Tests whether Smart Money Concepts and Inner Circle Trader trading concepts — translated directly from two LuxAlgo Pine Script indicators into Python, line by line, as one combined methodology rather than two separate ones — provide statistically significant, out-of-sample-honest trading edges on daily S&P 500 OHLCV from 2010-01-01 through 2026-06-13.
Not an attempt to replicate the indicators' visual output; boxes, lines and labels are irrelevant. Only the underlying trigger logic is tested.
Pipeline: pine_parser/ (shared primitives — pivot detection, leg and swing detection, Wilder ATR) → signals/ (every ICT and SMC detector across the full universe) → backtest/ (forward returns with no look-ahead, a full metrics suite, and multiple baselines: buy and hold, random entry, EMA crossover, RSI mean reversion, 52-week breakout, momentum) → analytics/ (bootstrap confidence intervals, significance testing with FDR correction, per-stock, per-concept, per-combination, per-regime and per-sector ranking) → dashboard/ (Gradio and Streamlit front ends).
Superseded by a published re-analysis
This repository once reported that "28 of 42 concepts survive a random-entry baseline" — a red flag caught early, since all 42 had cleared a much weaker zero-return null, which led to building a dedicated baseline-comparison engine in the first place. That 28-of-42 figure was itself wrong: a two-sided significance test had actually found 27 of those 28 concepts to be significantly worse than random entry, not better, and a random-entry baseline seeded from an unstable hash meant the count could not even be reproduced run to run. Both errors, plus a look-ahead outcome label stamped on the wrong bar, are documented in full in a follow-up paper published on SSRN, "Do Smart Money Concepts Predict Returns?" (September 2026), which formalizes 44 detectors rather than 42, replaces the flawed test with a calendar-time test validated against a purpose-built calibration study, and finds that none of the 44 beats a composition-matched random-entry baseline at any of eight holding horizons. Two rounds of self-correction, both on the commit history — this dashboard predates the second.