Crypto Data And Execution

Crypto data and execution

Crypto inverts several assumptions baked into equity tooling: there is no session, no holiday calendar, no consolidated tape, no corporate actions — and a far worse survivorship problem.

1. Pick a tool

✅ CCXT Pro is free. It was merged into the MIT ccxt package at v1.95 (2022) — WebSocket methods (watchTicker, watchOrderBook, watchTrades, watchOrders…) are no longer a paid product. Ignore any subscription-expiry notice or tutorial telling you to buy it. Today: ccxt (sync REST) · ccxt.async_support · ccxt.pro (async + WS) — one package. Since v4.5.66 it also covers prediction markets (Polymarket, Kalshi, Hyperliquid) through the same unified API.

2. 🚨 Survivorship is worse here than in equities

A pair list built from today's top-volume coins excludes every coin that died — and in crypto that is most of them. Exchanges delist aggressively and silently; a token that went to zero simply stops appearing in fetch_markets().

freqtrade, jesse, OctoBot and hummingbot are all chronically survivorship-biased in practice because their pair lists are constructed live. Backtesting 2021 on today's pairs is not a backtest of 2021.

Guards: snapshot fetch_markets() periodically and version it; reconstruct historical pair lists from your own archived snapshots; state explicitly when you could not, and treat the result as an upper bound.

3. What crypto breaks in equity-derived tooling

1. No session, no calendar. exchange_calendars and pandas_market_calendars do not apply. Annualization is 365, not 252. A library defaulting to 252 understates Sharpe and vol by 16.9% on a calendar-day series, and a calendar-CAGR-over-√252-vol ratio overstates by 20.4% (§8).
2. No corporate actions, but plenty of discontinuities — token migrations, redenominations, chain splits, ticker reuse. Price series are stitched by the exchange with no adjustment record.
3. No consolidated tape. Every venue has its own book and its own price. "The" BTC price does not exist; cross-venue backtests need explicit venue attribution.
4. Perpetuals carry funding. Funding is paid/received every 8h, 4h or 1h — per symbol (§9) — and can be larger than the alpha being measured. A perp backtest without funding is not a backtest.
5. Fees are large and tier-dependent. In Alpha Arena S1, fees alone consumed 13–17% of capital in ~2 weeks. Maker/taker asymmetry changes strategy viability outright.
6. 24/7 means no overnight gap — strategies keyed to opens/closes have no analogue.

4. freqtrade's bias detectors — worth using even from another framework

• freqtrade lookahead-analysis re-runs a backtest on progressively truncated data and flags indicators whose historical values change. Limits it states honestly: only checks triggered signals; can false-positive on limit orders with custom pricing callbacks.
• freqtrade recursive-analysis varies startup_candle_count and reports each indicator's last-row variance — the only off-the-shelf tool measuring the recursive warm-up problem (EMA/RSI/ ADX converge differently depending on how much history preceded them; backtest sees 5,000 candles, live sees ~1,000).

Both ideas are portable — see signal-construction §5 and plugins/fin-core/skills/signal-construction/scripts/assert_causal.py.

5. ccxt gotchas

• The unified API normalizes common fields only; venue-specific behaviour goes through params passthrough and implicit methods that are not portable across exchanges.
• enableRateLimit must be on.
• fetchOHLCV has per-exchange limit caps and inconsistent pagination — always loop and dedupe on timestamp.
• set_sandbox_mode(True) exists but only some exchanges have testnets. Verify exchange.urls['api'] actually points at a testnet host before assuming you are safe.
• Precision and rounding rules differ per venue and silently reject orders.
• Order-status vocabularies are only loosely unified.

6. Order safety

Everything in broker-execution-apis §3 applies. Crypto-specific:

• 🚨 Create API keys with trade scope only — never withdraw. Withdraw-enabled keys are how accounts get drained.
• stoploss_on_exchange (freqtrade) places the stop at the exchange so it survives your process dying. Off by default; the most important line in retail crypto risk management.
• IP-allowlist the key where the venue supports it.
• Deterministic client order IDs — retries on a timed-out order are routine on crypto venues.

7. Reference files

references/<library>.md for versions, licences, venue coverage and quirks.

