Deep dives into our methodology, the indicators we track, and the machine learning analysis behind our signal.
No articles match that search.
Each piece carries one of two labels. Methodology · Falsifiable is reproducible method work, built from the raw data and corrected in public when a number moves. Thesis · Directional opinion is an argued view about what is happening or will happen, grounded in reasoning and outside facts rather than the dashboard's validated machinery, and held to the scorecard. Newest first.
The oil test, turned inward on all 25 features including the dashboard's own 13. One indicator failed the same way (BTC Fees: 25.9pp pooled edge, 3.2pp within-year), seven came out stronger than pooled numbers ever showed, and the two indicators kept on economic reasoning alone both passed. Why fees stays, with the condition that would reverse the decision.
Read full article →Crude oil outranked most dashboard indicators in SHAP importance without being on the dashboard. Resolved: low-oil days saw bitcoin up 71.4% of the time pooled, and the edge collapses to zero within-year. The model learned when each row happened, not what happens next. A new caveat for every feature-importance chart.
Read full article →Since December the Fed has ended QT, cut twice, and bought $40 billion of T-bills a month under the name "reserve management purchases": an almost exact rerun of October 2019's "this is not QE". The framing half of the June prediction is behaving; the magnitude half ($189 billion in seven months vs $3 trillion in 2020) is not. The scorecard entry stays open, and bitcoin fell 31% anyway.
Read full article →Strategy's largest bitcoin sale ever: 3,588 BTC at roughly $60,000 against a $75,476 cost basis, to pay preferred dividends it had just raised to 12%. Principal, sold at a loss, to meet a fixed coupon: the sequence published here in June, on schedule. A new entry is on the scorecard.
Read full article →Rising stablecoin supply is crypto's favourite "dry powder" indicator. Tested against DefiLlama's full history three ways, it fails: the 0.84 level correlation is a trend artifact, the flow carries no lead, and the valuation version (SSR) is non-monotonic noise. Not added to the dashboard.
Read full article →An academic preprint (Baquero & Menezes, Porto) ran the full Clauset-Shalizi-Newman protocol on bitcoin's power law. It confirms the price "law" is a trendline, and corrects this website on one claim it should have tested: the return tail is better fit by a lognormal.
Read full article →Bitcoin treasury companies are pivoting to "5-25% yield machines." But bitcoin has no cash flows, so every yield on it is a risk premium in disguise. The four places the yield actually comes from, and why the 2022 lending graveyard is the version that broke.
Read full article →A pretrained tabular foundation model (TabICL) run on the same walk-forward protocol landed at 49.1% on bitcoin direction: a coin flip, the same wall the project's XGBoost and LSTM hit. The bottleneck is the target, not the model.
Read full article →Strategy (MSTR), its STRC preferred, and Terra-Luna are all reflexive machines. But only Terra carried an automatic, unbounded, code-enforced self-destruct. Why "the next Luna" is the wrong frame.
Read full article →The serious version of the Bitcoin Power Law, Giovanni Santostasi's, is a causal mechanism, not a curve fit. Credit where it is due, and the one fitted exponent (adoption as time cubed) the whole tower rests on.
Read full article →US net interest outlays passed defense spending in fiscal 2024 and cleared $1 trillion in 2025. When debt service crowds out the discretionary budget, the central bank stops setting the terms. Bitcoin sits at the bottom of the resulting gravity well, without needing to be named.
Read full article →US M2 ranked top-5 in only 3 of 12 rounds: liquidity's effects are already priced into Power Law and Fees. The model preferred the receipt to the cause.
Read full article →Equity indices contribute only mid-pack, macro-flavored signal: never leading, strongest in the 2022-23 window, redundant with DXY and M2.
Read full article →Three DCA strategies, 12 entry points. Nearly all of the edge is exposure, not timing skill, and the test is in-sample.
Read full analysis →6 scenarios from bull run to 75% crash, 40 heavy-tailed paths each (Student-t, GARCH), 3 DCA strategies through every one. Medians with ranges. The signal amplifies exposure: that's a feature and a risk.
See the scenarios →Jane Street cut IBIT ~71% and FBTC ~60% in Q1 2026. A 13F shows only the long leg, so a basis-trade unwind is the most plausible reading, not a fact.
Read full article →Bitcoin's famous price-vs-time chart is a power-function regression, not a statistical Power Law. But the daily return tail (index ~2.7) and volatility autocorrelation (decays as lag^-0.33) are genuine Power Laws. The label is on the wrong object.
Read full article →A Markov map of the composite: 84% of days bullish, 56% of moves reversed within three days. Includes a visible retraction where a pipeline fix killed the original finding.
Read full article →The in-kind ETF model hands Authorized Participants control over Bitcoin sourcing. Because a liquid CME futures market exists, APs can hedge with paper contracts and delay spot purchases. The argument: this is temporary, with the timing explicitly uncertain.
Read full article →XGBoost and LSTM across three cycles. Corrected ranking counts, plus the anomalies that don't fit, including oil ranking top-5 in 7 of 12 rounds.
Read full article →13 indicators, each graded on a 0-100 scale, combined with SHAP-informed, hand-fixed weights into a single score. From raw data to grade points to weighted average to verdict.
Read full article →The correlation matrix shows the 13 indicators are really 7 dimensions. Why cluster-correction was tested, then dropped.
Read full article →Pick any start date and base amount. Compare Standard DCA, Power Law DCA, and Signal DCA side by side with real historical data.
Run the backtest →Academic and institutional work that informed the approach.
LSTM + SHAP analysis by the Bank of Spain. Key finding: feature importance shifts across market cycles. Technology variables dominated early; attention variables dominate now.
Read paper (PDF) →Comprehensive review finding that most models "perform only marginally better than random guesses," validating our honest expectations about prediction difficulty.
S. Khedr et al., Decision Analytics Journal (2025), Open Access →Santostasi's own statement of the power-law model: a causal feedback loop deriving the exponents from adoption, Metcalfe's law, and mining. Engaged directly in The Power Law, Taken Seriously.
Read the theory →A power-law fit of price against block-time, framed through Taleb's Lindy effect (price ~ Block Years^5.3). Cited by Santostasi as independent confirmation, though the two fits land on different exponents.
Read the model →The full Clauset-Shalizi-Newman protocol applied to bitcoin. Confirms the price power law is a trendline, not a structural law (the fitted exponent runs 5.65 to 16.49 with the time origin); rejects the return-tail power law in favor of lognormal; and shows the trend is still the best long-horizon forecaster.
Read the paper (arXiv) →Every directional call here is dated and scored before the fact. See the predictions scorecard →
New research arrives by RSS feed.