moderate) — all consistent with the invalidation verdict.

Data

143,535 funding settlements + 48k daily klines. Available to Pro subscribers.

Code

Open-source on request.

Sharpe / CAGR / Drawdown are computed identically in IS and OOS — same simulator, same costs, same parameters. The only thing that differs is the data period.

“Efficient markets destroy fantasies. Our job is to find the small inefficiencies that survive.”

Apex Quant · Crypto Quant Intelligence