BIoTan — example backtest (sample data) Generated 2026-06-03 20:58 · 24 devices · 1 metric(s) · 2026-01-01 00:00 → 2026-03-01 21:00
Read me first. Lead time is an optimistic upper bound: it is computed by hindsight on data that already happened. In live operation the alert threshold must be fixed in advance, so real-world lead time may be shorter. This is a peer-relative, batch backtest: it prioritizes what to inspect — it is not a guaranteed failure predictor. Devices without enough peers are marked not evaluable , and failures with no sustained precursor are reported honestly as no clear precursor .
Metric: power_kw Cohorts (kmeans, silhouette 0.68): cohort 0 : 8cohort 1 : 8cohort 2 : 8
Cohort overview — peer-z (common mode removed) Healthy peers cluster near zero; flagged devices in red diverge.
5 0 -5 2026-01-01 2026-03-01 Flagged devices baseload-power_kw-03 change status: precursor 9.1 d before failure divergence: 2026-02-11 21:00 failure: 2026-02-21 00:00 lead time: 9.1 days before failure
7 0 -7 2026-01-01 2026-03-01 divergence failure baseload-power_kw-05 change status: detected only at/after failure divergence: 2026-02-27 03:00 failure: 2026-02-21 00:00
6 0 -6 2026-01-01 2026-03-01 divergence failure evening-power_kw-03 change status: detected only at/after failure divergence: 2026-02-28 21:00 failure: 2026-02-21 00:00
5 0 -5 2026-01-01 2026-03-01 divergence failure Backtest summary metric device flagged reasons divergence start failure lead (days) status power_kw baseload-power_kw-03 yes change 2026-02-11 21:00 2026-02-21 00:00 9.1 precursor 9.1 d before failure power_kw baseload-power_kw-05 yes change 2026-02-27 03:00 2026-02-21 00:00 — detected only at/after failure power_kw evening-power_kw-03 yes change 2026-02-28 21:00 2026-02-21 00:00 — detected only at/after failure
BIoTan-core · peer-relative batch backtest · 100% local, no telemetry, no data leaves your machine.