BTCUSDT — Data Overview

🎯 Mai-2026-Welle ABGESCHLOSSEN: Roadmap · Synthesis-Report — 9 Experimente, 4 promoted (BOCPD, ETF-Flow, DVOL/VRP, Master-LGBM, R²+10.6pp), 2 pursue, 3 dropped.

Generated by ml/overview.py on 2026-05-17 09:32 UTC. Source: Binance USDM 1m futures via backtesting/data/.

This report is the entry point for the ml/ pattern-discovery module. It is descriptive, not prescriptive — every conditional edge spotted here must be re-validated walk-forward in ml/experiments/ before feeding a strategy.

1. Data coverage

• Range: 2020-01-01T00:00:00+00:00 → 2026-04-28T23:59:00+00:00 (2,309 days, 6.3 years)
• Bars: 3,326,400 (expected 3,326,400, completeness 100.000%)
• Gaps > 60s: 0 (max gap 60s)
• Zero-volume bars: 369 (0.0111%)
• Price range: $3,707 → $126,087
• Funding: 2023-05-12 → 2026-05-13 (25,773 settlements)

Per-year coverage

2. Return distributions across horizons

Log returns at multiple horizons. Vol_annual is the per-bar std scaled by sqrt(periods_per_year). High excess kurtosis indicates fat tails; positive skew means more positive shocks than negative.

3. Volatility regimes over time

Annualised realized vol from 1m close-to-close returns, rolling over different windows. Look for regime breaks and clustering.

Annualised vol per year

4. Autocorrelation: returns vs |returns|

Returns themselves should have near-zero autocorrelation (efficient market). |returns| typically show strong positive autocorrelation (volatility clustering). The gap between them is the signature of GARCH-style dynamics.

• Lag-1 acf(return) = -0.0453 (mean-reversion)
• Lag-1 acf(|return|) = +0.3851 (strong vol clustering if > 0.1)

5. Time-of-day & day-of-week effects

Mean and std of forward 1h log returns grouped by UTC hour and by weekday. t-stat = mean / (std / sqrt(n)) — rough significance check.

⚠️ Caveat: consecutive 1m bars produce overlapping 60-bar forward windows, so observations within the same hour are heavily correlated. The reported n overstates the effective sample size by roughly 60×, which inflates t-stats. Treat large t-stats as a screening signal — confirm walk-forward in ml/experiments/ with non-overlapping samples before believing.

By UTC hour

By day-of-week (Mon-Sun)

6. Volume & activity patterns

Where does the action concentrate? Volume rhythms hint at when liquidity providers vs. takers dominate. Also a sanity check for time-of-day return effects.

Avg volume per 1m bar by UTC hour

7. Funding rate analysis

• Coverage: 2023-05-12 → 2026-04-28 (1,559,519 bars with funding attached)
• Mean per 8h: +0.153 bps
• Std per 8h: 0.417 bps
• Annualised (×1095): mean 1.68% / std 4.56%
• Min / Max single settlement: -8.18 bps / +5.98 bps
• % bars with negative funding: 13.81%

Forward-24h return by funding quintile (sampled at 8h settlements)

Quintiles use rank-based ties when funding-rate buckets collide on the default value. |t-stat| > 2 suggests the bucket's mean is unlikely zero — but settlements 24h apart can overlap, so treat as a hint not proof.

8. Bull / bear regimes & drawdowns

Top-10 drawdown episodes (peak-to-trough within sample)

Biggest 10 daily moves (non-overlapping 1440m windows)

9. Where to dig next

Concrete hypotheses worth testing in ml/experiments/ after reading this:

1. Vol-clustering exploit: ACF(|returns|) is the strongest non-zero autocorrelation we have. Test: predict realized vol over the next h hours from features, use that to size positions or filter entries.
2. Time-of-day conditional return: if any hour shows |t-stat| > 3 (see §5), check whether the effect is stable across walk-forward windows — overlapping samples make raw t-stats optimistic.
3. Funding extreme reversal: most-negative funding quintile vs. forward 24h return (see §7). Classical 'long when shorts are paying' thesis — test if it holds with walk-forward.
4. Regime clustering: HMM or k-means on (rv_1d, ret_24h, vol_z_1d) to find 3–5 distinct market states. Then look at conditional forward returns per state.
5. Volume-shock mean reversion: bars with vol_z_1h > 3σ — do they predict short-term reversal or continuation?
6. Range-compression breakout: low hl_range over N bars followed by directional move — quantify base rate and expectancy.

Each of these gets its own folder under ml/experiments/ with a README.md, the code, IC + bootstrap CI numbers, and a verdict.