The graveyard
Twenty-seven strategy families faced the gate. One is alive.
This page is the running scoreboard of everything we have tested: the claim, and the verdict, in one line each. The full postmortems — data, code decisions, every number the gate printed — live in the member library. We publish the deaths because they are the product: anyone can show you a winning backtest, which is exactly why you should not trust one.
Registered before results
The rule, the universe, the costs, and every variation we intend to try are written down before any result exists. A hypothesis conjured from its own failed test gets registered as new — and charged for the search that spawned it.
Charged for every variation
Test 27 things and one will look good by luck. Our gate (deflated Sharpe, combinatorial cross-validation, overfit probability) explicitly pays for the whole search that found a result, not just the winning row.
Costs and corpses included
Real spreads, real commissions, delisted stocks settled at their last price instead of quietly deleted. Two families earned fake passing grades from data artifacts this summer; the autopsies that caught them are in the library.
Untestable is a verdict
When the data cannot support an honest test, we say untestable and stop — we do not publish a number from a broken instrument. Several families below carry exactly that verdict.
The scoreboard
26 dead or untestable. 1 alive. Every verdict final until new data.
Options — volatility risk premium
The only band that has produced a survivor.
Defined-risk SPY put spreads (the survivor)
“index options overprice tail risk”
ALIVE — passed deep-sample validation, replicated out-of-sample on an index it never saw, capacity measured. Trades the live book.
Cash-secured puts / covered calls, high IV-rank
“premium selling on rich single names”
Died — the premium is real and the tail eats it. Every configuration Tier C.
Credit-spread regime grids
“condition spread entries on volatility regime”
Died — filtering entries on fear made the edge worse, three ways. The unconditioned version is the survivor.
Iron condors
“sell both tails at once”
Died — the call side subsidizes nothing; the put side alone was the whole edge.
Entry day-of-week, short-dated spreads
“does the weekday you sell matter?”
Died at the gate — Friday looked best but is the same premium re-observed, not a new edge. 96% win rates are exactly what the gate exists to distrust.
Earnings-week IV crush
“single-name straddles richest before earnings”
UNTESTABLE on free data — measured, not assumed: only 25% of events at the 20 most liquid names have usable chain coverage. We are now collecting the missing data ourselves.
US equities — days to two weeks
Six families, six mechanisms, one clean instrument. The band is closed: whatever lives here is smaller than costs plus honest multiplicity.
RSI washout reversal
“oversold liquid names snap back”
Died honestly — the premium exists (~+0.6% per event) and is far too thin to survive costs and the number of variations tried.
Overnight gap fade
“big gaps overshoot and fill”
First ruled untestable (unadjusted splits printed fake +694% events), then refuted on repaired data — 9 cells, all dead. The only pulse sat in shorts no real borrow would fill.
Volume-surge continuation
“follow the 3-8x volume day”
Died decisively — negative in all nine cells. By the next open the information is already in the price.
Post-earnings drift (PEAD)
“positive surprises keep drifting”
The instructive one: printed a fake Tier A on poisoned data (splits, duplicated events), then died honestly on the rebuilt panel — 43,616 events, every median negative. The autopsy is in the library.
Turn-of-month flows
“index ETFs rally around month-end”
Died — visible (+0.36%/window on SPY, 58% winners) and too small. Textbook decay of a published anomaly.
52-week-high momentum
“names pressing highs keep going”
Died with the sign inverted — buying highs on high volume lost in every cell.
Analyst estimate-revision momentum
“raised consensus drifts up”
Died — the Zacks-era edge is dust at institutional liquidity. Faint pulse at the strictest cut, negative medians everywhere.
Earnings announcement premium
“names rally into their expected report”
Died inverted — holding liquid names into earnings weeks cost money versus the index, 2021-2026.
Dividend run-up
“prices rise into ex-dates”
Died decisively negative across 38,000 windows per cell.
US equities — monthly horizon
Closed as not-provable-at-this-history: 55 months cannot prove a monthly factor, and we say so instead of pretending.
Low-volatility sleeve
“boring stocks quietly win”
Refuted for this era — lagged the index by 9-12% a year once low-vol became an ETF category. The famous anomaly, inverted.
Momentum 12-1 sleeve
“winners keep winning”
Right sign (+7.9%/yr on the decile) and unprovable on 4.5 years of history. Filed as open, not as a win — that restraint is the method.
FX
Verdict of the whole book: session structure and calendar effects are real and sub-spread at retail costs.
Carry on majors
“collect the interest differential”
Died — decomposed into a yen-trend bet wearing a respectable name.
H4 trend following
“trends persist at four-hour scale”
Died — 216 configurations, all Tier C.
London-open reversal
“fade the Asian session at the London open”
Died backwards — big Asian moves continue, and filtering to the biggest made every pair worse.
NY-overlap momentum
“the London morning persists through the overlap”
The pattern exists exactly where the literature says — worth 1-3bp gross where the spread costs 1-2bp. Retail intraday FX in one sentence.
Weekend gap
“Friday close to Sunday open mean-reverts”
Mechanism real, premium smaller than the reopening spread.
Intraday mean-reversion
“stretched hourly moves revert”
The one FX idea still breathing — trading on paper, earning its verdict in public, not yet promoted.
Machine learning & alternative signals
Tested with the same gate as everything else — including the ideas most shops would refuse to test at all.
KRONOS ML forecaster
“a trained model predicts returns”
Died on cross-validation — best config looked strong in-sample and could not survive honest resampling.
Insider cluster buying
“insiders buying together know something”
Died — the mean is positive, the median loses, and half the events were unpriceable.
Gematria / numerology signals
“cipher-and-date alignments carry information”
Tested seriously, twice: against 300 placebo ciphers across 3.7M stock-days, and against 99 years of market history with preregistered hypotheses and permutation nulls. Result: statistically identical to its own placebos; a classifier fed only these features scores a coin flip. Not detected — said without hedging.
Liquidity sweeps
“the ICT idea, measured”
Nothing there once measured instead of mythologized.
Reading the record
A 4% survival rate is the finding.
Two entire time horizons — days-to-two-weeks in US equities, and intraday FX at retail costs — are now closed as findings, not gaps: six mechanisms attacked each one on clean data and nothing survived. That is what an honest search of public market ideas looks like, and it is why the one edge that did survive means something. The gate that killed twenty-six families is the same gate that passed one — you cannot have the second without the first.
All figures on this page are research statistics from historical simulations, not performance claims. Past results — simulated or live — do not predict future performance.