Skip to content

Detection methods

A point is an anomaly when it falls outside what was expected. The Method decides how Detect computes that expectation. There are two families: methods based on a confidence interval, and methods based on a threshold you set.

All methods compare a point with a baseline of past points, except Forecast. See Baseline.

These methods compute an interval from your history. Its width depends on the Level of confidence (95% by default): the higher it is, the wider the interval and the more certain a flagged point is a real outlier.

The default and most generic method. Detect takes the points of the baseline, computes their average and their standard deviation for each KPI, and builds a confidence interval following a normal distribution. A point outside the interval is an outlier.

For a ratio, the deviation is weighted by the denominator of the step. The baseline needs at least 3 points.

Same as observed volatility, but on the variation of the KPI from one point to the next, rather than on its value. Use it for a KPI with a trend, where the value keeps moving but the variations stay regular.

It needs a date as the dimension to Detect anomaly on, and at least 3 points in the baseline.

A model forecasts the latest points from the history, taking seasonality into account (yearly, monthly, weekly), and builds a confidence interval around the forecast. A point outside the interval is an outlier.

Forecast always analyses the latest points only, and needs at least 10 points of history. Same forecast for all points makes one forecast for all the latest points, instead of one per point: fewer forecasts, faster.

These methods compare each point with the baseline and flag it when the gap crosses the Threshold you set.

Method A point is flagged when
Relative variation threshold Its variation from the baseline, in %, is above the threshold. For example +30%.
Absolute variation threshold Its difference from the baseline, in the unit of the KPI, is above the threshold. For example +50 units.
Absolute value threshold It crosses the threshold value while the baseline is on the other side. For example above 80 while the baseline is below.

An absolute threshold flags big values more easily than small ones: with a weekly seasonality, weekends get flagged less often. Use the relative threshold when the level of the KPI varies a lot.

You want to Use
Monitor a KPI with no prior knowledge Observed volatility
Monitor a KPI with a trend Variation volatility
Account for seasonality in the latest points Forecast
Get alerted past a business rule A threshold method