The detector suite
Every drift detector, explained honestly.
One page per detector: the exact update rule, a live playground running the same code the cloud suite ships, and a straight answer to the question vendors dodge — what is actually guaranteed, and what only holds if the textbook model happens to be true.
Warning tier — classical control charts
Warning tier: classical control charts — sensitive and fast, but calibrated against a model rather than a guarantee, so warnings arrive unbudgeted. These are the detectors practitioners already run — each earns its keep somewhere, and each textbook calibration quietly fails somewhere else, which is why they warn here rather than page.
- CUSUMWARNcumulative sum control chartCUSUM adds up small deviations from a baseline until their cumulative sum crosses a threshold — the classic fast detector for persistent mean shifts, with model-based calibration that is not a guarantee.Rule, playground & honest label
- EWMA chartWARNexponentially weighted moving average control chartAn EWMA chart smooths a metric with exponential weights and warns when the smoothed value leaves its control limits — sensitive to small persistent shifts, calibrated against a model rather than a guarantee.Rule, playground & honest label
- Static thresholdWARNfixed threshold / Shewhart-style limitA static threshold fits mean ± k·σ once on training data and fires on any crossing — the rule behind most dashboard alerts: unbeatable on big spikes, blind to slow drifts, and silently miscalibrated on real data.Rule, playground & honest label
- Rolling bandWARNtrailing-window anomaly band / adaptive thresholdA rolling band re-fits mean ± k·σ on a trailing window every point — the rule behind “adaptive” anomaly detection: great on sudden local anomalies, and structurally blind to slow regressions, which it absorbs.Rule, playground & honest label
Page tier — anytime-valid e-processes
Page tier: e-process alarms delivered inside a stated false-alarm budget that holds no matter how often you look. Their false-alarm control is a theorem about the statistic, not a simulation under a model — which is what makes a page from this tier news rather than noise.
- E-detectorPAGEShiryaev–Roberts e-detectorThe e-detector is an anytime-valid change detector: a Shiryaev–Roberts statistic built from e-values whose average time to a false alarm is guaranteed by theorem, with no Gaussian or independence assumptions on your metric.Rule, playground & honest label
- Coverage e-processPAGEanytime-valid coverage certificateThe coverage e-process is an anytime-valid monitor on conformal coverage: the probability it ever pages while coverage is genuinely healthy is capped at δ — a false-alarm budget that holds at every look simultaneously.Rule, playground & honest label
Why two tiers, not one winner?
Because the honest answer to “which detector is best?” is a trade, not a name: classical charts are usually faster on abrupt breaks but cannot budget their false alarms on realistic data, while e-processes deliver alarms inside a stated budget at the cost of some detection lag. ValidAnytime runs both tiers on every stream and labels every alarm with the tier that produced it — sensitive warnings you tune, budgeted pages you trust.