Correction (2026-10-06): The +64% trend-filter figure below left out trading fees and was measured on the same data used to choose the 50-day setting. Our own out-of-sample check, after fees, found the filter lost less than simply holding (−36% vs −52% median per coin), but it still lost. Lines like "'X' leads on auc" show the best of many coin-flip results, not a real edge. The grid-suitability result was also misread: the score ranks coins backwards, so the coins it rated best lost the most.
The Lab ran 21 experiments across 6 research threads this cycle (2026-07-08 → 2026-07-13) on the idle GPUs — every verdict below is published as the data said it, wins, losses and coin-flips alike. No cherry-picking.
Feature-family ablation
- majors pack, 5m bars — No single feature family is load-bearing — signal (such as it is) is diffuse
- majors pack, 15m bars — No single feature family is load-bearing — signal (such as it is) is diffuse
- broad pack, 15m bars — No single feature family is load-bearing — signal (such as it is) is diffuse
Confidence-gated trading sim
- majors pack, 5m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.6) still nets -33.2 bps/trade
- alts pack, 5m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.65) still nets 159.0 bps/trade
- meme pack, 5m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.6) still nets 20.7 bps/trade
- broad pack, 5m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.65) still nets 76.2 bps/trade
- majors pack, 5m bars, 3-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.65) still nets -28.2 bps/trade
- broad pack, 5m bars, 3-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.65) still nets 18.1 bps/trade
- majors pack, 15m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.65) still nets -1.4 bps/trade
- alts pack, 15m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.7) still nets 37.6 bps/trade
- meme pack, 15m bars, 1-bar horizon — POSSIBLE gated edge: short @ p>0.65 nets 209.6 bps/trade over 294 trades — needs leakage + robustness checks before anyone gets excited
- alts pack, 15m bars, 3-bar horizon — NO tradeable subset after 25bps/side fees — best cell (short @ p>0.65) still nets 111.4 bps/trade
- broad pack, 15m bars, 1-bar horizon — NO tradeable subset after 25bps/side fees — best cell (long @ p>0.7) still nets -0.3 bps/trade
- majors pack, 15m bars, 3-bar horizon — NO tradeable subset after 25bps/side fees — best cell (short @ p>0.65) still nets -7.8 bps/trade
- meme pack, 15m bars, 3-bar horizon — NO tradeable subset after 25bps/side fees — best cell (short @ p>0.65) still nets 30.7 bps/trade
- broad pack, 15m bars, 3-bar horizon — NO tradeable subset after 25bps/side fees — best cell (short @ p>0.65) still nets -0.0 bps/trade
Which crypto coins can a simple ML model actually predict?
- run of 2026-07-10 — 'ARB' leads on auc: +0.532
Would pausing grids in a BTC downtrend stop the bleed?
- run of 2026-07-11 — 'BTC-downtrend' leads on total_pnl: +118
Which coins are actually good for grids?
- run of 2026-07-12 — 'bottom (cull?)' leads on mean_grid_pnl: -2.83
Is a simple trend filter better than the grid?
- run of 2026-07-12 — 'trend_follow_50' leads on mean_return_pct: +64.00
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.