humarb ResearchEmpirical preprint · 005 · DOI 10.5281/zenodo.23917835

Forecasting Cross-Venue Funding Cash Flows

A Walk-Forward, Cost-Adjusted Benchmark

Download paper Read the abstract
Perpetual futuresFunding ratesForecast evaluationWalk-forward validationConformal calibrationTransaction costs
Abstract162 words

Funding predictability is not executable net carry.

Cross-venue perpetual-funding comparisons require forecasts of holding-period cash flows, not annualizations of the latest quote. We construct a public-data benchmark covering 40 crypto assets, five venues, and 800 directed venue pairs over 90 days. Six causal estimators are evaluated at 24-, 72-, and 168-hour horizons in a frozen 30-day holdout containing 72,800 pair-origins per horizon. A one-day exponentially weighted estimator reduces MAE from $1.762 to $1.574 at 24 hours relative to a seven-day median (Holm-adjusted p=0.0035); improvements at longer horizons are not uniformly significant after multiplicity control. The conservative endpoint's coverage falls from 80.13% to 75.86% as the horizon rises from 24 to 168 hours. Online-scaled central intervals approach nominal 90% coverage but retain days below 80%. Among positively forecast routes, median realized funding is $0.17–$2.79, far below a $50 illustrative round-trip cost. The results support horizon-specific forecast evaluation and show why funding accuracy cannot be equated with executable net carry.

阅读中文摘要

跨场所永续合约比较需要预测持有期资金费现金流，而不是把最新报价直接年化。本文使用五家场所公开历史接口，构建40种加密资产、800个有向场所对、90天窗口的可复现基准，并在冻结的30天留出集上比较六种因果估计器。以7日中位数为参照，1日指数加权估计在24小时视界将MAE从1.762美元降至1.574美元，经15项Holm校正后仍显著；在72小时和168小时视界，改善不再一致显著。保守端守住率随视界延长由80.13%降至75.86%，后两个视界低于80%。在线缩放使中央区间平均覆盖接近名义90%，但最弱单日覆盖仍低于80%。在预测为正的路线中，实际资金费中位数仅为0.17—2.79美元，显著低于两腿进出场合计50美元的示例成本。资金费方向具有预测信息，但点预测、尾部校准与可执行净收益必须分别验证。本研究不把历史盘口、真实冲击成本或未测基差损益填为零；扣成本结果不是完整套利盈亏，不构成投资建议。

40Common crypto assets
05Public-data venues
800Directed venue pairs
218,400Frozen-holdout observations
Paper contents

I · Research question

The latest funding quote is not a holding-period cash-flow forecast.

When two venues quote different funding rates for the same asset, subtracting and annualizing them looks immediate and comparable. It also assumes that current rates persist, settlement intervals are compatible, and execution does not depend on size. This paper instead predicts the signed funding cash that a directed two-leg position would have received over a fixed horizon, then evaluates that prediction on later observations.

The target is deliberately narrower than total trading income. Historical executable books, depth, transfer paths, financing, liquidation outcomes, and cross-venue basis changes are not present in public funding histories. They are treated as unresolved components, not as zeros.

Research 01 and 02 define executable accounting. Research 05 asks whether the funding component survives a frozen out-of-sample test.

Read the companion methods paper

II · Data & methods

One public snapshot. One frozen evaluation protocol.

The fixed window runs from 2026-07-23 00:00Z to 2026-10-21 00:00Z. Instrument lists and 361,140 funding records were retrieved from official public APIs on 2026-11-03. The 40-asset exact-name crypto intersection produces 20 ordered venue pairs per asset. Both directions remain in the statistical panel.

Forecast target · USD per $10,000 per leg
ŷ(m) = A·H·(r̂S(m) − r̂L(m))
I

Per-record interval normalization

Hourly rates are computed as settled rate divided by the interval attached to each record. Lighter's direction sign and percentage-point units are handled explicitly; no universal eight-hour assumption is made.

Units first
II

Six causal estimators

Latest rate, 3/7/14-day complete-UTC-day medians, and interval-weighted EWMA with one- or three-day half-lives. A 60-second boundary prevents same-settlement leakage.

