The Research Record
Around twenty candidate strategies have gone through the validation gate. None survived. The gate, the kill log, and the retraction are the work. Anyone can show a green equity curve; this page shows the process that refuses to lie to me.
The Gate (current)
These four statistical gates are mandatory for an alpha claim. For the flagship archive, the five due-diligence thresholds and the decision rule were recorded in the morning commit and the verdict in the evening commit of the same day, in that order.
Factor-adjusted alpha t-stat after regressing out market beta. Beta dressed up as alpha fails here.
Deflated Sharpe ratio, penalised for every trial ever registered against the dataset. 462 trials on the ledger.
After real fees and slippage, measured from live fill data. Not exchange headline fees.
Minimum sample before any metric is even computed. Below that, the harness refuses to answer.
Behind the gate: walk-forward splits, permutation tests, placebo event dates, a power analysis that refuses trials the data cannot answer, and a hash-chained provenance journal so no result can be quietly re-run until it passes. 5,500+ tests in the suite. 462 trials on the deflation ledger; sweep configurations (like the ~3,500-trial TA sweep) are counted separately as families, not individual ledger entries.
The Kill Log
Every entry was built in full, tested in full, and killed by a specific, named test. One entry is a kill that was itself wrong, and retracted.
BUILT:A three-edge US equity and ETF sleeve. Its first pass showed OOS Sharpe 1.58; the accompanying DSR of 0.9709 counted only 13 trials. Looked like the one.
KILLED BY: Honest trial accounting. The registry actually held 462 candidates, not 13. Recomputed: DSR 0.000 vs the pre-registered 0.95 threshold, placebo percentile 0.913 vs 0.95, full-period Sharpe 0.64 against SPY at 0.93. Three of the five pre-registered DD checks failed. That battery predates the current four-gate framework above; the report documents both.
Full due-diligence report below.
BUILT:4-hour mean reversion on crypto perps. Ran live. Was labelled VALIDATED on an earlier version of this page.
KILLED BY: Re-audit at real fees: alpha_t 1.03, DSR 0.00. The earlier validation used optimistic cost assumptions. The re-audit harness now gates every candidate.
reaudit.py is now a mandatory gate.
BUILT:Post-earnings announcement drift on large caps. Built, walk-forward validated, ready to arm.
KILLED BY: Placebo test: random event dates produced the same "edge". The signal was a low-volatility mirage, not earnings drift.
Placebo nulls are now standard for every event study.
BUILT:The whole family: momentum, reversal, breakout variants across timeframes and assets.
KILLED BY: A ~3,500-trial walk-forward sweep. Turnover costs kill every configuration. Family-level verdict, not cherry-picked survivors.
The search space itself is closed.
BUILT:Cross-sectional funding-rate carry on Hyperliquid, delta-hedged.
KILLED BY: Only 8 hedgeable pairs exist. Basis P&L dominates the funding income. The carry is real; the hedge you need to collect it costs more than the carry.
Killed before a dollar was risked.
BUILT:The harness produced a kill (iteration 4) and a negative sweep verdict (sweep 5). Both were then audited.
WHAT HAPPENED: The harness gets audited too. The iteration-4 kill was retracted for two reasons. The fixed 5% and 7% barriers were not volatility-scaled, so the test measured asset volatility rather than the mechanism. The signal also fell inside a matched-placebo band. The revised verdict was inconclusive. The same audit separately corrected a cost model that overstated some costs by roughly 3x. A separate sweep-5 audit corrected a roughly 32x capital-allocation overstatement in the economic gate. Both revised verdicts are in the commit history.
Both retractions are in the commit history.
Read the Due-Diligence Report
The full post-mortem of combined_sleeve_v1: the point-in-time universe rebuild (1,206 symbols, 2,015 membership-change events), the honest trial count, the gate-by-gate verdict, and the residual-risk register. This is the document I would want to read before trusting anyone's backtest, so I wrote it about my own.
combined_sleeve_v1_DD.pdfWhat Still Runs
Small capital, deliberately, because nothing has earned size. What runs is infrastructure, not conviction.
A Uniswap v4 CLMM LP hedged with a Hyperliquid ETH short. Net delta held near zero. The position is a fee-collection structure, not a directional bet.
A crisis-protective momentum sleeve with regime-gated sizing that de-risks to zero in stress. Factor attribution says it is beta, not alpha (alpha_t 0.35), and the site says so too. The protection layer (hard 3% stop, 8% trail from peak) is armed and live.