The bar, written down.
Running thousands of trials against one price history will always produce something that looks like an edge. The entire design problem is telling that apart from an edge that is real.
Every threshold below is read from the engine's live configuration when this page is served. If a gate moves, this page moves with it.
Below this a strategy is broken, not merely weak: Sharpe ≥ 0.2, max drawdown ≤ 40%, 10–500 trades, win rate ≥ 20%, profit factor ≥ 0.8.
Worth putting on the paper book: Sharpe ≥ 1.2, max drawdown ≤ 20%, 30–300 trades, win rate ≥ 35%, profit factor ≥ 1.3.
Every backtest reserves a trailing 252 days it never sees. Holdout Sharpe must clear 0.3 and reach 40% of the full-window figure. A strong full window with a negative holdout is the overfit signature — a hard reject, not a warning.
An edge has to beat market beta, not borrow it. Annualized alpha against SPY must be positive, and |beta| above 3.0 is treated as a leverage artifact rather than a result.
The window is split into 3 contiguous regimes and the worst one must hold Sharpe ≥ -0.5. A strategy that only works in one market is a strategy that has memorised one market.
5 purged folds with a 5-day embargo at each edge; the worst fold must clear 0.0 across at least 2 informative folds. Sharpe is then deflated against a 8,000-trial budget — because that is roughly how many attempts stand behind any one result.
Minimum 21 days on the paper book, reconciled daily against real fills. Paper Sharpe must hold 65% of backtest Sharpe and drawdown must stay within 1.5× the backtest figure.
Of 20,612 strategies evaluated, 18 have cleared everything above — 0.09%. The gates are mandatory, not weighted, so a strategy cannot trade a failed one against a strong result elsewhere.