Every edge on the catalog started here. Some make it. Some don't. This page is the log — open research, rejected hypotheses, and edges graduating next. Updated as runs complete.
buy the open after a bottom-quartile close, Mon-Thu — 6/6 walk-forward windows, p=0.001
7-year mine on 1,784 RTH sessions: n=307, 58.6% WR, PF 1.51, +27pt/trade net. Permutation-tested against matched random buy days. Scale-outs and breakeven stops tested and REJECTED — the edge is the fat right tail. Live shadow tracker posting daily since 2026-07-08. Short mirror rejected (bear-window-only).
which factors predict gamma-level trade outcomes: flow, regime, breadth, SPY agreement
Running on live alerts (n≈50 and growing). Early read: flow strongest single factor; flow+regime cell ≈ 90%+; ≤1 aligned factor = no-trade. Thresholds feed the alert scoring; re-weighting deferred until out-of-sample n≈100.
two survivors of a 62-variant Bonferroni-gated mechanism campaign — on watch, not shipped
Both passed 5/6 walk-forward folds but carry small samples. Watching live before any catalog entry. 60 sibling variants killed in the same campaign, including the Monday-after-OPEX drift (artifact of Monday drift generally).
does true holding absorption (tf_* features) predict rejection moves?
Original March study flagged as regime artifact; scheduled for re-test with fresh tick-features data once the live engine relaunches. The absorption polarity booster (SELL+absorption) remains in shadow validation.
VAH/VAL/POC location, dynamics, and LVN reactions — all fail placebo-controlled testing
5,864 level touches tested against shifted fake levels. Location effects = placebo-level. Legacy prev-week VAH/VAL sells (once on this site at 69.8%) = generic rally-fade artifact, retracted from the catalog. LVN bounce claims didn't reproduce under clearance-gated first-touch testing; traversal-velocity claim null. Entire family closed — we won't re-test it.
trend-flips, oscillators, squeezes, ICT combos — ~200 walk-forward tests, zero ship
Every popular mechanism family walk-forwarded on MNQ with drift-guard controls and 1m-path fills. 0 of 102 survived. An ML interaction hunt (56 features, 6-fold WF) also died: AUC ~0.55, top decile loses after costs. If it's on every YouTube channel, assume it's priced out.
OR/IB both-sides break setups, 18 variants — all lose to time-matched random entries
Every published version tested (p≈1.0 vs random). The '70% resolve by noon' premise measures ~31-37% on MNQ. One descriptive keeper: IB close-half predicts break side (~80%) — a prior, not a trade.
30-50pt down impulse + counter ≤15pt — retired from the catalog
Original validation had an outcome-tracker bug; live replay measured ~41% WR at 30/20 brackets — negative expectancy after costs. Removed rather than re-tuned.