The Graveyard
Rejected ideas, failed tests, and lessons from the research process. The public sees what died and why. Exact parameters and live-edge mechanics stay private.
Bowie V-pattern reversal — KILLED
Thesis: Does Bowie's V-pattern reversal work mechanically?
What killed it: No — negative across the universe; mechanical reversal-catching loses. Killed.
Evidence: Real-price replay
Lesson: No — negative across the universe; mechanical reversal-catching loses. Killed.
Bowie 5-min MACD trend — KILLED
Thesis: Does Bowie's 5-min MACD trend strategy work mechanically?
What killed it: No — negative on every symbol over 2 years; classic crossover chop. Killed.
Evidence: Real-price replay
Lesson: No — negative on every symbol over 2 years; classic crossover chop. Killed.
Bowie ORB tested — range-day filter SURVIVES, mechanical retest KILLED
Thesis: Do Bowie's ORB rules — retest entry and the range-day skip — hold up mechanically over 2 years?
What killed it: The range-day filter is real and transferable; the 'enter on the retest' rule fails mechanically because it misses the runners. Kept one, killed the other.
Evidence: Real-price replay
Lesson: The range-day filter is real and transferable; the 'enter on the retest' rule fails mechanically because it misses the runners. Kept one, killed the other.
Indicator crossover day-trade family (RSI/MACD/MA)
Thesis: Time entries with RSI, MACD, and moving-average crossovers.
What killed it: Lagging transforms of price everyone already watches — the crowd competes the edge away; costs finish it.
Evidence: Real-price replay
Lesson: Lagging transforms of price everyone already watches — the crowd competes the edge away; costs finish it.
5-min BTC Polymarket favorite-scalp
Thesis: Buy the near-decided favorite on 5-min crypto prediction markets.
What killed it: The favorite is priced fairly — the spread eats the whole edge. A guaranteed loss after costs.
Evidence: Real-price replay
Lesson: The favorite is priced fairly — the spread eats the whole edge. A guaranteed loss after costs.
Basket tier-tilt weights
Thesis: Tilt the research basket toward its undervalued names and away from the overvalued — let the valuation score move the weights.
What killed it: A look-ahead artifact: scored honestly out-of-sample, the tilted version underperforms the plain rule it was meant to improve. The flattering number came from measuring momentum over its own evaluation window.
Evidence: Real-price replay
Lesson: A look-ahead artifact: scored honestly out-of-sample, the tilted version underperforms the plain rule it was meant to improve. The flattering number came from measuring momentum over its own evaluation window.
Fade-as-diversifier
Thesis: Keep the mean-reversion sleeve not to earn, but to pay on the days the breakout bleeds — a shock absorber for the book.
What killed it: It doesn't pay when needed: the model's rescue rate was fiction on real fills (0-for-10), the correlation benefit evaporates out-of-sample, and standalone it earns nothing. Tail insurance that never pays out isn't insurance.
Evidence: Real-price replay
Lesson: It doesn't pay when needed: the model's rescue rate was fiction on real fills (0-for-10), the correlation benefit evaporates out-of-sample, and standalone it earns nothing. Tail insurance that never pays out isn't insurance.
Turn-of-month flows
Thesis: Calendar-forced institutional money should concentrate market drift around the month turn — arm the machine during whale week.
What killed it: Whale week isn't in the tape: a sliver of the drift on a fifth of the days, and flat for at least fifteen years. The plumbing is real; the price effect was arbitraged away — textbook post-publication decay.
Evidence: Real-price replay
Lesson: Whale week isn't in the tape: a sliver of the drift on a fifth of the days, and flat for at least fifteen years. The plumbing is real; the price effect was arbitraged away — textbook post-publication decay.
Turtle Soup + volume tell (Wyckoff)
Thesis: Fade the failed breakout — the trapped crowd has to unwind — and let a century-old volume tell separate real springs from noise.
What killed it: The trap is real (three-quarters of level breaks fail back inside) but fading it doesn't pay intraday, and the volume tell only manufactured in-sample beauty that flipped out-of-sample. Knowing about the trap is worth more than trading it.
