Leakage screens
Purged, embargoed chronological folds plus truncated-tape and correlation screens — future data can't leak into past tests.
Now in early access
QUANT RESEARCHFOR CRYPTO PERPETUALS
Research → Backtest → Paper → Live
One disciplined pipeline for crypto perpetualsBuilt for the people who measure before they act.
A pattern earns the catalog by surviving a gauntlet — not by looking good once.
Nine live data lanes from Binance and Hyperliquid feed one pipeline: collect, label on two decision horizons, compute 26+ features daily, and validate through eight fail-closed gates. The verdict is net post-cost Sharpe on two independent holdouts — everything else is a diagnostic.
CONNECTED VENUES
Live data across both,
execution on your own accounts.
01 / Collect9lanes
02 / Label2horizons
03 / Features26+computed
04 / Gauntlet8gates
05 / Catalogvalidatedsurvivors
THE PIPELINE
01 / Collect
9 data lanes
Nine live streams from Binance and Hyperliquid, across twelve perpetual-futures symbols. What the venue printed is what gets tested.
INTAKEBINANCE · HYPERLIQUID · 12 PERPETUAL-FUTURES SYMBOLS
01Tradesevery print the venue tape producedtick-levelRAW
02CandlesOHLCV aggregated from the trade tape1m · 5m · 1hDERIVED
03Top of bookbest bid and ask, continuously sampledcontinuousRAW
04L2 depthresting size beyond the touchsnapshotsRAW
05Fundingevery settlement, to the millisecond8h · 4hRAW
06Open interestoutstanding contracts across the venue5m · 288/dayRAW
07Liquidationsforced exits, as they firedevent-drivenRAW
08Oracle pricesthe venue's resolved reference price5m · resolvedRAW
09Mark-index basisthe perp's premium over its index1m · 1440/dayDERIVED
RAW captured as the venue printed itDERIVED computed on our side from the raw feeds
THE GAUNTLET
Eight ways to get rejected
Each gate fails closed: a candidate without complete evidence is refused, and the serving boundary refuses anything without a passed packet. These are rejection bars — not promises.
Factor concentration
If one latent factor explains more than 80% of the variance across symbols, the candidate is rejected.
Decision dossier
Reference class, base rate, pre-mortem, and the economic rationale — written down before the verdict, not after.
Capacity ceiling
Per-symbol notional ceilings sized to real book depth — a pattern must hold up at the size customers actually run.
Regime scorecard
Scored across volatility, trend, and liquidity-stress regimes — no single-weather patterns.
Overfitting controls
Deflated Sharpe and combinatorial overfitting probability on purged folds — luck gets discounted, not rewarded.
Trial ledger
Every attempt is counted in a persistent ledger. The more we try, the higher the bar automatically gets.
Dual-tier post-cost Sharpe
The decisive test: portfolio Sharpe net of fees, funding, and slippage — on two independent holdouts, both re-run at 1.5× costs.
CURRENT BARS · VERSIONED IN THE DOCS
Both tiers must pass, and both must survive a 1.5× cost-stress re-run. Fail either, and the candidate never reaches the catalog.
Tier 1 · recent & broadSharpe ≥ 0.70 · DD ≤ 15%
Three-month holdout across all twelve symbols — the first bar every candidate must clear.
Tier 2 · deep & longSharpe ≥ 0.50 · DD ≤ 20%
Six-month holdout on the longest-history symbols — consistency over time, not just lately.
DIAGNOSTICS — THEY EXPLAIN, NOT DECIDE
IC rank correlation with the forward outcomeIC IR consistency of IC across foldsPositive folds share of folds above zero
Test a candidate against history — and get the same answer twice.
Run a candidate against frozen historical data — pinned to a manifest, partition IDs, and the exact commit, so any result is re-runnable and auditable. Empirical evidence, not a forecast.
Deterministic replay
Same candidate, same partitions, same commit — same equity curve. No hidden randomness, no look-ahead.
Reproducible artifacts
Equity curve, drawdown, per-trade ledger, and behaviour by regime — pinned to the run that produced them, so anyone can verify.
Variant sweeps
Sweep configurations in one pass, rank by post-cost Sharpe, stress under heavier costs. Survivors earn the next rung: paper.
From paper to live — on your own accounts, on your terms.
A validated candidate climbs a one-way ladder — paper, then testnet, then live — only when you advance it. Statly is non-custodial: your own API keys, no withdrawal permission, no discretionary trading on your behalf.
Paper first
Run forward on live data with realistic costs before a dollar is at risk — then testnet against the exchange’s own matching engine.
Your keys, your control
Your own Binance and Hyperliquid keys, scoped without withdrawal. Statly never holds your funds — not an exchange, broker, or wallet.
You decide
Nothing is promoted to live automatically. Statly runs only what you activate, inside the limits you set — never at its own discretion.