Funding Methodology

• Realized — a settled funding payment, from venue history endpoints, keyed to its settlement timestamp.
• Current — the venue's indicative rate for the running interval. It moves until settlement.
• Predicted — the venue's own forecast for the NEXT settlement, only where the venue publishes one. A prediction can change any number of times before settlement and is never displayed as realized or guaranteed.

Storage, API payloads, schema and every table keep the three apart; cross-venue aggregates and divergence figures are computed within ONE type at a time.

Notable verified semantics: OKX's nextFundingRatefield is empty by design under its current formula — the venue no longer publishes a prediction, so we don't either. Kraken's ticker rates are absolute USD figures and are divided by mark price before use. Other venues appear with current rates via our aggregated feed; their intervals are the venue's standard schedule for majors and are marked “assumed” until read per-instrument. Capability is distinct from a temporarily failed observation.

• Linear per-period conversion: rate1h = rawRate × 60 / intervalMinutes; 8h likewise. This matches the periodic-payment model all covered venues use on BTC/ETH majors; any venue with compounding or caps gets venue-specific code documented here before inclusion.
• The raw venue rate and original interval are always stored and shown — normalized values never replace them.
• An unknown interval yields NO normalized value (null), never an assumed 8h conversion.
• Annualized figures are a linear extrapolation of the 1h rate (×24×365), labeled as such everywhere — funding resets every interval, so this is a comparison aid, not an expected yearly return.

• Sweeps every 15 minutes (desired 5-minute cadence documented as a platform-budget trade-off), storing verbatim raw envelopes (append-only) with sourceTimestamp separate from ingestedAt, quality-prioritized idempotent upsert, advisory lease and per-venue failure isolation — CEX and DEX fail independently.
• Realized settlements are recorded keyed (venue, settlement timestamp), append-only. Prediction accuracy uses the LAST prediction observed before that settlement and reports the observation lead time; early predictions are never scored without it.
• Rows with future-dated source timestamps or parse-error-class rates are dropped as invalid; rates older than 30 minutes at ingest are flagged stale. Missing data is null/“—”, never zero.
• History starts at our first stored sweep (2026-07-16) — no synthetic backfill; historical charts and trend claims wait for an established series (age AND observation count). Storage growth ≈3 MB/day compressed.

• Divergence compares the SAME rate type on comparable linear BTC/ETH contracts within one sweep, normalized to the same period. Spot and perp data are never blended.
• Published figures are gross spreads. A net opportunity requires fees, spread and slippage on both legs (from the order-book cost engine), holding-period and interval-uncertainty assumptions — until all inputs are present the model reports incomplete rather than a precise-looking net number. Funding strategies are not risk-free: rates flip, positions carry liquidation and venue risk, and costs often exceed short-window funding income.

Errors are handled under the correction policy. Pipeline: PerpFinder Data Team · review: PerpFinder Editorial Review. Sources: venue public funding APIs (Binance, OKX, Bybit, Kraken Futures, Hyperliquid, dYdX directly; further venues via PerpFinder's aggregated feed), attributed per row.