Evidence: Real-price replay
Lesson: The trap is real (three-quarters of level breaks fail back inside) but fading it doesn't pay intraday, and the volume tell only manufactured in-sample beauty that flipped out-of-sample. Knowing about the trap is worth more than trading it.
Oops! gap-reversal (L. Williams)
Thesis: Larry Williams' classic — when an overnight gap fails and price reclaims yesterday's range, fade it toward the gap fill.
What killed it: Every setting made money in-sample and lost out-of-sample — the signature of a fitted artifact. Five trades were 107% of two years' profit; the median trade was a coin toss.
Evidence: Real-price replay
Lesson: Every setting made money in-sample and lost out-of-sample — the signature of a fitted artifact. Five trades were 107% of two years' profit; the median trade was a coin toss.
NR7 arming (Crabel)
Thesis: Crabel's coiled spring — a narrow, quiet day loads the spring for the next day's breakout, so only trade after the quietest sessions.
What killed it: Armed days were worse than unarmed across the entire narrow-range family, in-sample and out. A quiet yesterday alone doesn't load the spring — compression measured as range says nothing.
Evidence: Real-price replay
Lesson: Armed days were worse than unarmed across the entire narrow-range family, in-sample and out. A quiet yesterday alone doesn't load the spring — compression measured as range says nothing.
ORB-retest on crypto
Thesis: Does the ORB 'retest holds' edge transfer from stock sessions to 24/7 crypto?
What killed it: No. The retest-holds tell is a SESSION-STRUCTURE effect — it needs a daily open/close to anchor the crowd's memory of a level. 24/7 crypto has none.
Evidence: Real-price replay
Lesson: No. The retest-holds tell is a SESSION-STRUCTURE effect — it needs a daily open/close to anchor the crowd's memory of a level. 24/7 crypto has none.
Fast crypto (100 trades/day) — definitive hunt
Thesis: Can a fast crypto strategy make 100+ trades a day profitably after costs?
What killed it: No — definitively. The intraday signals are real but too thin to survive even near-zero fees at high turnover; 100 trades/day means 4–50% of capital churned daily in costs. The only crypto edges that survive are slow.
Evidence: Real-price replay
Lesson: No — definitively. The intraday signals are real but too thin to survive even near-zero fees at high turnover; 100 trades/day means 4–50% of capital churned daily in costs. The only crypto edges that survive are slow.
Standardized earnings-surprise momentum
Thesis: Stocks with the biggest earnings surprise, scaled by their own historical surprise volatility, keep drifting in that direction for weeks after.
What killed it: No monotonic pattern across surprise-size buckets, the point estimate ran backwards from the thesis, and a shuffle test matched or beat the real result most of the time.
Evidence: Null model · ~950 earnings events · 2.7 years
Lesson: A 35-year-old published anomaly needs the same shuffle test as anything else. This one didn't survive first contact with recent data.
Nearness-to-52-week-high momentum
Thesis: Rank stocks by how close price sits to its own 52-week high and buy the names nearest their highs.
What killed it: Beat a buy-and-hold benchmark, but a same-universe random-portfolio control matched it almost exactly — the 'edge' was the universe's own bull-market drift, not the ranking.
Evidence: Null model · 774-name universe · 3 years
Lesson: Beating the index proves nothing if a random slice of the same names beats the index too. Always test excess over the universe's own drift, not just over a passive benchmark.
Mean-reversion fade
Thesis: Buy the snap-back instead of the break — a strong long-horizon model.
What killed it: Withdrawn by the founder; recent real days disagreed with the model.
Evidence: Model-only
Lesson: A glittering backtest is a reason to test, never a reason to trade.
Single-name, mornings only
Thesis: A fast single name's entire edge appeared to live before midday.
What killed it: Held out of the live roster pending real-price proof at size.
Evidence: Model-only
Lesson: A time-segmented edge is a hypothesis until live fills confirm it.
Opening-range breakout (ORB)
Thesis: Trade the break of the first fifteen minutes' range.
