The strategies we killed.

Every signal provider has a graveyard. Ours is public. Below are the trading edges we tested with real data, real walk-forward, real bootstrap CIs — and decided NOT to ship to subscribers because the math didn't survive scrutiny. This is the work most providers hide. We lead with it because honest failure data is how trust is built.

12
Hypotheses killed
2
Look-ahead bugs caught
7
Survivors deployed
~32%
Hypothesis pass-rate

METHODOLOGY BUGS WE CAUGHT

Volume-Profile σ-Imbalance Fade × GC

2026-04-25 · GC · RETRACTED — LOOK-AHEAD BIAS

Original sweep produced a striking result on Databento ohlcv-1s 2-year sample.

ORIGINAL CLAIM PF 2.47 · CI [1.48, 4.06] · n=180 · 6/6 anchored WF folds positive · cross-symbol replication on NQ

Pre-deploy audit caught the bug: the High-Volume-Node bins for "today" were computed using today's full-day volume. At backtest time T (e.g. 13:00 ET), the σ-imbalance feature could see all of today's trading — including bars that hadn't happened yet at T.

CLEAN REPLAY (yesterday's HVN bins only) PF 0.84 - 1.11 across 5 configs · all sub-gate · cross-symbol replication was the same bias propagating

Lesson: Anchored walk-forward folds DO NOT catch within-day look-ahead. Every fold inherits the same bias. We now require every signal feature to be "computable at bar T from bars 1..T" with explicit audit of groupby-day transforms, centered rolling windows, bfill, and joins on day-keys.

Aramid Pro EMA-Cross Indicator

2026-04-25 · NQ · RETRACTED — LOOK-AHEAD BIAS

Pine v6 indicator imported from a TradingView source. 1-year backtest showed PF 1.38 — looked deployable.

4-YEAR REVALIDATION Code had implicit look-ahead bias worth ~$680k/yr in claimed but non-realizable PnL. Killed before any DLL was built.

RESEARCH HYPOTHESES — KILLED

orbFade × ES (intraday breakout fade)

2026-04-22 · ES · RETRACTED — DATA CORRUPTION

Initial sweep showed PF 1.82 on a wide range of bracket configs. Looked like the foundation of a fade-fleet.

CLEAN DATA REPLAY PF 0.63 across the same configs — a NET LOSS. Original result had been inflated by corrupted ES bars (one had range = 3,118 points). We added hygienic-bar filtering to load_bars() so this class of bug can't recur.

L2 Imbalance Fade (TopStepX SignalR)

2026-04-22 · NQ · REAL BUT NOT EXPLOITABLE

Found a statistically real L2-imbalance signal. Zero-slippage simulation: avg +$0.62 per trade. Smaller than the NQ bid-ask spread. HFT-only edge. Not retail-tradeable on a 100ms TopStepX pipe.

GC London → NY Handoff Fade

2026-04-22 · GC / ES / NQ · NULL ACROSS ALL 3

Hypothesis: London session momentum mean-reverts when NY opens. Tested directly across GC, ES, NQ. None passed gate. Killed entire seam.

boxRange × all symbols

2026-04-22 · NQ / ES / GC · NEGATIVE EXPECTANCY

Range-breakout strategy. Negative expectancy across 3,642 trades. Tuning was not the answer — the core signal had no edge.

ORB Breakout (Open-Range)

2026-04-24 · NQ / ES / GC · 1,728-CONFIG SWEEP NULL

Comprehensive parameter sweep — 1,728 configurations. Zero passed the rigorous gate. A phase-1 result with CI_lo 1.14 did NOT replicate under clean methodology. Strategy permanently disabled.

NQ → GC 5-Min Lead-Lag

2026-04-25 · NQ → GC · NULL AFTER COSTS

Hypothesis: NQ momentum at 5-min bars predicts GC return. Tested on 2-year Databento ohlcv-1s sample, 39,091 RTH bars.