No future inputs
III

Frozen 30-day holdout

The protocol fixes the estimators, horizons, cost cases, missing-data rule, and inference before new holdout computation. Each horizon contains 72,800 pair-origins; row counts are not treated as independent sample size.

24 · 72 · 168h
IV

Costs and omitted basis kept separate

Four executions across two legs produce $20, $50, and $105 total-cost cases after transfer. Historical mid-spread P&L is not measured and is not silently set to zero.

F − C ≠ total P&L

III · Main results

Recency helps conditionally; costs dominate the median.

01 / 04
Figure 1

One-day EWMA has the lowest 24-hour MAE; no estimator dominates every horizon and loss function. USD per $10,000 per leg. Source: public funding histories, authors' calculations.

Figure 2

Conservative coverage is 80.13%, 76.87%, and 75.86% at 24, 72, and 168 hours. The dashed line is 80%, not a certified target.

Figure 3

Median funding after cost is negative even in the low-cost case. Historical basis P&L is not measured and is not set to zero.

Figure 4

Online scaling approaches average nominal coverage but weak days remain. Source: 72-hour holdout origins and authors' calculations.

Scroll horizontally or use the arrows. Select a figure to inspect its details.

IV · Evaluation & robustness

Statistical improvement is horizon-specific and does not imply economic significance.

At 24 hours, one-day EWMA reduces MAE from $1.762 to $1.574 relative to the seven-day median; the Diebold–Mariano result remains significant after Holm adjustment across 15 comparisons (p=0.0035). At 72 hours, one- and three-day EWMA both have adjusted p=0.0959. At 168 hours, no EWMA improvement over the seven-day median survives adjustment.

Direction is stronger than level

One-day EWMA sign accuracy is 69.60%, 75.24%, and 72.93% at 24, 72, and 168 hours. Magnitude errors remain large, especially when forecasts are near zero.

Coverage is time-varying

Online central intervals reach 91.14%, 90.86%, and 90.39% average coverage, but their worst daily coverage is 77.42%, 76.83%, and 79.50%.

Long tail and late period are weaker

At 72 hours, long-tail MAE is $4.991 versus $3.626 for majors. Conservative coverage falls from 79.30% in the first 15 holdout days to 74.09% in the last 15.

PBO is not profitability

CSCV/PBO is low for estimator-rank stability in this holdout. It does not establish positive cash flow, executable capacity, or total-return robustness.

V · Evidence boundary

What this paper does not claim.

  • No total-arbitrage P&L. Funding-after-cost is not realized net income. Basis changes, executable depth, financing, and liquidation are not measured.
  • No full-market representation. The sample is five venues and 40 current-common assets. Current-list intersection creates survivorship and identifier-mapping limitations.
  • No universal estimator winner. Responsive methods are strongest at short horizons; the evidence does not support replacing every longer-horizon median with one EWMA rule.
  • No coverage guarantee. Average calibration hides weak days and correlated misses. Day-block uncertainty is primary; iid Wilson, Kupiec, and Christoffersen results are secondary.
  • No capacity or annual-return estimate. Scenario annualization is an arithmetic scale conversion, not a repeatable, financed strategy result. Raw-data redistribution requires venue-terms review.

This paper is methodological and empirical research. Historical forecasts and scenario calculations are hypothetical evidence, not trading instructions or investment advice.

VI · Conclusion

Recent funding history contains information, but total cash-flow and implementation questions remain open.

The benchmark supports selective use of responsive estimators, explicit horizon-specific evaluation, and separate measurement of tail coverage. It does not support equating sign predictability with executable income. The next evidence step is a preregistered extension with point-in-time contract membership, more regimes, historical executable books, and measured basis changes—not additional tuning on this holdout.

VII · Citation

Cite the evidence.
Preserve its boundaries.

Cite the concept DOI, which resolves to the latest version. The v1 version DOI and Zenodo record are listed below.

Concept DOI10.5281/zenodo.23917835
Zenodo record · version DOI 10.5281/zenodo.23917836

Liu, H. (2026). Forecasting cross-venue funding cash flows: A walk-forward, cost-adjusted benchmark [Preprint]. humarb Research. https://doi.org/10.5281/zenodo.23917835

Download .bib Download PDF 中文摘要

Figure