Cross-venue perpetual-futures funding spreads are often presented as arbitrage opportunities, yet a quoted differential is neither a forecast nor net trading income. We study when settled funding differences can cover four-fill execution costs. Public APIs supply 121,864 funding records for eight major contracts on Gate, Hyperliquid, and Pacifica, aligned into 521 eight-hour blocks from May to November 2026; current public books on those venues and Paradex provide a separate cost snapshot. The median absolute spread is 0.45 bp per eight hours and the 99th percentile is 2.99 bp. Out of sample, a capped forecast reduces MAE from 0.430 to 0.399 bp relative to last-value persistence, with significant Diebold–Mariano improvements. This statistical gain does not create economic feasibility: no candidate generates a 24-hour entry at a 34 bp hurdle, while sparse 72-hour entries at 20 bp lose money in the funding-cash proxy. Probability forecasts underperform a training-frequency baseline. The evidence supports a break-even discipline rather than a claim of persistent arbitrage.
中文摘要
本文检验跨所永续合约资金费价差能否覆盖四次成交的执行成本。公开接口提供三个历史场所、八个合约的 121,864 条资金费记录，并形成 2026 年 5 月至 11 月的 521 个 UTC 8 小时区块。绝对价差中位数仅为每 8 小时 0.45 bp，第 99 百分位为 2.99 bp。样本外 capped 预测将 MAE 从 0.430 降至 0.399 bp，但这一统计改善并未带来经济可行性：24 小时、34 bp 主情景下所有方法均未入场；72 小时、20 bp 情景下的稀疏入场全部亏损。概率预测也逊于训练频率基准。证据支持严格的盈亏平衡与弃权纪律，而不是持续套利的主张。
Three filters separate a visible spread from a trade
A funding spread must be large enough to cover four executions, persistent enough to survive the holding period, and predictable enough to select the correct direction before settlement. A contemporaneous scanner addresses none of these conditions completely. This paper joins mechanical break-even, out-of-sample directional prediction, and data-snooping control in one decision framework.
0.45 bpmedian absolute spread per 8h
2.99 bp99th-percentile absolute spread
0primary-case entries for every estimator
28.72 bpmedian current core-venue cost
The historical sample covers 11 May–3 November 2026. Cost estimates use a separate current-book snapshot acquired on 3 November 2026.
Data and method
Native settlement histories are aligned before any prediction
The sample contains BTC, ETH, SOL, XRP, DOGE, ADA, LINK, and AVAX on Gate, Hyperliquid, and Pacifica. Gate contributes native eight-hour records; the other venues contribute hourly records. An hourly block is admitted only when all eight distinct hours are present. Missing observations are never interpolated or zero-filled.
Table 1 · Sample construction and chronological evaluation
Native public records
121,864
Complete venue–asset 8h blocks
18,904
Warmed-up pair–asset–block rows
12,504
Training / test blocks
312 / 209
Test panel observations
5,016
Four fixed estimators; no test-set tuning
The candidates are last-value persistence, a trailing nine-block mean, a trailing nine-block median, and a capped rule that retains the smaller magnitude of persistence and the mean when their signs agree. Entries occur only when forecast cumulative magnitude exceeds the stated cost. The primary case uses 24 hours and 34 bp; 8/24/72-hour horizons and 20/34/60 bp scenarios form the sensitivity grid.
Forecast quality and economic feasibility are evaluated separately
Point forecasts use MAE, RMSE, and Diebold–Mariano tests with cross-sectionally averaged losses and 18-lag Newey–West standard errors. A low-cost statistical scenario evaluates Brier score and log loss. White’s Reality Check uses 2,000 moving-block bootstrap replications. Current public books provide depth-walked four-fill cost scenarios; they are never treated as historical fills.
Results
The whole observed spread distribution sits below one-settlement costs
The median absolute differential is 0.45 bp per eight hours, the 90th percentile is 1.34 bp, and the 99th percentile is 2.99 bp. None of the 12,504 aligned observations reaches 20 bp. Multi-settlement persistence is therefore central, not an optional refinement.
Figure 1 · Spread distribution. Absolute paired differentials in bp per 8h; dashed references are 20, 34, and 60 bp cost scenarios. Source: public Gate, Hyperliquid, and Pacifica settled funding histories, 11 May–3 November 2026.
A capped forecast improves accuracy without creating a trade
The capped rule lowers held-out MAE from 0.430 to 0.399 bp and RMSE from 0.666 to 0.634 bp relative to persistence. Diebold–Mariano tests favour it under absolute loss (t = −4.510, p = 0.0000065) and squared loss (t = −2.561, p = 0.0104). Yet it generates no entries in any decision case.
Table 2 · Out-of-sample point forecasts, bp per 8h
Estimator
MAE
RMSE
Sign accuracy
Zero
0.5397
0.8118
0.1946
Last value
0.4296
0.6658
0.7444
Three-day mean
0.4589
0.6768
0.6820
Three-day median
0.4493
0.6963
0.6842
Capped
0.3988
0.6340
0.6974
Figure 2 · Alternative estimators. Mean net proxy per entered trade at 72 hours and 20 bp. Capped has no entries and no defined bar. Source: deterministic out-of-sample decision replay; fixed-notional funding-cash proxy, not complete P&L.