RESULT Strongest signal (NQ z > +1.5, 60-min forward): t-stat = +4.16 (highly significant statistically) BUT mean magnitude was only 3.78bp ≈ $1.70/trade on GC vs ~$20 transaction cost. Net-negative after costs.

CME Close-Location × Symbol × Timeframe

2026-04-XX · NQ / ES / GC · 258 days × 4 timeframes — NULL

Full sweep: 258 trading days, 3 symbols, 4 timeframes (5m/15m/30m/60m). Zero combinations passed winsorize. Hypothesis space exhausted.

arity-2 / arity-3 / arity-4 Primitive Conjunctions

2026-04-23 · 49-primitive library · 18,492 TRIALS NULL

Genetic search across 49 stateless primitives, conjunctions of 2/3/4 arity. 18,492 trials. Zero passed CI_lo > 1.0. Confirmed the stateless primitive library has no exploitable AND-edge in this market regime.

Mega ORB v2 / fade_breakout

2026-04-14 · NQ / ES / GC · PF CI_LO = 1.00

Entire historical-bar ORB sweep retired. fade_breakout's PF lower bound clamped to exactly 1.00 — indistinguishable from random. Nothing in the sweep cleared the gate after honest methodology.

ON THE PAPER-TRACK / NOT YET DEPLOYABLE

NQ-Daily → GC-Daily Continuation

2026-04-25 · GC · PAPER-TRACK · n=56 too thin

Prior-day NQ RTH return > 1% predicts +ve GC RTH return today. PF 1.62 · 95% CI [1.09, 2.94] · 6/6 WF folds positive · 3-year stable.

WHY NOT SHIPPING YET n=56 fails Bonferroni correction (12 cells tested → adjusted CI_lo drops to 0.51). Structural narrative is weak ("risk-on momentum carries to gold") — could be regime-specific to 2024-2026 AI-bubble + weak-USD. Forward-collecting; re-evaluate at n >= 100 (60-90 days).

GC ORB + L2 Concentration ≥ 15%

2026-04-24 · GC · SUGGESTIVE · n=8 too thin

Adding a break-second volume concentration filter to GC ORB flipped 11-day net from -$3,550 to +$5,050. Real aggressor delta alone matched the proxy. n=8 too small for confident deployment. Patient accumulation; re-validate in 2-3 months.

GC London Gap-Fade

2026-04-23 · GC · PAPER-TRACK · 2nd real edge candidate

Structural mean-reversion at the London open. PF 1.49 · 95% CI [1.16, 2.11] with 1.5-tick slip baked in. Paper-tracking; promotion to live pending more live samples.

THE METHODOLOGY THAT KILLED THESE

Every claim above was killed by the same set of gates:

  1. Anchored walk-forward (6 folds). The strategy must produce ≥4/6 positive folds. Killed orbFade, ORB, boxRange.
  2. Bootstrap profit-factor CI (10k moving-block resamples, block=√n). 95% CI lower bound must exceed 1.0. Killed GC handoff, NQ-GC lead-lag.
  3. Bonferroni correction for multiple-testing. If we tested K cells, the surviving CI must clear 1-α/K. Killed several "passed" candidates.
  4. Look-ahead audit. Every signal feature must be computable at bar T using only bars 1..T. Killed volume-profile + Aramid Pro EMA-cross.
  5. Cost-realism baked in. 1.5-tick slip per side + $5 commission. Killed L2 imbalance and NQ-GC lead-lag (statistically real, economically dead).
  6. Live broker validation. Even if all five gates pass, the strategy ships to a sim/PRAC account first. Live decay-monitor auto-HALTs if live-PF / backtest-PF drops below 0.5 for 5 consecutive days.

If you're a paying customer, you can verify any of the survivors at aramid-live.pages.dev (live broker P&L, not screenshots). The full methodology is documented at aramid-lab.pages.dev.