crypto trained

BTC price is the clearing price of flows through a fragmented, levered, 24/7 venue graph. It is not the output of fundamentals.

• Slow layer (months to years): supply schedule and miner economics, holder cost basis (realized cap, MVRV), stablecoin and ETF rails, and the global liquidity and risk regime. This layer sets the regime; the market-mechanics guide owns it in depth.
• Fast layer (minutes to days): perp leverage (OI, funding, basis), order book depth, and forced flows (liquidations, margin calls, redemptions). This layer sets the path and nearly all tail days, and market-mechanics covers it in the same depth.
• Narrative is the story told after the fact. Treat it as a hypothesis that must be linked to flows, never as the explanation itself.
• The measurement is the product: fake volume, venue quirks, timestamp units, and revisions decide whether a "signal" exists.

The working method is the reasoning-framework guide: explain the move, judge the signal, verify the mechanism.

The market-lab examples run all of this end to end and are verified. The minimal leak-free funding join below was verified:

The epoch fix for data.binance.vision: ts // 1000 if ts > 1e14 else ts, then unit="ms" (the to_utc helper).

• Explain moves in sigma units and in UTC minutes before you tell a story (framework A of reasoning-framework).
• Treat funding, basis, and OI as crash-risk and positioning variables, not as direction signals. This agrees with BIS WP 1087 and with H1 below.
• Always test the reverse direction and the base rate, and count independent episodes.
• Use data.binance.vision for bulk history (no rate limit, reproducible files) and REST only for the recent tail, as the binance public data readme recommends.
• Pin every claim to a venue, a sample window, and an access date. For features and models, route to feature-engineering, ml, and xgboost. For storage and pipelines, route to olap and data-pipeline.

• The data is unusually open. Full tick and kline history, funding, and 5-minute OI metrics come free from Binance. On-chain data is public. There is a free Coin Metrics community tier.
• Leverage mechanics are exchange-documented formulas, so a cascade can be verified minute by minute (H2).
• 24/7 trading gives a large number of observations for microstructure work.

• Cycle-level questions (halvings, MVRV bands) have n≈3-5. No statistics can rescue them.
• Regimes shift. Funding's predictive IC decayed toward zero, and macro correlation rose after 2020.
• Vendor on-chain metrics (LTH, SOPR, exchange flows) depend on proprietary attribution that gets revised. Glassnode definitions were not opened here.
• The best causal data (liquidation feeds, order book depth history, ETF flows by the hour) is paid or real-time only. Binance has no historical liquidation dump for USD-M: liquidationSnapshot returned 404 for BTCUSDT.

1. data.binance.vision spot timestamps are microseconds from 2025-01-01 (1760140800000000, 16 digits). Futures files are still ms. A mixed concat silently lands spot data in the year 57,000 or produces NaT.
2. Spot vision CSVs have no header row, futures vision CSVs do. Reading both with the same header= drops a row or turns the header into data.
3. Funding fundingTime has ms jitter (1790496000001). An exact join on the 8h grid misses rows, so floor to the hour.
4. startTime=0 on /fapi/v1/fundingRate returned only the latest page (500 rows, from 2026-04). A naive paginator "finishes" with 5 months of data and no error. Start from the perp launch (2019-09-10). The docs say the default without time params is the latest records.
5. Funding is clamped. 23.4% of days sit exactly at 0.01%/8h (the interest-rate default, Binance FAQ). pd.qcut fails on duplicate edges, so rank first. Funding can also switch to hourly in stress, so check interval counts.
6. Levels correlate spuriously. Stablecoin supply vs price shows +0.85 in levels. The prediction is +0.14 and not significant.
7. Naive p-values on overlapping windows lie. 7d funding IC: naive p = 0.017, Newey-West t = -0.93.
8. BTC's daily close (00:00 UTC) is asynchronous with the US equity close. Daily correlations are understated, and weekend BTC moves land on Monday equity rows.
9. "Active addresses" are not users, and "exchange outflows" include custody and ETF reshuffles. Treat both as attribution-dependent.
10. Aggregator volume is mostly fake on unregulated venues (more than 70% wash, NBER w30783).
11. Paper findings are sample-bound. "No exposure to stocks" (NBER w24877, 2011-18) is false for 2020-26 (H7).

