Three research rounds on free daily data (2017–2026, 3,547 days): brute-force weight sweep across 1,771 combinations, a cycle-event study, and a probabilistic regime model reproduced across three walk-forward folds. Every number below is computed by the script that generated this page — no hand-typed results.
Composite vs 30-day forward returns; holds out-of-sample across folds. A thermometer with real information — not a timing bell.
Flagged extreme at major tops/bottoms; best at legendary extremes (2017 top 86, 2024 top 81) and capitulation lows via on-chain + attention.
Confirms trends, does not call them: P(bull)>0.7 predicted up-90d only 37% at cycle maturity (F3). Honest ceiling for the live regime task.
Attention IC is +0.24 (greed-dir) inside bull regimes and inverts inside bear regimes — attention must be regime-conditional.
Log price (ink), composite temperature v0.3 (amber, 0–100 right axis), regime probability P(bull) (blue), bear-market ground truth (red wash: price<SMA200 & SMA50<SMA200). Cycle events marked ◆. The P(bull) line is produced by the model trained only through 2024-06 — in-sample before that date, out-of-sample after.
The full pipeline (fit → evaluate on unseen window) was repeated three times with different train/test cuts, plus leave-one-feature-out refits. A conclusion only counts if it survives folds.
| Fold | P(bull)>0.7 → fwd90 positive | P(bull)<0.3 → fwd90 negative | q5−q1 fwd90 spread | IC30 composite | Regime lag |
|---|---|---|---|---|---|
| F1 · 2017–21 → 2022–23 | 62% (n=21) | 58% (n=233) | +22% | 0.226 | -47 days after SMA200 cross |
| F2 · 2017–23 → 2024–25H1 | 39% (n=152) | — | -45.3% | 0.142 | -37 days after SMA200 cross |
| F3 · 2017–24H1 → 2024H2–26 | 18% (n=38) | 68% (n=31) | -2.4% | 0.065 | -25 days after SMA200 cross |
What reproduces: positive-but-modest composite IC (every fold), bear-side regime detection, attention's regime flip. What does not: confident bull-side calls at cycle maturity (F2/F3 show <50% hit when P(bull)>0.7) — momentum models buy tops. We report both. Leave-one-feature-out refits kept the failure pattern stable (no single feature creates or fixes it — test rank-IC of the model vs 90-day returns with each feature dropped): sma200: -0.02 · mom90: -0.03 · mom30: -0.02 · dd: -0.02 · vol: -0.03 · mvrv: -0.03 · wiki: -0.02.
| Date | Event | Composite | Reading |
|---|---|---|---|
| 2017-12-17 | top $19.6k | 86 | flagged hot ✅ |
| 2018-12-15 | bottom $3.2k | 38 | mild only 〽️ |
| 2021-04-14 | top $64.9k | 44 | missed ❌ |
| 2021-07-20 | low $29.8k | 20 | flagged cold ✅ |
| 2021-11-10 | top $69k | 57 | warm only 〽️ |
| 2022-11-21 | bottom $15.5k | 50 | missed ❌ |
| 2024-03-14 | top $73.8k | 81 | flagged hot ✅ |
| 2024-08-05 | low $49k | 54 | missed ❌ |
| 2025-01-20 | top $109k | 42 | missed ❌ |
| 2025-04-07 | low $74.5k | 34 | mild only 〽️ |
Verifications used in this dossier (jev typed contracts, run today as deterministic stand-ins — at M2 the identical contracts execute via live jev calls with probability gates): (1) schema gate — every snapshot row passes a typed schema before scoring; malformed source payloads are rejected to neutral, never averaged in; (2) regime task — the P(bull) line is exactly jev's {regime: bull|bear|chop, probabilities} contract; this backtest is the accuracy bound the live model must beat; (3) calibration check — probability outputs are stretched/calibrated on train windows and never consumed raw; (4) anomaly triage — 3σ jumps in the series are flagged with a typed {likely_cause, action} rather than silently smoothed.
Where jev shines here: the historical layers (price, MVRV, attention counts) are deterministic and jev adds nothing there — that is deliberate. jev is load-bearing exactly where data is unstructured text: classifying thousands of headlines and Reddit posts per run into typed sentiment with probabilities (pillar D's live version will be far richer than the attention proxy testable retroactively), tagging divergence causes (news-driven vs leverage flush vs data issue), triaging anomalous readings before they touch the index, and producing the fact-bound regime line. The research shows why the typed-probability shape matters: uncalibrated, free-text sentiment would have been swept into the index with the wrong orientation (attention inverts sign by regime). jev's contract — typed output, probabilities attached, gated, capped at 12 index points — is what made every failure above detectable and every discovery reproducible.
Overlapping forward returns inflate apparent significance (no transaction-cost or strategy claims are made); pillar D is an attention proxy — retro-classified news/Reddit sentiment does not exist, so live D may test differently; funding history (167d) and ETF flows (2024→) are too short to backtest; three folds + feature drop-outs reduce but do not eliminate multiple-testing bias; ground-truth "bear" is defined by moving-average rules, not exogenous truth.