What killed it: Strong in the long-horizon model; a straight run of real-price failures.
Evidence: Real-price replay
Lesson: The model-vs-real gap IS the lesson. Benched until it can earn its way back at real prices.
Predicting chop days
Thesis: Detect a chop day early and stand down.
What killed it: No signal separated chop from trend ahead of time.
Evidence: Real-price replay
Lesson: Chop days don't announce themselves through your own P&L.
Daily loss limits & win locks
Thesis: Quit while you're behind — or lock in the win early.
What killed it: Loss limits scored as noise; win locks cut the runners.
Evidence: Real-price replay
Lesson: The runners pay for everything. Don't cut them to feel safe.
Looser exit family
Thesis: A long-horizon model said looser exits were worth several times more.
What killed it: Every variant lost to the tighter exits on real option prices.
Evidence: Real-price replay
Lesson: Models can't price patience; real option premiums tax it. The biggest model-vs-reality gap on the board.
Morning / late-day entry buffers
Thesis: Skip the open, or avoid the close.
What killed it: Every variant deleted more good trades than bad.
Evidence: Real-price replay
Lesson: Time-of-day filters cost more than they saved.
Breakeven stop after a small gain
Thesis: Never let a winner become a loser.
What killed it: A wash — no measurable improvement.
Evidence: Real-price replay
Lesson: Complexity that doesn't pay is removed. The simpler rule stayed.
Anticipatory entries
Thesis: Enter just before the level breaks, at a better price.
What killed it: Same gross profit, roughly double the trades.
Evidence: Real-price replay
Lesson: The spread ate the difference. Execution cost, not signal, decided it.
Retest-confirmation entries
Thesis: Wait for the pullback to confirm before entering.
What killed it: Large negative vs. the simpler entry.
Evidence: Real-price replay
Lesson: The best breaks never retest. Waiting for comfort forfeits the move.
VWAP alignment filter
Thesis: Only trade in the direction of the day's VWAP.
What killed it: Net loss across the test window.
Evidence: Real-price replay
Lesson: A directional overlay that sounds prudent removed edge, not risk.
Volume / relative-volume filters
Thesis: Only take breakouts confirmed by a volume surge.
What killed it: Worse at every threshold tested.
Evidence: Real-price replay
Lesson: The best breaks often leave on quiet tape. A confirmation filter deleted more good trades than bad.
Double-bottom entries
Thesis: Buy the retest of a prior low that holds — the classic 'W'.
What killed it: Positive on the data it was fit to, negative on instruments it had never seen.
Evidence: Transfer test (held-out instruments)
Lesson: A pattern that only works where you fit it isn't an edge. Transfer to unseen names is the honesty test it failed.
Reversal / short chart patterns
Thesis: Fade strength — upthrusts, double tops, supply-zone rejections, trend-band flips to the short side.
What killed it: Bled across a 22-year, secular-uptrend universe; only positive inside isolated bear windows.
Evidence: OOS + transfer + bear survival
Lesson: Shorting patterns are regime bets, not standalone edges. In a market that mostly rises, fading it is a slow leak.
Gann price-time 'squaring' cycles
Thesis: A move reverses when elapsed time 'squares' the price traveled — the 144 master square, the number-9 cycle, geometric angles.
What killed it: Only a plain mean-reversion leg cleared the bar; the 'squaring' geometry added nothing, and its other half was outright noise.
Evidence: OOS + transfer · 56 markets / 22 years
Lesson: Strip the mysticism and a Gann signal is just 'buy after a sharp drop turns.' The geometry was decoration, not mechanism.
How an idea earns its way in
Ideas must pass Skeptic → Quant → Risk → PM before any capital is at risk. Every idea runs a long-horizon backtest first — but backtests flatter, so nothing is believed until it survives a replay against real option prices, and nothing trades until it survives that. Live authorization does not mean an edge is proven. It means the idea survived the current gates, at tiny size, and remains under review. Killed strategies are archived, not deleted — the graveyard is part of the research asset.
Learn how the desk works →