8. 🚨 Annualisation: the factor is rows per year, and the error runs both ways

✅ scripts/perp_mechanics.py §1, one synthetic calendar-day series (1,460 rows, seed 0):

Which way the 252 error goes depends on the formula. mean/std·√252 understates; a ratio whose numerator is annualised by calendar time and whose denominator is √252 vol overstates by √(365/252) − 1 = 20.4%. ✅ Live: quantstats 0.0.81 and empyrical 0.5.5 both default to 252 and reproduce the understated 0.5736 — pass periods=365 / annualization=365. ✅ quantstats 0.0.81's cagr() uses years = len(returns) / periods, not calendar time: a 365-row year is 1.45 "years" at the default, and it returns 15.97% for a series whose calendar CAGR is 23.94%.

The mistake in the other direction. A strategy flat at weekends, kept as weekday rows only (1,042 of them), annualised with √365 overstates its Sharpe by 20.4% (0.9407 vs 0.7816). The same P&L padded with weekend zeros (1,460 rows) and annualised with √365 gives 0.7946 — the weekday/√252 answer to 0.0129. The factor is the number of rows per year in the series you hand the function: 365 for calendar days, 252 for weekday-only rows, 8,760 for hours, 1,095 for 8h funding intervals.

9. 🚨 Funding is the carry a spot backtest omits

The crypto twin of ../../../fin-futures-fx/skills/fx-markets/SKILL.md §1: a spot series has no carry, a position does. For a perp the carry is funding, exchanged between longs and shorts at every settlement. payment = notional × rate per interval. ✅ OKX docs (public/funding-rate): positive rate = longs pay shorts; negative = shorts pay longs. It is charged on notional, so leverage multiplies it against margin: $100,000 at 0.01% per 8h is $10 per settlement, $30/day, $900 (0.90%) per 30 days — 9.0% of margin per month at 10x.

🚨 The interval is per symbol, not "8h". ✅ OKX, 2026-09-08, all 644 live swaps: 357 settle every 8h, 286 every 4h, 1 every 1h. OKX's docs say to read nextFundingTime − fundingTime; Bybit's docs: each symbol has a different funding interval, read instruments-info. ⚠️ Binance's docs and API answered HTTP 202/451 from here — nothing about Binance is verified in this file.

✅ A real year — Deribit BTC-PERPETUAL, hourly interest_1h summed per UTC day, 2025-09-08 → 2026-09-07: a long paid 2.77% of notional while the index fell 29.0%; 280 of 365 days positive; worst day 7.36 bps, best −3.39 bps. The largest 30-day bill was 0.825% (window ending 2025-10-28) against a +1.15% price move; |funding| exceeded |price move| in 10 of 336 windows and flipped a long's sign in 3. Median 30-day |move| 7.97% vs median 30-day funding 0.128%: in this low-funding year funding was a tax, rarely the whole story — and still 27.7% of margin at 10x. ✅ OKX BTC-USDT-SWAP (linear), last 92 days: 1.11% of notional = 4.40%/yr, price +23.1%; 29 of 277 settlements printed exactly 0.0100%, the instrument's interestRate field, and none exceeded it (⚠️ the premium-plus-clamped-interest formula predicts exactly that mass point). ✅ OKX caps each settlement at ±0.375% — pinned there, 33.75% of notional per month, a hypothetical. ⚠️ Illustrative, not observed in either sample: a bull-market path at 0.05% per 8h costs a long 4.5% of notional in 30 days, 45% of margin at 10x. Fetch the history; do not assume the regime.

A perp backtest is price return minus funding paid (long) or plus funding received (short), settlement by settlement, on the notional at each settlement. ✅ fetch_funding_rate_history is True in ccxt 4.5.78's has for binance, bybit, okx and deribit, False for coinbase and kraken; the base class raises NotSupported. Depth served: ✅ OKX 3 months (documented), Deribit ≥ 420 days hourly (pulled today), Bybit 200 rows/call (documented). Archive it yourself — the venue's window is shorter than your backtest.

10. Basis and cash-and-carry

A dated future trades away from spot by the basis, which is what a perp's funding stream approximates. annualised = (F/S − 1) × 365/days (simple) or (F/S)^(365/days) − 1. ✅ Deribit marks, 2026-09-08 13:25 UTC, index 78,449.67:

✅ OKX BTC-USD-261225 minutes later: 1.444% → 4.89%/yr. Two venues, one curve, both near the 2.77%/yr the perp actually paid (§9).

