SML EDGE COURT

Methodology edge-court-1.0 · calibrated:false

Every verdict follows the rules on this page. Thresholds are stated design values, not fitted to a benchmark. Every number in a report comes from the caller's real trades or real market data.

Verdicts

VerdictRule
ROBUST EDGEscore >= 75 AND full_history_edge, statistical_significance, temporal_stability pass AND parameter_robustness passes when applicable
CURVE-FITprofitable on full history, but a parameter-spike pattern or a non-positive walk-forward out-of-sample result shows the profit depends on the exact settings
FRAGILE EDGEprofitable on full history but fails one or more robustness requirements
NO EDGEnot profitable on full history
NOT ENOUGH DATAtoo few trades or too short a span to judge

Gates

INSUFFICIENT_DATAfewer than 30 closed trades, or less than 365 days between first entry and last exit
NO_EDGEfull-history net profit <= 0 or profit factor < 1.0

Scored checks

CheckWeightPass rule
full_history_edge20full-history profit factor >= 1.10 and net profit > 0
statistical_significance15Probabilistic Sharpe Ratio (per-trade, vs 0) >= 0.95
temporal_stability15>= 60% of 5 chronological folds are net-profitable AND >= 60% of calendar years with >= 5 trades are net-profitable
monte_carlo10stationary block bootstrap probability of net <= 0 is <= 10%
regime_breadth10profit factor > 1 in at least 2 trend regimes (BULL/BEAR/RANGE) that each hold >= 10 trades; skipped without price history
parameter_robustness15>= 70% of +/-10-20% input perturbations keep profit factor > 1 and no isolated-peak (spike) pattern; skipped for trade-list audits
walk_forward0informational unless parameter sweep exists: anchored walk-forward out-of-sample net must be > 0 (failure forces CURVE_FIT)
profit_not_concentrated5final 10% of the time span contributes < 50% of net profit
not_underwater5cumulative P&L is below zero on < 50% of trades
win_rate_vs_payoff5win rate exceeds the payoff-implied breakeven win rate by >= 3 percentage points

Score = weighted share of applicable checks passed. Checks that cannot apply (no price history, no parameters) are dropped from the denominator, never counted as passes.

Statistics

profit factorgross profit / gross loss over closed trades (net of the commissions applied by the strategy or platform).
snapshot windowRolling calendar windows stepped daily; window length holds ~12 trades on average (min 28 days, max 1/3 of history); windows need >= 8 trades. The best window by PF is the "snapshot"; share_of_windows_pf_ge_2 is how often a random snapshot would have looked like a monster.
psrProbabilistic Sharpe Ratio, Bailey & Lopez de Prado (2012), per-trade Sharpe vs 0 with skew/kurtosis correction.
dsrDeflated Sharpe Ratio, Bailey & Lopez de Prado (2014): expected max Sharpe under the null across N trials. N = completed sweep settings + caller-declared prior trials; variance from the sweep when available.
min trlMinimum track record length (trades) for 95% confidence that the true Sharpe > 0.
folds5 chronological folds with equal trade counts.
walk forwardAnchored, 5 equal time folds; at each step the setting (base or sweep) with max in-sample net profit is applied to the next fold. Uses only already-computed real runs.
parameter sweepUp to 4 numeric strategy inputs (visual/display inputs excluded) perturbed by x0.8, x0.9, x1.1, x1.2 (ints rounded, min/max respected).
bootstrap5000 iterations, deterministic seed from the trade P&L; iid and stationary block (mean block length = cube root of trade count).
cost stressNet and PF after subtracting 5/10/25% of the median absolute trade P&L from every trade; breakeven extra cost = mean trade P&L.

Regimes

trendDaily close vs SMA(100); SMA slope over 20 days. BULL: close>SMA and slope>0. BEAR: close<SMA and slope<0. RANGE: otherwise.
volatility20-day realized volatility of daily log returns, split into sample terciles (LOW/MID/HIGH).
lookaheadA trade is labelled with the regime of its entry day, computed from data up to the previous daily close.
warmup_days120

Data

Binance spot klines via https://data-api.binance.vision (public market-data mirror). Only closed bars. SHA-256 of the exact bars is in every receipt. Caller-supplied OHLC CSV, hashed into the receipt; provenance is the caller's. Bar cap: 30000.

Execution

Pine runs in PineTS (AGPL-3.0, LuxAlgo) inside a sandboxed Node child process (permission model: no fs writes, no child processes; network globals removed; memory cap; hard time budget).

Receipts

Every audit returns an Ed25519 signature over a canonical JSON receipt (verdict, score, key metrics, input hashes, data hash, report hash). Public key: /.well-known/edge-court-keys.json.

Limits