foxalgoGraveyard
Most tested ideas don't survive contact with an adversarial review — lookahead bias, confounded stratification, or a placebo control that quietly fails. We publish the ones that died here instead of quietly dropping them.
TF timeframe agreement DEAD
Hypothesis: Cross-timeframe candle agreement predicts next-move direction.
Agreement of the timeframes is not a directional signal: across 2,011,483 range reads in our own data (2010-2026), price moved the agreed way 48.8-49.0% of the time (baseline 50.1%), and it changes the size of the next move by about 1%. Strong-looking hits were window-boundary artifacts (the TF candle sat inside the London/Asia session window).
Source: Research archive #01
Cross-sector / intermarket cues DEAD
Hypothesis: Rates move stocks, USD moves metals/energy — cross-sector signals carry information.
Raw deltas of 15-17 bps shrank to 1-5 bps once stratified by each market's own position (pos930_atr). Daily-level intermarket analysis adds nothing beyond a market's own state. Only the peer-Asia-sweep pattern inside the same sector survives.
Source: Research archive #02
Event days (FOMC / CPI / NFP) DEAD
Hypothesis: Scheduled macro events shift conditional touch/breakout probabilities.
Conditional probabilities did not move on event-flagged days in any tested pair.
Source: Research archive #03
VWAP-side pairs as new information DEAD (as a combo)
Hypothesis: print8 & vwap_side930 combined carries information beyond either signal alone.
Raw delta looked like +50 bps versus +27 bps for pos930 alone, but the increment over stratified singles with shared weights was only ~+1.5 bps (52/54 draws positive, 0/54 above 5 bps). Lesson: stratified deltas must be subtracted using shared weights, or three different band-weight distributions produce a compositional artifact.
Source: Research archive #04
Sweep-behind-price setup FALSIFIED — lookahead
Hypothesis: A liquidity sweep behind price plus room ahead shows a +20.6 bps edge.
The +20.6 bps result used information not yet known at decision time (lookahead). A placebo control did not catch it — the key lesson was that a naive placebo does not reveal lookahead bias by itself.
Source: Research archive #05
Expansion as a directional signal FALSIFIED — lookahead
Hypothesis: Range expansion predicts next-bar direction, +18.4 bps.
+18.4 bps was intrabar lookahead (the signal used the fill bar's own outcome). After the fix, the effect dropped to +0.2 bps.
Source: Research archive #06
HTF anti-persistence timing DEAD — lookahead
Hypothesis: Anti-persistence against the higher timeframe improves selection timing.
The construction required future information to compute; retired as lookahead.
Source: Research archive #07
Calendar-day flag FALSIFIED
Hypothesis: Day-of-week / calendar position explains part of the observed ratio.
The apparent calendar effect was volatility repricing, not a calendar effect.
Source: Research archive #08
The 4.045x opening-range target NOT NEAR-CERTAIN
Hypothesis: Projecting 4.045x the width of the 09:30:00-09:30:30 candle beyond its high and low tags one of the two levels almost every day, and once a trigger 15 points away is hit the target is near-certain.
Measured on the exact construction over 7,821 days of 15-second bars: in 2025 either side was touched on 86.5% of ES days [81.6-90.3] and 82.5% of NQ days [76.5-87.2] - well below 'almost every day'. Once the 15-point trigger was hit, the target was reached 66.3% up / 64.4% down, and conditioning further on a 20-point retrace lowered it to 57.9% / 57.8%. Touch odds also track the opening candle's width, not the level itself: 91-93% on the narrowest quintile against ~80% on the widest.
Source: Research: the 4.045 opening-range level
Entries after an hourly sweep DEAD
Hypothesis: Entries taken after an hourly high or low is swept capture 40-50% more favourable excursion per unit of stop risk than entries taken at any other time.
Measured per direction on 4.07M synthetic entries across six futures, comparing the four bars after a sweep against every other bar. The improvement came out between -0.7% and +0.8% depending on target distance - two orders of magnitude below the hypothesised 40-50% - and the two placebo passes carried opposite signs, which is what noise looks like. Stratifying by volatility, session and sweep direction moved nothing.