The trade: buy spot, sell the dated future, hold to expiry, collect the basis whatever the price does. On $100,000 via BTC-25DEC26: $1,356 locked, $35 futures taker fee (✅ 0.035%), 4.47%/yr net of that fee. The risks are not in the arithmetic:

• Margin on the short leg. A 30% rally marks the short −$30,000 and a linear venue wants it in cash now; the spot gain is a coin, not cash. At 3x isolated (mmr 0.4%) the short is liquidated at +32.8%, at 5x at +19.5% — the hedge becomes a naked long. On an inverse future with BTC as collateral a 1x short is a synthetic dollar and cannot be liquidated (§12).
• Exchange risk — both legs and the collateral sit on venues (§13).
• A perp instead of the dated future does not lock anything: the carry is the funding stream, sign-changing, never expiring into the basis. This is the roll analogue in ../../../fin-futures-fx/skills/futures-continuous-contracts/SKILL.md §5 — a dated future rolls and earns roll yield, a perp never rolls and pays funding. Neither is in a spot series.

11. 🚨 Liquidation is not a stop-loss

Maintenance margin is the equity floor below which the venue closes you. ✅ OKX BTC-USDT-SWAP tier 1 (positions up to 10 BTC = $784,622 at the fetched index): mmr 0.4%, initial 1%, max 100x; tiers rise with size. Isolated linear position, fees ignored: long liquidates at (1 − 1/L)/(1 − mmr), short at (1 + 1/L)/(1 + mmr), bankruptcy at ∓1/L. ✅ Computed:

The simplified entry × (1 − 1/L + mmr) differs by a few bps; funding and fees move it more.

Why it is not a stop: (1) you keep almost nothing — at 10x it fires with 3.6% of margin left. (2) A liquidation fee — ✅ Deribit's instrument spec carries max_liquidation_commission = 0.01: 1% of position value = 10% of margin at 10x, more than remains, so the loss is 100% of margin; an on-exchange stop at the same price pays 0.035% × 10 = 0.35% and keeps 3.3%. (3) The price is not yours — ✅ OKX docs: liqPx is the estimated mark price at which this position would be forcibly liquidated, and it can change quickly due to funding rate accrual: the trigger is the mark, not the last trade, and the engine's market order fills wherever the book is. (4) Funding walks the line toward you — a 10x long paying 0.01% per 8h for 30 days sees its liquidation move from −9.64% to −8.73% with the price unchanged. (5) Cascades — every liquidation is a market order into the next cluster of liquidation levels; the mechanism is impact, ../../../fin-core/skills/execution-cost-analysis/SKILL.md §5, and a fixed-slippage backtest cannot see it. stoploss_on_exchange (§6) is the retail defence; sizing so the liquidation price sits far outside the stop is the professional one.

12. Crypto options and inverse contracts

../../../fin-core/skills/derivatives-pricing/SKILL.md prices them, ../../../fin-core/skills/options-backtesting/SKILL.md owns the lifecycle; this is only what is crypto-specific. ✅ Deribit public/get_instruments, 2026-09-08:

• BTC-PERPETUAL: contract_size 10 USD, settled in BTC, instrument_type: reversed, max leverage 50x, maker 0.015% / taker 0.035%.
• BTC options: 908 listed, every one settled in BTC, contract 1 BTC, tick 0.0001 BTC, 11 expiries all at 08:00 UTC. ⚠️ European, cash-settled to the index — not fetched today.
• USDC-settled linear options exist too — 3,538 listed across SOL, BTC, ETH, HYPE, XRP, AVAX and TRX — and BTC_USDC-PERPETUAL is linear. "Deribit is inverse" is no longer the whole story.

Inverse P&L lives in the coin. Long N USD of an inverse contract: P&L = N(1/P₀ − 1/P₁) BTC (✅ the formula OKX's docs print for inverse upl). ✅ $100,000 at 78,449.67, 1x:

The USD column is exactly 2× linear: the collateral is itself long BTC, so a "1x long" on an inverse venue is a 2x long in dollars. The BTC column is asymmetric — gains cap at N/P₀ = 1.2747 BTC, losses are unbounded as the price falls. A 1x inverse short with BTC collateral is a synthetic dollar: equity constant at every price, no liquidation level at any maintenance rate. A BTC-settled call pays (S − K)/S BTC: 0.2 BTC at S = 100,000 for K = 80,000, approaching 1 BTC and never exceeding it, and the premium you paid in BTC is worth more dollars exactly when the call wins.

