Every signal we serve, backtested live and out-of-sample. We show exactly how well each one predicts outcomes — hit rate, lift over base, and sample size. No black boxes.
How often a token actually dumps (peaks <15 min after deploy) when its top-20 early buyers include known dump-cluster wallets. Higher count → stronger signal.
Calibration: when a deployer's as-of runner-rate sits in a bucket, how often do their next tokens actually run (peak ≥60 min)? A trustworthy score should track the diagonal.
Rolling train/test split: signals are derived from older tokens and graded on the most recent settled window — never on data they were built from (no lookahead). We only label tokens with ≥15 early buyers (sparse, barely-traded tokens are unlabelable). Labels are peak-timing based: a token “dumped” if it peaked <15 min after deploy, “ran” if ≥60 min. Numbers update daily and will drift with the market — that's the point.
Programmatic access: GET /api/v1/signals/{name}/performance