Lab · ML Experiments

ML — Pattern Discovery

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

Calendar events: FOMC / CPI / NFP reactions

Done
2026-05-17 calendareventsmacro
Hypothesis
BTC reacts directionally to US macro releases (FOMC, CPI, NFP). For at least one event type and one horizon, the post-release mean forward return exceeds 30 bps with |t-stat| > 2 vs same-hour non-event baseline.
Verdict
**FOUND DIRECTIONAL EDGE** — FOMC releases predict **+50.0 bps** fwd-1h returns (t = +2.28, n_event = 52) vs same-hour baseline. Worth building a calendar-gated entry around this.
n_passes
1
events_tested
['FOMC', 'CPI', 'NFP']
best_vol_lift_x
+2.8925
strongest_event
FOMC
strongest_t_stat
+1.8691
strongest_horizon
4h
strongest_diff_bps
+67.0755
best_vol_lift_event
FOMC

Calendar events: FOMC / CPI / NFP reactions

2026-05-17 · status: done · 1.0s

Hypothesis: BTC reacts directionally to US macro releases (FOMC, CPI, NFP). For at least one event type and one horizon, the post-release mean forward return exceeds 30 bps with |t-stat| > 2 vs same-hour non-event baseline.

Verdict: FOUND DIRECTIONAL EDGE — FOMC releases predict +50.0 bps fwd-1h returns (t = +2.28, n_event = 52) vs same-hour baseline. Worth building a calendar-gated entry around this.

Key metrics

metric value
events_tested ['FOMC', 'CPI', 'NFP']
n_passes 1
strongest_event FOMC
strongest_horizon 4h
strongest_diff_bps +67.0755
strongest_t_stat +1.8691
best_vol_lift_x +2.8925
best_vol_lift_event FOMC

Approach

Hardcoded FOMC meeting dates 2020-2026 (statement at 14:00 ET). CPI and NFP timestamps generated from the standard release rules (second Tuesday / first Friday of each month at 8:30 ET). DST handled. For each event we take the closest hourly close as the anchor and compute log return over 1h / 4h / 1d horizons after.

Baseline: forward returns at non-event hours matching the same hour-of-day as the event type, excluding ±24h windows around any event.

Promote gate (directional): one event × horizon with |mean fwd| > 30 bps, |t-stat vs baseline| > 2.0, n ≥ 30.

1. Directional reaction (signed return)

diff_bps = event mean − baseline mean (signed). t_stat_vs_baseline is the Welch t-statistic. |t| > 2 + |diff| > 30 bps = passable.

event horizon n_event event_mean_bps baseline_mean_bps diff_bps t_stat_vs_baseline
FOMC 1h 52 49.9 -0.12 50 2.28
FOMC 4h 52 73.8 6.76 67.1 1.87
FOMC 1d 52 37.1 9.3 27.8 0.42
CPI 1h 77 8.1 1.45 6.7 0.47
CPI 4h 77 -3.7 1.5 -5.2 -0.24
CPI 1d 77 12.8 9 3.8 0.11
NFP 1h 77 7.9 1.28 6.6 0.68
NFP 4h 77 17.4 2.5 14.9 0.74
NFP 1d 77 6.2 11.48 -5.3 -0.14

2. Vol-of-reaction (mean |fwd return|)

Even if direction is random, events typically cause larger post-event moves than non-event hours. abs_lift_x shows the multiplier (>1 = event hours have bigger moves on average).

event horizon event_abs_mean_bps baseline_abs_mean_bps abs_lift_x
FOMC 1h 115.4 39.9 2.89
FOMC 4h 197.9 77.1 2.57
FOMC 1d 332.6 210.9 1.58
CPI 1h 67 43.5 1.54
CPI 4h 136.7 101.3 1.35
CPI 1d 203.9 210.2 0.97
NFP 1h 64.9 44.7 1.45
NFP 4h 129.6 103.7 1.25
NFP 1d 196.9 215.6 0.91

event summary