How easy is it to "discover" a winning trading strategy that is pure luck? Trivially easy — and this repo lets you feel it.
▶️ Live demo
Pull the lever: it searches hundreds of moving-average crossover rules, keeps the one with the best in-sample Sharpe, then tests it out-of-sample. Do it on a pure random walk (where no edge can exist by construction) and the "winner" still shows a Sharpe above 1 — then collapses to ~0.
Not financial advice — the exact opposite. The lesson is that a strategy selected from its own backtest tells you almost nothing about the future.
Searching K strategies and reporting the best in-sample Sharpe is multiple-hypothesis testing. The more rules you try, the higher the best in-sample Sharpe climbs — even on noise. Averaged over many random seeds (K = 200 rules):
| Data | Best in-sample Sharpe | Out-of-sample Sharpe | Deflated Sharpe |
|---|---|---|---|
| Pure random walk (no edge) | ~0.75 | ~0.08 | ~0.42 |
| Market-like noise | ~0.80 | ~0.11 | ~0.43 |
The Deflated Sharpe Ratio (Bailey & López de Prado, 2014) prices in the number of trials, the track-record length, and the skew/kurtosis of returns. Below ~0.95 it's statistically indistinguishable from luck — and the casino's "winners" never clear the bar.
Across 200 independent searches on pure random walks, the in-sample "winner" Sharpe and its out-of-sample Sharpe correlate ~0.07 — i.e., not at all. The deflated Sharpe falls as you search more rules.
pytest asserts the mirage reproduces: in-sample beats out-of-sample by a wide margin, out-of-sample Sharpe ≈ 0 on a random walk, and the Deflated Sharpe never clears 0.95. GitHub Actions runs it on every push.
experiment.py— simulates returns, runs the MA-crossover grid search, computes in-sample vs out-of-sample Sharpe, the expected-maximum Sharpe under the null, and the Deflated Sharpe Ratio; aggregates over 30 seeds.index.html— a fully client-side re-run of the same experiment (seeded RNG, grid search, deflated Sharpe with an Acklam inverse-normal); the in-browser numbers reproduce the Python experiment's statistics.
pip install -r requirements.txt
python experiment.py