dsr-gate 0.1.0

Certifier — an honest backtest gate whose default answer is NO. Deflated-Sharpe + ruin-floor certification.

Certifier — an honest backtest gate whose default answer is NO

Most backtest tooling is built to make a strategy look good. This does the opposite: it is engineered to reject a backtest — including its authors' own — unless the edge survives every discount a skeptical statistician would apply.

Free and MIT-licensed. numpy + the Python standard library, nothing else.

Try it in 30 seconds (browser, nothing uploaded)

https://rocomas.github.io/certifier-demo/ — the whole gate runs client-side. Click Load an overfit, then Certify, and watch the best of 400 pure-noise strategies get REJECTED.

What it checks

Feed it a return series and the honest size of the search that produced the strategy. It returns CERTIFIED only if:

1. Deflated Sharpe ≥ 0.95 (Bailey & López de Prado), deflated for
   • the best-of-N selection that produced the edge — a bigger search raises the bar,
   • the return distribution's skew and kurtosis,
   • an optional publication-decay haircut (--haircut 0.5 for known anomalies),
   • an optional overlapping-window correction (--overlap-lag h-1);
2. it survives a ruin floor — a $1 book compounding the returns never draws down past the floor (default 51%).

Both must hold. Thin samples, overfit selections, decayed edges, and ruinous paths are all rejected.

Install

pip install dsr-gate                 # PyPI name: dsr-gate; import name: certifier
# or straight from source:
pip install git+https://github.com/Rocomas/certifier

(The PyPI name certifier was already taken by an unrelated TLS package — hence dsr-gate. The module and CLI are both called certifier.)

Use

certifier returns.csv --trials 200                 # one return per line; exit 0/1 for CI
certifier returns.csv --trials 200 --haircut 0.5 --overlap-lag 4
echo "0.01,0.02,-0.005" | certifier - --trials 50
python -m certifier.webdemo                        # local copy of the browser demo

from certifier import certify_returns
res = certify_returns(my_returns, n_trials=200, haircut=0.5, overlap_lag=4)
print(res["verdict"], res["dsr"], res["ruin"]["survives"])

Reproduce the money shot

python examples/demo.py     # dredges the best of 500 noise trials; the certifier kills it

overfit   sharpe=+0.17   DSR=0.360/0.95   ruin_ok=True   ->  REJECTED
real      sharpe=+1.96   DSR=1.000/0.95   ruin_ok=True   ->  CERTIFIED

Same family of Sharpes, opposite verdicts. examples/overfit.csv and examples/real.csv are committed so you can verify from the CLI.

The honest caveat (and what's coming)

The verdict is only as honest as your n_trials. Most researchers genuinely don't know how many strategies they tried — every discarded variation counts. We're building the layer that fixes that: a trial ledger that hooks your research loop, records every backtest automatically, and certifies against counted trials, not claimed ones — plus a CI gate and a signed certification report. The calculator here is free forever; the ledger/workflow layer will be a paid tier. Watch/star this repo to hear when it lands.

What it is not

It does not tune, fit, or optimize anything. A CERTIFIED verdict is a statement about historical evidence, not investment advice and not a guarantee of future returns.

License

MIT. Math: Bailey & López de Prado, The Deflated Sharpe Ratio (2014), implemented inline in ~200 lines — audit certifier/dsr.py and certifier/gate.py yourself.