Paper contents
I · Research question
Ranking by funding gap is not ranking by complete scenario value.
Current rates, order books, funding schedules, valuation prices and fees describe a conditional scenario. Future rates, exit basis and execution costs need not remain unchanged. Selecting the largest displayed spread adds a second problem: the selected estimate belongs to a different error population from the full candidate set.
The research question is deliberately narrow: how should current cross-venue funding scenarios be ranked, how much joint overestimation can a selected scenario tolerate, and what additional evidence is required before that ranking becomes a defensible decision rule?
Research 01 measures executable net carry; Research 02 frames net-carry evidence; Research 10 supplies the event-time ledger. Research 11 begins after that ledger closes and studies its error budget and selection layer.
Read Research 02: Framing Net Carry
II · Framework
One accounting boundary. Two distinct ranking layers.
D = S0 − Π ⇒ (Π > 0 ⇔ D < S0)
Current scenario baseline
Complete scenarios share quantity, currency and reporting basis, holding view, event schedule, fee basis and information version. Same-book funding, basis, crossing, fees and other costs enter once.
Joint realization error
The error budget decomposes funding-rate, funding-price, rate–price interaction, basis, crossing and fee differences. Missing evidence remains NA; covariance is never assumed to be zero.
Preference-dependent decision layer
J requires the joint error distribution, a loss weight and a declared capital-hour contract. If these inputs are unavailable, J is unavailable—not zero and not a reason to suppress the descriptive S₀ ranking.
Evaluation after selection
Chronological holdouts, interval scores, Newey–West comparisons and Holm adjustment are reported alongside selected-subset errors. Overall marginal coverage does not automatically cover the selected maximum.
Expectations require a joint distribution; A is reserved capital integrated over time. λ and ρ are not calibrated in this paper.
NA is not zero, and marginal coverage is not selected coverage.
A missing book, record or official account cash flow cannot be replaced by zero. A verified empty event set is different from a scheduled event with missing evidence. Combined amounts use an explicit one-for-one USDC/USD parity scenario, not realized account P&L.
III · Figures
The identity holds; calibration and selection require separate evidence.
Scroll horizontally or use the arrows. Select a figure to inspect its details.
IV · Evaluation and robustness
Funding-rate evidence is not complete strategy evidence.
The paper uses chronological prediction only to measure funding-rate components. It does not convert those results into a complete backtest, a probability of profit, or a calibrated decision objective. Every attempted specification remains in the reproduction outputs.
There are 3,720 holdout predictions across six venue–asset series. The analysis reports MAE/RMSE, 23-lag Newey–West Diebold–Mariano tests, interval scores, Kupiec and Christoffersen diagnostics, and Holm-adjusted families.
Two holdout halves, two venues, three assets, five rolling windows, absolute and squared loss, four dependence lags, and four one-sided levels are retained. Selected maxima are reported separately from all candidates.
A future confirmation must freeze D0/D1/D2/D3, purge overlapping labels, preserve shared-event blocks, and score complete capital paths. The present retrospective sample and one-hour snapshot do not meet that standard.
The first two statements concern accounting and measurement. The third requires joint error, loss preference and capital inputs that this study does not have.
V · Limitations
Public proxies and reproducible arithmetic—not live performance.
- Decision scope. No joint closed-position error distribution, λ, ρ, capital contract, positive-expectation estimate, market-wide optimum, or validated J is supplied.
- Statistical scope. The retrospective sample covers two venues, three continuing assets and about two months. Asymptotic tests and chosen lags do not eliminate regime shift or shared-event dependence.
- Execution scope. Public-book VWAP excludes queue position, reaction, hidden liquidity, partial fills, financing, transfer, liquidation and account-specific fees.
- Currency and evidence scope. Combined amounts are parity-denominated scenario units, not realized USD, USDT or USDC account cash. Public oracle/mark snapshots are not official account funding records.
- Protocol and review scope. The forward snapshot is locally hashed, not externally preregistered. Internal multi-role review and local similarity screening are not external peer review or commercial plagiarism certification.
This paper is for research and methodological information only. It does not provide investment, legal, accounting, or tax advice; does not recommend any trade, venue, leverage, or position size; and does not promise income, execution, capital protection, or protection from liquidation.
VI · Conclusion
A testable error budget—not a validated arbitrage recommendation.
The exact contribution is a reconciliation rule: Π > 0 if and only if D < S0. The evaluation contribution is equally important: current-scenario ranking remains separate from decision-ready J selection, and selected-subset error remains visible rather than hidden inside a full-candidate average.
The appropriate next step is a frozen, complete forward evaluation—not another universal threshold. Validate the scenario and error ledger first, estimate joint error only on isolated data, and test J on locked capital paths only after the required distributions and preferences exist.