Source: Research archive #09
Hourly sweep-and-reclaim as a direction signal DEAD - geometry
Hypothesis: When an hourly candle sweeps the previous hour's high and closes back inside, the next move goes the other way.
508,481 hourly events across six futures, 2010-2026. Next-hour direction after a sweep-and-reclaim landed within half a point of the base rate - a coin flip. The one large-looking effect (which side of the hour's own range gets touched first, -0.42) collapsed to -0.04 once the close's position inside its own candle was controlled for: 91% of it was the mechanical fact that a candle closing near its high is already next to that high.
Source: Research archive #10
Range-breakout continuation DEAD
Hypothesis: When price closes outside a confirmed swing range, the move continues, and the recent sequence of resolutions sharpens the odds.
477,611 breakouts on six futures across seven timeframes, fitted 2010-2017 and measured 2018-2026. Continuation ran at 47.4% - slightly under a coin flip, before costs. An order-3 Markov state over recent resolutions added nothing out of sample (Brier moved by -0.0004, the wrong way), and no event in half a million reached even a 70% probability. The predictions were well calibrated; they were calibrated noise.
Source: Research archive #11
Roster rotation by recent performance REAL BUT NOT ENOUGH
Hypothesis: Rebalancing daily into whichever strategy legs performed best over their last 20 trades turns a pile of weak legs into a smooth book.
Tested on 84 of our own legs whose individual edge is a coin flip after costs. The rotation is genuinely selective - it beat a frequency-matched random roster at the 100th percentile of 200 draws, and the improvement survives a jackpot check. It still could not make the book positive: hold-everything lost heavily, rotation cut the loss to roughly a twelfth of it, and that was all. Rotation reduces and reshapes variance; it does not create edge. The anti-roster, deliberately picking the worst legs, finished ahead of the winner-picking one on this substrate.
Source: Research archive #12
Day direction from the first hours DEAD
Hypothesis: How the morning trades tells you where the session closes.
22,423 sessions. A model given the morning's move scored 0.521 by area under the curve - below its own null. The high agreement that looks so convincing after the initial balance is the move that has already happened; what is left of the day averages 0.007 ATR, in either direction.
Source: Research archive #14
Time-based quarters (quarterly theory) DEAD
Hypothesis: Time divides into quarters with fixed roles - accumulation, manipulation, distribution, continuation - so the second quarter of a session, day or week tends to form the period's extreme, a sweep of Q1's range in Q2 reverses, and price relative to the 'true open' gives directional bias.
Tested at three scales (90-minute quarters of the NY morning, four 6-hour quarters of the day, weekday quarters of the week) on 34 futures, 2010-2026, with a volatility-clock null of 200 resampled paths per market. None of nine pre-registered hypotheses survived multiple-comparison correction in either half of the data. The extreme-timing claim inverts: the day's high or low forms in the 00:00-06:00 quarter least often of all four (~11-16% of days), and the concentration in the morning-RTH quarter is fully reproduced by the volatility clock. After Q2 sweeps Q1's edge, price returns to Q1's midpoint less often than distance and remaining time alone predict, not more. A separate autopsy of 20 earlier conditional findings built on these windows attributed all 20 to proximity - distance to the level, not the quarter, carried the number. Re-testing with the roles reassigned - Q1 forming the period's cap, Q2 continuing it - changed nothing: Q1's share of extremes sits on the volatility-clock null (+0.2pp in-sample, -0.8pp held-out), and quarter-to-quarter sign continuation is under one percentage point at every transition. The one consistent deviation is not quarterly at all: days print their extreme by noon more often than a reshuffled ordering of the same returns would, and afternoons under-extend. That one deviation was then attacked separately with pre-registered kill criteria: it survives a volatility-matched null (91% of the effect remains), holds in both markets tested and in 16 of 17 years, and persists after excluding opening-drive and gap days - but its finer 22.5-minute version turned out to be the opening print itself, not additional structure. What remains is a small, verified fact about timing, not about the theory: the session's extreme lands in the opening quarter of regular hours about 2.4 points more often than the volatility clock implies.
Source: Research archive #15