Research — full grids · held-out windows untouched
Long-form write-ups of what survived validation and what died. Every number comes from a pre-registered protocol: full grids reported, held-out windows untouched until adoption, permutation nulls on any survivor, and a permanent ledger of rejections.
The validation gauntlet every candidate strategy must survive before trading — walk-forward splits, permutation nulls, cost stress, and virgin holdouts. Most edges die there. That is the point.
2026-08-24 · read →A walk-forward ML rotation over 59 S&P names, blending two prediction horizons and gating entries by cross-sectional rank. +27.9% CAGR on an untouched test window, Sharpe 1.26 — with the Monte-Carlo receipts to back it.
2026-08-24 · read →A full-S&P expansion that collapsed, a stat-arb book undone by costs and a backtest bug, thirteen intraday families killed by fees, and an ML prediction program closed by its own poison control. Negative results, fully documented.
2026-08-24 · read →One rule — long when price is above its 200-day average AND 60-day momentum is positive — was tested across 28 liquid ETFs spanning equities, sectors, rates, gold, FX and commodities. It was net-positive on all 28.
2026-08-24 · read →