Crypto market structure
Name the venue, instrument, quote currency and bar boundary before comparing prices or returns. Continuous crypto trading is a calendar convention; it does not guarantee continuous liquidity, uninterrupted exchange access or a single settlement convention for every derivative.
All measurements below were reproduced on 2026-09-14 by scripts/market_structure.py, seed
20260910. They use synthetic prices and are not empirical estimates of venue behavior.
Annualization follows the observation frequency
Use fin_skills.api.conventions.annualization_factor() and annualize_sharpe() when the
package is installed. The standalone script falls back only when fin_skills itself is absent;
it does not silently mask broken package imports.
For ordinary calendar-day crypto returns, the library's annualization convention is 365. It uses 8,760 for hourly observations and 1,095 for eight-hour funding observations. A weekday strategy still needs weekend portfolio returns accounted for if risk is measured in calendar time. Choosing 252 merely because an equity library defaults to it is not a conversion.
✅ Measured on the same 1,460 calendar-day rows:
| Statistic | Factor 365 | Factor 252 |
|---|---|---|
| Sharpe, mean / sample SD times square-root factor | 0.8432 | 0.7006 |
| Annualized volatility | 65.81% | 54.69% |
Their ratio is sqrt(252/365) = 0.8309; the 252 result is 16.9% smaller for this positive
Sharpe. For negative Sharpe the magnitude shrinks rather than performance becoming worse.
For a series with 252 observations per year, applying 365 inflates magnitude by 20.4%.
Dropping weekends alone does not produce that frequency: the script's weekday subset has
no exchange-holiday filter, so its 252/365 comparison is sensitivity analysis, not a claim
that either factor matches the subset's actual observations per year.
These are conventional scaling identities, not proof that returns are independent or that
square-root scaling is a valid forecast at longer horizons.
Rolling rows are not calendar durations
✅ Measured: on this synthetic daily series, 252-row and 365-row momentum calculations disagree on sign in 17.7% of 1,095 common observations; correlation is 0.7791. A 252-row lookback on an unbroken calendar-day series does not span a year. Missing observations require explicit coverage checks; silently filling outages with zero returns changes both the signal and risk.
✅ Measured: with an assumed desk watching 13:00–21:00 UTC on weekdays, 1,926 of 2,000 hourly paths cross a 5% downside level. Of those first crossings, 71.3% occur outside desk hours. This is a homogeneous-volatility example, not evidence that real activity is uniform by hour. Price jumps and local reopening gaps remain possible even in a continuously traded asset.
There is no single venue-independent price or daily close
✅ Measured: five synthetic venue feeds of one latent asset over 730 days have median cross- venue price range 21.4 bp, 99th percentile 43 bp and maximum 48 bp. A lagged thirty-day momentum rule produces a 15.4 percentage-point return spread across those feeds before transaction costs. Noise near the entry threshold changes trades; these are not executable arbitrage profits.
The same hourly path sampled at different daily boundaries produces Sharpe values from −1.2856 to −1.2473. Both the intervals and sample endpoints differ. Align UTC opening/closing instants, quote currency, candle completion, instrument type and venue before calling a difference a premium. Historical OKX and Deribit boundary examples in the script retain the original 2026-09-08 observation date; confirm the chosen endpoint's current bar convention.
Venue failure, socialized losses and ADL
✅ Source-verified 2026-09-14, web text: Bybit's ADL rules describe automatic reduction of opposing positions and settlement at a bankruptcy price. The current trigger conditions include insurance-fund drawdown thresholds, so fund exhaustion is not a universal prerequisite. ADL is position reduction; a generic account-wide haircut or socialized-loss allocation is a distinct mechanism. Read the venue's instrument and account rules rather than treating all of them as the same fee.
✅ Measured illustration: a short entered at 30,000 and marked at 21,000 shows a 30% gain. If it is instead force-closed at an assumed 27,000, the gain is 10%: two-thirds of that marked gain is given up. This is a settlement-price example, not an ADL probability or a venue forecast.
✅ Measured: with assumed mean four-hour outages, the simulated fifth-percentile price move while unable to act is −2.39%. Summing only negative moves over six assumed annual outages gives −3.33%; that is not expected annual drag, because favorable moves have been discarded. The function exposes the unconditional mean separately. Real outages can coincide with stress, so independent synthetic returns are insufficient to price exchange exposure.
Stablecoin depegs change the unit of account and collateral value
For an asset worth P_USD and a stablecoin worth S_USD, the stablecoin-quoted price is
P_USD / S_USD. An apparent asset gain can therefore come from the quote currency falling.
Convert cash balances, fees, funding and collateral with the same timestamped exchange rate.
A depeg can reduce borrowing capacity or create a financing need; spot conversion alone cannot
estimate actual borrow rates, liquidation haircuts or funding costs.
✅ Measured: the synthetic six-day depeg reaches a stablecoin value of 0.917 USD and introduces a 9.0% peak quote-conversion premium. After recovery the window's conversion effect is 0.00%, yet annualized tracking error between USD and stablecoin returns is 1,284 bp. High daily correlation, 0.9805 here, does not remove intraperiod collateral risk.
✅ Source-verified 2026-09-14, web text: Circle's March 2023 notice reported $3.3 billion of USDC reserves at Silicon Valley Bank, about 8% of total reserves, and subsequent availability of those deposits. This is a dated banking event; the script's depeg path is not fitted to the historical USDC price.
Scripts and routing
Run python plugins/fin-crypto/skills/crypto-market-structure/scripts/market_structure.py.
Functions cover annualization, rolling windows, off-desk breaches, cross-venue dispersion,
sampling-hour effects, hypothetical outages/forced closes and stablecoin conversion. They
perform no network access, credential handling or order submission.
../crypto-data-and-execution/SKILL.md— clients, candles and exchange access.../crypto-token-events/SKILL.md— historical universes and token identity.../perpetuals-and-funding/SKILL.md— funding, margin and liquidation triggers.../defi-and-amm-mechanics/SKILL.md— LP exposure when a stablecoin leg changes value.../../../fin-core/skills/backtest-validation/SKILL.md— return definitions and Sharpe limits.../../../fin-core/skills/broker-execution-apis/SKILL.md— execution state and reconciliation.