How score_v1 is built
This page describes the live board formula at a high level. The detailed recipe stays on the server. Past results do not guarantee future returns.
Bayesian win-rate floor
The score uses a Bayesian lower bound of the win rate (the 5th percentile of a posterior), not the headline number, so a short lucky streak cannot carry it.
Permutation test
Each wallet’s picks are compared with a stated null universe: this week’s board, plus closed-trade records of other active wallets when that stored sample is large enough. Otherwise the test uses the board only, and the page says so. p-values are adjusted across the board (Benjamini–Hochberg). Edge is marked only when the adjusted p-value is below 0.05.
Hidden Markov regimes
A two-state (hot/cold) model with binomial emissions is fitted on the whole board. When a wallet has at least two periods and eight picks, we show P(hot) with a 10th–90th percentile band; otherwise the regime is n/a.
Walk-forward and decay
Disjoint earlier versus later windows compare in-sample and out-of-sample hit rates (wins/picks). A documented decay penalty reduces the edge term when the later hit rate falls. A separate two-sample test compares the last 7 days with the wallet’s own earlier record.
God Mode (score_v1) is a 0–100 summary when a window has at least 8 picks; otherwise it is n/a. Missing Copy ROI is omitted, not treated as zero. Copy ROI figures are published snapshot estimates, not this site’s fill simulation. A blank or n/a means there is no value or not enough data.
The public track record stays on “insufficient data yet” until eight weekly snapshots exist.