Results

Generated from docs/RESULTS.md on 28 Sep 2026 14:25 UTC. Negative results are published too.

All numbers are out-of-sample unless stated otherwise. "Hit" = ×2 within 1 h of a follower's entry 2 s after the wallet (fast profile). "Expected" = the market's hit rate for buyers at the same curve positions.

Run 1 — 5 days of history (26–27 Sep 2026), before the full stream

Run 2 — headline (full 30 days, 28 Sep 2026)

Run 2 — full 30-day stream (28 Sep 2026)

8,011,903 signals over 31.4 days; validation = signals after 21 Sep 12:13 UTC (748,074 signals), never used for selection. Fast-profile passports use curve prices and, from 22 Sep, Uniswap v4 pool prices after graduation; the pool-price backfill for the earlier days is pending, so the picker profile (which depends on 24-h peaks after graduation) is measured conservatively here.

Fast profile (×2 within 1 h)

Market hit rate: selection 19.9 %, validation 18.3 %; by curve bucket on validation: b0: 7.2 %, b1: 15.5 %, b2: 23.5 %, b3: 28.9 %, b4: 7.0 %.

H1 · Portfolio. Top-N by score on the validation period vs the market at the same curve positions and vs placebo lists of the same size (50 random draws).

ListVal. signalsWalletsHitExpectedExcessPlacebo excess (mean / p95)
top-255921420.4 %11.8 %+0.086-0.010 / +0.097
top-5032862021.5 %14.5 %+0.070-0.010 / +0.126
top-10033962921.8 %14.8 %+0.070-0.002 / +0.062
top-30054418723.5 %17.8 %+0.057-0.001 / +0.044

Verdict: supported for top-100, 300: excess above the placebo 95th percentile; not for top-25, 50.

H2 · Magnitude. Top-100 vs the market re-weighted to the same curve buckets.

Threshold≥1.3≥1.5≥1.8≥2≥3≥5median
peak 1 h · top-10047.6 %37.1 %26.1 %21.8 %10.5 %4.0 %×1.27
peak 1 h · market33.1 %24.8 %17.7 %14.6 %7.0 %3.1 %×1.07
peak 24 h · top-10042.3 %32.7 %23.0 %19.3 %9.2 %3.8 %×1.28
peak 24 h · market30.8 %23.4 %17.1 %14.4 %7.3 %3.4 %×1.08

drawdown below 0.5× within 10 min: top 21.6 % vs market 14.6 %; median time to the 1-h peak: top 29 s vs market 45 s.

Verdict: supported — the top-100 leads the market at every threshold, i.e. the edge is in the whole distribution, not only at ×2.

H4 · Confirmation. A second top-100 wallet buying the same token within 60 s.

Confirmed signals: 407 with hit 26.3 %; single signals: 2989 with hit 21.1 %.

Verdict: supported — confirmed signals hit more often.

H5 · Stability. Spearman correlation of wallet excess week → next week (wallets with ≥ 10 signals in both).

week 0→1: ρ = 0.34 (n = 12065), week 1→2: ρ = 0.37 (n = 11760), week 2→3: ρ = 0.34 (n = 7163), week 3→4: ρ = 0.37 (n = 2183).

Verdict: supported — skill persists week to week (ρ ≥ 0.2).

Picker profile (pool or ×5 within 24 h, ≥ ×1.5 from entry)

Market hit rate: selection 9.5 %, validation 11.5 %; by curve bucket on validation: b0: 3.0 %, b1: 6.7 %, b2: 11.5 %, b3: 22.5 %.

H1 · Portfolio. Top-N by score on the validation period vs the market at the same curve positions and vs placebo lists of the same size (50 random draws).

ListVal. signalsWalletsHitExpectedExcessPlacebo excess (mean / p95)
top-25130391815.5 %11.8 %+0.036+0.024 / +0.183
top-50161603816.0 %11.9 %+0.041+0.010 / +0.092
top-100217687916.1 %11.8 %+0.043+0.016 / +0.078
top-3002939420417.1 %12.2 %+0.049+0.012 / +0.052

Verdict: not supported — no list beats the placebo 95th percentile on the validation period.

H2 · Magnitude. Top-100 vs the market re-weighted to the same curve buckets.

Threshold≥1.3≥1.5≥1.8≥2≥3≥5median
peak 1 h · top-10043.3 %33.0 %24.0 %20.3 %10.1 %3.8 %×1.20
peak 1 h · market44.5 %34.1 %24.2 %19.8 %8.9 %2.9 %×1.23
peak 24 h · top-10045.8 %36.2 %27.4 %23.7 %13.1 %5.5 %×1.24
peak 24 h · market46.3 %36.2 %26.4 %22.1 %10.8 %3.6 %×1.25

drawdown below 0.5× within 10 min: top 18.5 % vs market 30.8 %; median time to the 1-h peak: top 87 s vs market 62 s.

Verdict: partly — top-100 ahead at 3 of 6 thresholds.

H5 · Stability. Spearman correlation of wallet excess week → next week (wallets with ≥ 10 signals in both).

week 0→1: ρ = 0.51 (n = 10114), week 1→2: ρ = 0.56 (n = 10068), week 2→3: ρ = 0.55 (n = 6026), week 3→4: ρ = 0.51 (n = 1362).

Verdict: supported — skill persists week to week (ρ ≥ 0.2).

H3 (pickers) is the picker block above.

Raw numbers: docs/data/analysis-run2.json (output of src/analyze_full.py).