• Binance REST 429 / 418. Weight is 2400 per minute per IP. Ignoring a 429 escalates to a 418 IP ban. data.py honors Retry-After (API docs).
• /futures/data/openInterestHist serves only the last 30 days. Older OI comes from vision metrics daily files (5-minute rows, create_time as a string).
• Binance docs and FAQ pages return HTTP 202 (a JS challenge) to curl. Read them with a browser or WebFetch.
• fred.stlouisfed.org timed out (HTTP/2 INTERNAL_ERROR, then a read timeout) from this network. Yahoo chart returned 429, and stooq served a JS proof-of-work page. The keyed API api.stlouisfed.org/fred/series/observations is reachable (400 "api_key is not set" if no key); a free key is the path, fred is not blocked.
• mempool.space returned 429 and timed out. Use the Coin Metrics HashRate/FeeTotNtv instead, or retry with back-off.
• The Coin Metrics docs URL .../market/mvrv is a 404 (moved). The community catalog endpoint lists the free metrics (CapMVRVCur, FlowInExNtv, SplyExNtv, and others).
• Deribit DVOL history IS free: https://www.deribit.com/api/v2/public/get_volatility_index_data?currency=BTC&start_timestamp=<ms>&end_timestamp=<ms>&resolution=3600 (duckdb read_json_auto + unnest(result.data) → 721 hourly rows for 30 days).
• Binance quarterly (dated) futures are archived on data.binance.vision under their contract symbol (e.g. BTCUSDT_261225). Resolve live symbols from fapi/v1/exchangeInfo (symbols starting BTCUSDT_). Guessed expired symbols return empty listings. Gives a free dated-futures basis.
• CFTC CoT bitcoin (financial futures): the financial futures file (https://www.cftc.gov/files/dea/history/fut_fin_txt_2026.zip, FinFutYY.txt) has 5 bitcoin-named markets; only BITCOIN-CME and MICRO BITCOIN-CME are CME, the other 3 are Coinbase Derivatives. Filter with ilike: a case-sensitive LIKE '%BITCOIN%' silently drops "Nano Bitcoin" (172 vs 134 rows for 2026).
• Kalshi api.elections.kalshi.com/trade-api/v2/markets?series_ticker=KXBTCD is duckdb-readable (unnest(markets)), no custom user agent needed.
• Paywall drift: defillama /emissions and bridges now return 402 (paid); beaconcha.in returns 401 without a key. Coin Metrics community free tier has 9 usable btc metrics (AdrActCnt, CapMVRVCur, FeeTotNtv, FlowInExNtv, FlowOutExNtv, HashRate, IssTotNtv, SplyExNtv, TxCnt); no sopr/nupl/lth/puell free.
• Free liquidation history: none. Binance liquidationSnapshot archive and allForceOrders both 404. OKX public liquidation orders are recent-only.

All cases ran with python 3.12, pandas 3.0.6, numpy 2.5.3, scipy 1.18.1, and statsmodels 0.15.0; the scripts are the market-lab examples.

