Digital Assets Research (Research Preview)
RESEARCH PREVIEW — NOT VALIDATED FLAGSHIP. A prototype Crypto Cycle Index (CCI) was built end-to-end on 2026-04-26. Three pre-registered configurations (v1, v2A, v2C) were tested against a locked validation battery; all three failed at least one institutional gate. The model output below is published as research transparency — the regime classification it produces is intellectually informative but should not be cited as a validated index. Full failure analysis: kill log · v2 pre-registration. Path to a validated v3 likely requires paid on-chain data (Glassnode / CryptoQuant) to add the pre-2022 variance the free-tier indicators lack.
A 2-axis crypto regime classifier — Cycle phase × Correlation regime — built on an 11-indicator monthly panel via a 2-factor walk-forward DFM. Methodologically peer-reviewed (Mariano-Murasawa / Stock-Watson DFM, statsmodels), data fully transparent. The model produces meaningful classifications, just not yet at the institutional bar that BCI cleared. Updated
Current model state (research preview — not validated)
Cycle factor history
Solid line = walk-forward Cycle factor z-score from the v1 model (2-factor DFM, 11 indicators, quarterly refit, expanding window from 2018-01). Dashed lines mark the phase cutoffs (±0.25 = EXPANSION/ACCUMULATION boundary, ±0.75 = EUPHORIA/DISTRIBUTION boundary). 28 quarterly refit windows since first honest classification 2019-07-31.
Correlation factor history
The Correlation factor loads on cross-asset indicators (BTC-SPY 30d rolling corr, Fear & Greed). High values = STRUCTURALLY-CORRELATED regime (crypto trades like a risk asset). The 2025-26 transition to STRUCTURAL coincides with the documented cross-correlation rise from 0.57 to 0.85 — model is reading the regime shift the user diagnosed independently.
Validation status — three configurations, three kill rationales
This is not a methodology page hidden behind a passing gate — these are the actual validation results from the locked battery. Read this as the transparency that justifies the "preview" label above.
What each gate tests
- G1 — Factor sign stability (≥ 95% pre-reg threshold). Fraction of quarterly refits where the latent factor's sign-resolved orientation is consistent. v2A failed at 55.56% — the 12-month minimum train was too short to identify the cycle factor reliably across data windows.
- G2 — Sub-sample stability (≥ 70% phase agreement). Fits the model on a pre-2022 sub-sample and on the full sample, then compares phase classifications on the overlap window. v1 failed at 61.59%, v2C at 58.33% — the pre-2022 sub-sample fit only ever produces ACCUMULATION / EXPANSION because the underlying data variance can't push the expanding-z into EUPHORIA / DISTRIBUTION territory. v2A passed (100%) because its overlap was tiny by construction.
- G3 — Informativeness vs random (Romano-Wolf p < 0.10). Tests whether the phase distribution differs significantly from a 50/50 random labeling. All three configurations passed this gate — the model output is statistically informative even where it fails the stability gates.
Why all three failed
The unifying lesson: the free-data crypto indicator universe lacks the pre-2022 variance needed for stable 4-phase sub-sample factor identification. The pre-CME-futures era (2017 and earlier) and the pre-institutional-ETF era (pre-2024) behave structurally differently from the post-ETF era. A factor model trained on a sub-sample misses the extremes the full-sample model captures.
What we'd need for v3
Three plausible v3 angles (no commitment — decision belongs in the next conversation):
- Add Glassnode / CryptoQuant on-chain metrics — specifically NUPL, RHODL, Realized Cap. These have richer pre-2022 variance because they're derivative-of-price-with-on-chain-structure rather than purely market-data. Roughly $30–100/mo subscription.
- Drop to a 2-phase classifier (RISK-OFF / RISK-ON). Pre-2022 data CAN support a binary classification reliably. Trade-off: less institutionally distinctive than a 4-phase model.
- Wait. Each year of additional crypto history adds meaningful variance to the pre-sub-sample. By 2027 the pre-2024 sub-sample will include the full ETF-era expansion + correction cycle.
What's published vs. what isn't
Published as research
- The 11-indicator monthly panel and underlying data feeds (CoinMetrics on-chain, DefiLlama stables, Farside ETF flows, CFTC TFF crypto positioning, Deribit DVOL, yfinance macro extension) — refreshed daily
- The walk-forward DFM model code with full pre-registration and validation harness — auditable end-to-end
- Three kill-log documents transparently recording every configuration tested and the gates each failed
- The model's current cycle phase + correlation regime read — labeled "research preview" on every reference
NOT published
- CCI as a flagship validated index — the regime call below the fold is research, not a signal subscribers should size off of
- The Repricing family's planned Crypto Futures Sleeve — depends on a passing CCI as input; remains blocked pending v3
- Any quantitative claim that lifts CCI's classification into the methodology pages of the validated sleeves (Core / SN / CB / AW)
Brand-honest position
Macro Sentinel publishes its kill logs. We told subscribers the failed v1, then ran two more pre-registered fixes, then wrote down why those failed too. The model isn't ready for the institutional bar — and the way you can trust we mean it is that we said so on the page where we'd otherwise have launched it.
When v3 ships (paid on-chain data, smaller phase taxonomy, or another year of history makes the substrate work), it'll have the same validation discipline this preview's failure was held to.
Underlying data refreshed nightly via
scripts/refresh_all.py (LaunchAgent at 16:30 ET weekdays). Source caches: cache/alternative_data/coinmetrics/, cache/alternative_data/defillama_stables/, cache/alternative_data/farside_etf/, cache/alternative_data/cftc_tff_crypto/, cache/alternative_data/deribit_dvol/, cache/alternative_data/macro_extended/. Model code: site/python/bci_compute/cci.py. Validation harness: scripts/cci_validation.py. Pre-reg + kill docs: docs/cci_scope.md, docs/cci_v2_pre_reg.md, docs/cci_v1_kill_log.md.