The primary case has no entry; relaxed cases have only losing entries
Every estimator faces 1,656 primary decisions and enters zero times. The future-informed two-sided reference also finds no 24-hour window covering 34 bp. At 72 hours and 20 bp, persistence enters seven times, the mean six, and the median three; all entries lose. The ex post reference identifies only two cost-covering windows, with perfect future direction imposed by construction.
Figure 3 · Net proxy through time. Cumulative cross-sectional mean net bp per independent route, 72h at 20 bp. Source: out-of-sample public-data replay. Shared routes are not a portfolio equity curve.
Probability scores do not rescue tail-event selection
The primary 34 bp training case has zero positive events, so a two-class probability model is unidentified. A separate 24-hour, 5 bp statistical scenario produces Brier skill of −1.01%, with a moving-block 95% interval of [−2.69%, +0.04%]. Log loss is also worse than the constant training-frequency baseline, while the Brier-skill interval includes zero and therefore does not establish worse population skill. This scenario is a diagnostic, not an executable-cost assumption.
Figure 4 · Probability calibration. Out-of-sample predicted probabilities and observed frequencies in the 24-hour, 5 bp statistical scenario. Source: training-only logistic model evaluated on the chronological holdout.
Current books establish cost scale, not historical execution
Across the 24 routes involving only Gate, Hyperliquid, and Pacifica, median four-fill cost is 28.72 bp for a USD 10,000 quote budget per fill, with a 17.61–47.60 bp range. Paradex-involving routes are reported separately because their median scenario is approximately 2,204 bp under thin current depth. The maximum cross-book timestamp skew is 14.392 seconds.
Figure 5 · Current cost snapshot. Depth-walked four-fill scenarios by asset on a logarithmic scale; dashed reference is 34 bp. Source: public books acquired 3 November 2026. Snapshot quotes are not historical fills.
Evaluation and robustness
Time, venue, parameter, and estimator slices do not reverse the result
Table 3 · Median estimator at 72 hours and 20 bp
Slice
Decisions
Entries
Mean net bp
Gate–Hyperliquid
184
1
−11.517
Gate–Pacifica
184
0
—
Hyperliquid–Pacifica
184
2
−16.248
Early test half
264
2
−14.534
Late test half
288
1
−14.945
Chronological: early and late test halves both contain only losing median-rule entries.
Venue: no pair contains a cluster of successful forecast-selected trades.
Parameters: every 8h and 24h case has zero entries; activity appears only at 72h and remains unsuccessful.
Estimators: capped is best on point loss but abstains; persistence enters most often in the relaxed case and loses on all seven entries.
Selection: White’s Reality Check gives p = 1.0 against zero cash for the four-candidate 72h, 20 bp family.
Figure 6 · Cost sensitivity. Three-day median at 72 hours; left axis is entry share and right axis is mean net bp per entered trade. The 60 bp case has no entry and no defined per-trade mean.
Twelve unit tests—eight accounting tests and four analysis regression tests—cover common base quantity, actual exit-notional fees, book-walk direction, asset-specific fees, floating-point tolerance, insufficient depth, duplicate and missing events, and settlement marks. These tests prevent implementation errors; they cannot supply missing historical information.
Evidence boundary
The paper rejects these tested signals; it does not prove universal no-arbitrage
What is identified
Settled public funding-rate differentials on a strict common grid, out-of-sample forecast accuracy, and fixed-notional funding cash under disclosed cost scenarios.
What is not identified
Historical four-way execution, complete price/basis P&L, exact oracle-price cash, margin and transfer costs, liquidation, and account-level portfolio returns.
The eight currently available common contracts create survivorship and external-validity limits.
Gate’s public trailing-history boundary restricts the common sample to roughly six months.
Current books are one snapshot, not simultaneous guaranteed fills or historical costs.
The ex post two-sided reference uses future information and is not a feasible oracle strategy.
Maker, rebate, stress-period, and materially different contract designs remain outside the evidenced conclusion.
Not investment advice. This is an empirical research preprint, not an offer, solicitation, or recommendation to trade. Example parameters are disclosed research settings, not production rules.
Conclusion
Disciplined abstention is the supported decision in this sample
A small improvement in average point forecasts does not bridge the gap between tightly anchored funding rates and four-fill taker costs. Future work needs synchronized historical books, complete basis and margin accounting, more heterogeneous venues, and prospective stress-period samples. Until those tests are passed, a visible funding difference remains a quotation to investigate, not a profit to recognize.
The paper differentiates itself from humarb Research 01 on executable net-carry measurement and humarb Research 02 on settled-history accounting by adding an independent multi-month public-data forecast and decision evaluation.
Downloads and citation
Paper, summary, metadata, and reproducible evidence
Liu, H. (2026). When is a cross-venue funding-rate spread worth trading? Public-data evidence on forecastability and cost coverage (humarb Research 04) [Preprint]. humarb Research. https://doi.org/10.5281/zenodo.23917371
BibTeX
@misc{liu2026crossvenuefunding,
author = {Liu, Hang},
title = {When Is a Cross-Venue Funding-Rate Spread Worth Trading? Public-Data Evidence on Forecastability and Cost Coverage},
year = {2026},
institution = {humarb Research, London, United Kingdom},
type = {humarb Research 04},
url = {https://humarb.ai/research/cross-venue-funding-spread/},
doi = {10.5281/zenodo.23917371},
note = {Empirical preprint, not peer reviewed.}
}
