The Lab ran 10 experiments across 8 research threads this cycle (2026-07-27 → 2026-07-31) on the idle GPUs — every verdict below is published as the data said it, wins, losses and coin-flips alike. No cherry-picking.
Wide-feature XGBoost direction
- broad pack, 5m bars, 6-bar horizon (fresh-data re-run) — Marginal: AUC slightly above coin-flip; not bankable on its own
- alts pack, 15m bars, 1-bar horizon (fresh-data re-run) — Marginal: AUC slightly above coin-flip; not bankable on its own
- majors pack, 15m bars, 1-bar horizon (fresh-data re-run) — Marginal: AUC slightly above coin-flip; not bankable on its own
I tried to get paid without predicting anything. It cost me money.
- run of 2026-07-28 — Cash-and-carry needs no forecast — you just collect funding. Across 23 coins and a year of real rates, the median paid -2.41% APR.
I tested four ways to make money in crypto. The best one made 1.5% a year.
- run of 2026-07-28 — Prediction, trend-following, ranking, and carry — all measured, all normalised to the same unit. Every strategy that beat the market beat it by losing less.
I checked whether my own backtests were lying to me.
- run of 2026-07-28 — Every backtest assumed that if price touches your order, you get filled. I never verified it. 156 real orders later: 97.1%.
My research bot published the same discovery six days in a row.
- run of 2026-07-28 — Six identical findings, all labelled 'fresh'. The data underneath hadn't moved in two months and every health check was green.
Ranking coins beats picking them. Both still lose money.
- run of 2026-07-28 — Cross-sectional momentum across 90 configurations. Diversification helps monotonically — and the benchmark still lost 20%.
We hand-tuned 221 stocks. One shared setting beat all of it.
- run of 2026-07-30 — Per-stock parameter tuning lost to a single generic setting on both strategies. The tuning was memorising noise.
The models said we'd lose money. They were right
- to within twelve dollars. — Matched 53 live trades against backtest predictions from exchange records. Gap-fade landed within $12 of its model.
Auto-published weekly by Gordon Lab. Each entry is a real backtest or model run on real market data; methods and full result tables live in the lab archive. This is research, not financial advice.