13. 🚨 Venue and counterparty risk is a first-order cost

• ⚠️ Exchanges fail with client assets on them: Mt. Gox (2014), QuadrigaCX (2019), FTX (2022 — withdrawals halted, then bankruptcy), Bybit's hot-wallet theft (2025). Short withdrawal pauses at solvent venues are routine. A backtest with all its capital on one venue has an unmodelled tail.
• ⚠️ Proof-of-reserves shows assets at a snapshot, not liabilities, and the accounting firm that produced several of the 2022 attestations withdrew from the work the same year.
• ✅ Access is jurisdictional. From this location on 2026-09-08 Binance's API answered HTTP 451 and Bybit's 403; OKX, Deribit, Coinbase, Kraken, Bitstamp and Gemini answered 200. A strategy that assumes a venue is assuming it will serve you.
• Survivorship has two layers — dead tokens (§2) and dead venues. Every bar is one venue's book; attribute it, and expect the venue with the best 2019 data not to exist.
• Keys: §6. Collateral: keep on-venue only what the strategy needs; the rest is counterparty exposure with no coupon.

14. Data: there is no official close, and "daily" means three things

✅ The close of 2026-09-07 as each venue's 1D candle labels it: Coinbase 79,091.97, Kraken 79,090.30, Bitstamp 79,090.41, Gemini 79,101.21 — 00:00 UTC bars, 1.4 bps apart. OKX 78,834.10: its 1D bar opens at 00:00 UTC+8, i.e. 16:00 UTC the previous day (✅ OKX docs — 1D is UTC+8, 1Dutc is UTC+0). Deribit 78,453.50: its 1D chart bars open at 08:00 UTC, its settlement hour. That is −32.6 and −80.7 bps from Coinbase's for the same date, and none of it is a venue difference. ✅ Across 60 closed UTC days (Coinbase/Kraken/Gemini) the cross-venue close range had median 1.77 bps, max 4.30 bps — small, non-zero, USD venues only; USDT pairs add the stablecoin premium. "00:00 UTC" is a convention you choose; a backtest of "the daily close" must say whose and when — the crypto restatement of fx-markets §4.

Funding history: ✅ OKX GET /api/v5/public/funding-rate-history (3 months), Deribit public/get_funding_rate_history (hourly, ≥ 420 days), Bybit GET /v5/market/funding/history (200 rows/call, docs). Binance GET /fapi/v1/fundingRate ⚠️ unverifiable from here. Candles: every venue's last bar is unclosed (✅ OKX marks it confirm=0) — lib-ccxt covers fetch_ohlcv.

15. Scripts and cross-links

scripts/perp_mechanics.py — every number in §8–§14 from constants fetched 2026-09-08 (endpoint named beside each) and one seeded synthetic series; verifies §8 against quantstats and empyrical when they are installed. No network at run time.

• ../../../fin-futures-fx/skills/fx-markets/SKILL.md — §1 the carry a spot backtest omits (FX twin of §9), §4 no official close
• ../../../fin-futures-fx/skills/futures-continuous-contracts/SKILL.md §5 — roll yield; a dated future rolls, a perp funds
• ../../../fin-core/skills/derivatives-pricing/SKILL.md — pricing, Greek scaling, vol surfaces
• ../../../fin-core/skills/options-backtesting/SKILL.md — assignment, expiry and margin lifecycle
• ../../../fin-core/skills/execution-cost-analysis/SKILL.md §5 — impact, the cascade mechanism
• ../../../fin-core/skills/broker-execution-apis/SKILL.md §3 — order safety (§6 here)
• ../../../fin-core/skills/signal-construction/SKILL.md §5 — causal indicators (§4 here)
• ../../../fin-libraries/skills/lib-ccxt/SKILL.md and ../../../fin-libraries/skills/lib-freqtrade/SKILL.md — the per-library deep dives