We publish the failures too.

Five research rounds on free, point-in-time data — hunting for edge in Bitcoin sentiment. Some found it, some refuted our own hypotheses. Every number on these pages is computed by the script that generated the page.

RoundQuestionMethodVerdict
R1Which pillars matter? (weights by brute sweep)1,771 combos · walk-forwardAttention belongs in the mix at ~25% — but flips sign by regime. On-chain is the long-horizon validator.
R2Can attention time tops?cycle event study 2017–25Flagged 3/10 extremes; attention is systematically contrarian at capitulation. Thermometer, not bell.
R3Does it reproduce?3 folds · feature drop-outComposite IC positive in every fold (0.23/0.14/0.07). Regime model lags SMA200 by ~26d — honest ceiling.
R4Does structure beat holding?A–E level engine · 6/6 suite+8,209% vs +6,570% B&H — same return, half the drawdown.
R5Where else could edge live?carry bands · jev NOUL/CHOICE/SCORECarry = cash-plus (Sharpe 3, tiny abs). Synthesis: park the level system's idle cash in strict-band carry.

Dossiers

R1 · WEIGHT DISCOVERY

Six pillars, 1,771 weight combinations, train/holdout discipline. Where v0.3 weights come from.

R2 · CYCLE EVENT STUDY

Every major top and bottom since 2017, and what the index actually read on those days.

R3 · REPRODUCTION FOLDS

Three walk-forward folds plus leave-one-feature-out. What survives, what doesn't.

R4 · THE $10,000 BOOK

Five levels of aggression backtested nine years — full reliability suite included.

R5 · CARRY VS EVERYTHING

Seven years of 8h funding, five bands, and the honest size of the carry edge.

R6 · BENCHMARKS

The level system against the S&P, gold, 60/40 and trend-following on one chart.