sharpe-gate

Everyone computes a Sharpe ratio. Nobody stores the honest trial count behind it. Without that count N, every multiple-testing statistic — the Deflated Sharpe Ratio above all — is decoration: it deflates by however many trials you remembered running, not by how many you actually ran across every session, notebook, and parameter sweep.

This package is the machinery that keeps N honest, and the gates that spend it:

• Trial registry (sharpe_gate.registry) — every backtest persists one append-only row keyed by a code-agnostic config_hash, plus its daily PnL/exposure stream. SQLite (WAL); concurrent writers wait for each other instead of failing, including on the first touch of a fresh store. Studies are pre-registered (declared trial count, kill threshold) before the search; you cannot record a trial without one.
• Canonical dataset fingerprints (sharpe_gate.fingerprint) — the pooling key fingerprints the data, never the hypothesis, and is validated structurally at the write boundary. A spelling drift or a hypothesis suffix cannot carve a private, smaller-N pool out of shared data. Venues ship as prefixes; register_source("bybit") adds your own without turning the validation off.
• Overfit gate (sharpe_gate.overfit) — DSR (Bailey & López de Prado) on the registry-resolved pool, PSR fallback for N=1, expected-max-Sharpe under H0, PBO via CSCV on purged+embargoed folds, and a shuffled-label held-out AUC floor for ML overlays. The entry point takes a trial_id, not a bare float — an unrecorded Sharpe cannot reach the gate. Kurtosis is Pearson (normal = 3.0) throughout; passing excess kurtosis raises rather than quietly inflating the result.
• minBTL (sharpe_gate.min_btl) — the cheapest kill: is the sample even long enough to support the hurdle Sharpe given N? Pure arithmetic, runs before any backtest.

None of this contains alpha. It is machinery against self-deception — the part that tells you the unpleasant truth before you pay for it.

Install

pip install sharpe-gate

Quick start

from sharpe_gate import register_study, record_trial, overfit_screen

study = register_study(
    study_id="my_screen_v1",
    hypothesis_id="H42",
    generator="grid",
    dataset_keys=["binance_btc_perp_4h"],
    statistic="sharpe",
    n_declared=24,
    kill_threshold=0.95,
)
trial_id = record_trial(study, config={"lookback": 20}, stream=(daily_pnl, in_position))
result = overfit_screen(trial_id)   # -> OverfitResult(probability, method="dsr"|"psr", ...)

The store location resolves in order: explicit store_dir= argument → SHARPE_GATE_DIR env var (HL_REGISTRY_DIR is honored as a legacy alias) → sharpe_gate.registry.DEFAULT_STORE_DIR (settable by a host repo) → ./data/registry under the current working directory.

What it does not do

It contains no alpha, no strategies, and no data. It never asks the network, holds no keys, and cannot place an order. The registry is a local SQLite file plus Parquet stream sidecars.

Development

git clone https://github.com/bond-labs-dev/sharpe-gate
cd sharpe-gate
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest -q

See CONTRIBUTING.md for the checks CI runs and the invariants a change must not break.

Status

Alpha. The API is stable in practice but not promised across 0.x.

License

MIT — see LICENSE.