Lab · ML Experiments

ML — Pattern Discovery

Inverted workflow: find conditional edges in BTC data first, build strategies second.
55 experiments

Funding extreme reversal

Inconclusive
2026-05-17 fundingreversalmacro
Hypothesis
Extreme funding rates predict directional reversals in the next 24h: very negative funding → squeeze higher; very positive funding → mean revert lower. Tested walk-forward on funding z-score quintiles.
Verdict
**INCONCLUSIVE** — Q0=-34.7 bps (t=-0.8), Q4=-8.9 bps (t=-0.3). Magnitudes hint at vol-driven returns (both tails positive) rather than clean directional mean-reversion. Re-examine: is this an absolute-return effect rather than a directional one?
Q0_t_stat
-0.7690
Q4_t_stat
-0.2876
n_windows
7
Q0_mean_bps
-34.7361
Q4_mean_bps
-8.9396

Funding extreme reversal

2026-05-17 · status: inconclusive · 1.2s

Hypothesis: Extreme funding rates predict directional reversals in the next 24h: very negative funding → squeeze higher; very positive funding → mean revert lower. Tested walk-forward on funding z-score quintiles.

Verdict: INCONCLUSIVE — Q0=-34.7 bps (t=-0.8), Q4=-8.9 bps (t=-0.3). Magnitudes hint at vol-driven returns (both tails positive) rather than clean directional mean-reversion. Re-examine: is this an absolute-return effect rather than a directional one?

Key metrics

metric value
Q0_mean_bps -34.7361
Q0_t_stat -0.7690
Q4_mean_bps -8.9396
Q4_t_stat -0.2876
n_windows 7

Approach

We attach the 8h funding rate to 1m bars, compute its 30-day rolling z-score, then sample one observation per calendar day (so consecutive fwd-24h windows do not overlap). Each daily observation is bucketed by funding z-score quintile, computed within each walk-forward training window only (no in-sample leakage) and applied to the corresponding test window.

Daily observations: 1,067

Walk-forward windows: 7

Pooled OOS per quintile

Each test window contributes one mean per quintile. We then pool across windows: mean = mean of window-means, se = std/sqrt(n_windows). t_stat measures consistency across windows.

q mean_bps se_bps t_stat n_windows total_obs
0 -34.74 45.17 -0.77 7 98
1 -26.32 34.33 -0.77 7 121
2 -11.3 24.05 -0.47 7 148
3 -13.28 30.69 -0.43 7 156
4 -8.94 31.09 -0.29 7 118

Long-Q5 / short-Q1 spread per window (bps fwd-24h)

window 0 1 2 3 4 spread
1 -261.93 10.6 -7.76 2.34 -39.93 222
2 121.53 81.29 29.88 -14.74 85.98 -35.55
3 -76.52 65.18 4.15 -167.58 19.26 95.78
4 -65.66 -114.29 74.13 88.42 -27.89 37.78
5 16.45 17.89 -40.79 55.05 -143.32 -159.77
6 17.55 -120.28 -8.38 -33.12 90.71 73.16
7 5.43 -124.65 -130.36 -23.32 -47.39 -52.82

pooled quintile returns

spread per window