SSRN · Working paper · September 2026

Do Smart Money Concepts Predict Returns? Evidence from 44 Formalised Detectors on S&P 500 Daily Bars, 2010–2026

Harsha Tangirala · Independent Researcher

Translates two widely used open-source Pine Script implementations of Smart Money Concepts (SMC) and Inner Circle Trader (ICT) methods into 44 formally specified event detectors, tested on daily data for 496 S&P 500 constituents from 2010 to 2026 — 2.20 million events and 17.6 million simulated trades across eight holding horizons — measured against each ticker's own unconditional return with calendar-time inference, calibrated against a known-zero-edge baseline.

Detectors formalised

44

S&P 500 constituents

496

Simulated trades

17.6M

Sample period

2010–2026

• Zero concepts beat a fair benchmark

  Against a composition-matched null — entering the same tickers the same number of times on random dates — with false-discovery-rate control across all 352 concept × horizon hypotheses, none of the 44 detectors beats random entry at any horizon, and none is significantly worse.

• The benchmark decides the headline number

  Against a naive zero-return null instead, 38 of the same 44 concepts appear significant — a benchmark that credits any long-biased signal with 16 years of bull-market drift it did nothing to earn.

• A corrected re-analysis

  An earlier version of this pipeline reported that 28 of 42 concepts beat random entry. A two-sided test had in fact found 27 of those 28 to be significantly worse than random entry, not better; the paper documents this and two further mechanical errors — a look-ahead outcome label and an unseeded random baseline — that each converted a null result into an apparently positive one.

• Selection does not survive out of sample

  Concepts ranked best on three years of training data earn −31 basis points the following year, barely better than an unselected concept's −35 bp, with a train-to-test rank correlation of just 0.32.

• Power, not certainty, is the limit

  For ten concepts the data rule out, at 95% confidence, any edge large enough to cover even the lowest modelled trading cost. For the two best-measured concepts — both liquidity-sweep detectors — a true edge of 30–50 basis points cannot yet be ruled out.

Gross returns from these detectors are large because they are long a rising market, not because they carry information — the same random-entry baseline earns as much. The apparent edge in an earlier version of this analysis came from benchmark choice, test direction, and an uncalibrated variance formula, not from the underlying signals.