Does options flow actually work?
Numbers as of July 17, 2026 · updated nightly on the live Performance dashboard · every figure below is self-graded against real closing prices, never estimated
Most options-flow products show you the winners. We log every signal our scanner publishes — 79,245 so far — wait for real market closes, and grade all of them. Wins, losses, and the cohorts that turned out to be noise. This page is the summary; the app carries the always-current version with confidence intervals on every number.
The headline, honestly
Translation: buying every flow signal blindly is a coin flip. We publish that because it's true — and because it's the honest baseline that makes the cohort-level findings below meaningful instead of marketing.
Where edge shows up
The signal isn't "flow good." It's which flow, held how long. Cohorts with statistically meaningful spreads on our ledger (each shown with sample size and confidence in the app):
| Cohort | Win rate | vs. baseline | n |
|---|---|---|---|
| Positioning-window DTE (2–8 wk) | 50.0% | +4.2pp | 4,182 |
| Flow into earnings | 48.5% | +2.8pp | 3,330 |
| Golden sweeps (+5d vs SPY) | 47% | +0.40% excess | 18,358 |
| Highest-liquidity names | 41.3% | −4.0pp | 3,703 |
| Small caps (recent window) | 17.7% | collapsed from 50% | 141 |
Those last two rows matter most: the most-watched flow is the most faded, and small-cap flow edge died in the recent tape — our decay tracker caught it and demoted the cohort automatically.
What we've proven doesn't work
- Raw signal scores don't rank. The rank correlation (IC) of our own trade score vs realized returns is −0.03 — statistically nothing. We publish that on the dashboard and treat the score as an admission filter, not a ranking. Learned cohort adjustments carry the ranking instead.
- Holding long kills it. +5d excess is ~zero; +10d is −1.0%; +20d is −1.9%. Flow information decays in days. No stop/target combination we tested makes 20-day holds positive.
- Most hypotheses fail. Our hypothesis lab has auto-tested 268 feature combinations against out-of-sample data. 258 are in the graveyard. 10 survived, and each survivor ships with its sample size and OOS numbers attached.
How the grading works
- Every published signal is logged the moment it appears — no retroactive additions, no deletions.
- Returns are graded against real closing prices at +1/+3/+5/+10/+20 days, benchmarked vs SPY.
- Next-day open-interest confirmation tells us whether flagged positions were actually opened and held — a check most flow products skip.
- Cohort statistics gate at minimum sample sizes and ship with 95% confidence intervals. Cohorts below the gate say "accruing," never a made-up number.
- A nightly calibration check compares trained expectations vs the live tape and publishes the drift — currently −2 to −8pp depending on band, shown in the app.