Label hit-rate ledger
Backtest-derived calibration for the breakout and divergence labels rendered on /convergence. Hit = a published signal landed for that entity within 24h of the label being computed. Generated by scripts/backtest-convergence-labels.py; re-run to refresh.
backtest window: 14d · generated 2026-08-11 09:48 UTC
baseline — any converged entity (≥3 sources, 24h window)
78.2%43 / 55 entity-days had a signal in the next 24h
| label | n | hits | rate | lift | vs baseline |
|---|---|---|---|---|---|
breakout ≥3 sources fire AND attention surging ≥15% — mainstream catching up | 2 | 2 | 100.0% | 1.28× | |
divergence ≥3 sources fire AND attention dropping — pre-news lead-time | 28 | 21 | 75.0% | 0.96× | |
unlabeled-converged ≥3 sources fire but no attention trend / flat — the residual | 25 | 20 | 80.0% | — |
Bar: solid = label hit-rate, vertical tick = baseline. To the right of the tick means the label produces signal above what you'd get from any converged entity at random.