• H1: high funding predicts lower forward returns. Weakly supported in 2019-21, not supported since. The sample is Binance BTCUSDT, 2019-09-10 to 2026-08-28, 2,545 days. Median annualized funding was 8.4% (p95 48.8%), and 12.7% of days were negative. Funding autocorrelation is 0.83 at lag 1, so it is highly persistent. Spearman IC: 1d -0.022, 7d -0.047 (naive p 0.017, Newey-West t -0.93), 30d -0.078 (NW t -0.84). The 7d quintiles are non-monotonic (q1 +1.9%, q3 -0.95%, q5 +0.55%). The reverse direction is stronger: past 7d return vs funding is +0.238, so funding mostly records what price already did. By era, the 7d IC went -0.148 (2019-21), -0.049 (2022-23), -0.018 (2024-26). In the top 1% of funding days (n=26), mean fwd7 was -0.5% vs +0.56% overall. Verdict: use funding as crowding and crash-risk context, not as a return predictor.
• H2: 2025-10-10 was a leverage cascade. Supported. Binance perp OI went from $10.78b at 20:45 UTC to $9.36b at 21:30 (-13.2%), and to $8.20b a day later (from $11.78b the day before). Price fell 116,976 to 101,516 at 21:20 (-13.2%). The spot low was 102,000. Perp-spot basis reached -51 bp at 21:19. Peak 1m volume was $1.14b, 143x the median. The taker ratio fell to 0.49. Price was at 113,597 by 23:00, 78% retraced. The 5m tape shows OI falling in step with price (10.60 → 10.14 → 9.40). The tariff-headline trigger is background knowledge, not measured.
• H3: the halving cycle peaks about 12-18 months after. Pattern held 4/4 in timing. Not a tradable law. Peaks came 371, 525, 546, and 534 days after the halvings. Peak multiples were 92x, 30x, 8.5x, and 1.9x (log10 1.96 → 0.28, shrinking every cycle). Drawdowns from peak were -85%, -84%, -78%, and -53%. The 2024 cycle peaked on 2025-10-06 at 124,824, fell to 58,525 on 2026-06-30, and stood at 84,386 on 2026-09-26. The base rate is that 62-72% of all days had a positive 1y forward return. The pattern rests on n=3 complete cycles, cannot be separated from global liquidity cycles, and is a public calendar.
• H4: "MVRV < 1 = bottom, > 3.5 = top". Direction supported, thresholds drift. Median forward 1y return by bucket: <1 +123% (99% up, 775 days), 1-1.5 +112%, 1.5-2.5 +76%, 2.5-3.5 +24%, >3.5 +4% (51% up). But MVRV > 3.5 happened only in 2011, 2013, 2017, and 2021. The 2024-25 top peaked at MVRV 2.78, so the classic top line never fired. There are about 4-5 independent bottoms. Verdict: a valuation regime gauge, not a timer.
• H5: hash rate leads price. Not supported. On 155 non-overlapping 30d samples, spearman(hash-rate growth, forward price) = -0.043 and the reverse is +0.053. Both are noise at this horizon.
• H6: stablecoin supply growth is dry powder that predicts BTC. Not supported as prediction. Monthly, 2019-03 to 2026-08, n=90. Same-month co-movement is +0.294 (p 0.005). Next month is +0.137 (p 0.197). BTC last month → supply this month is +0.262 (p 0.013), so supply reacts to price. The tercile means look monotonic (-0.8%, +3.7%, +7.2%) with n=30 each, which is suggestive and not significant. Supply was $312b.
• H7: BTC trades as a risk asset. Supported since 2020, as a regime. On 2,288 aligned days, BTC vs Nasdaq COMP correlation was +0.31 (by year: 2017-19 about 0, 2020 +0.46, 2022 +0.59, 2025 +0.46, 2026 +0.45; rolling 90d from -0.32 to +0.67). Beta was 0.90, with volatility 2.9x the Nasdaq. BTC vs UUP was -0.08 overall (rolling from -0.54 to +0.36; 2022 -0.35). Weekend mean absolute return was 1.4-1.9% vs 2.4-2.8% on weekdays.

• Free data: data.binance.vision, Binance spot and USD-M REST, Coin Metrics community, DefiLlama, blockchain.info charts, mempool.space, and nasdaq.com JSON.
• Paid data: Glassnode, CryptoQuant, Kaiko, Amberdata, and Tardis for depth and liquidation history.
• Tooling: ccxt for multi-venue REST (the awesome-quant-crypto list catalogs more), DuckDB for Parquet analytics (olap), and the awesome-bitcoin list for protocol tooling.
• Community correction: the digital-assets guide says "higher NVT suggests overvaluation" and presents the halving as "scarcity" without noting n=4 and the shrinking multiples. Treat both as hypotheses, per H3 and H4.