Loading

HYBRIDTEST


outlier gts
Since v1.0.0
Available on all platforms
See also

The HYBRIDTEST function detects outliers in a GTS (or a LIST of GTS) which has a seasonal part.

Like STLESDTEST, HYBRIDTEST performs an ESDTEST onto a GTS that have been relieved of its seasonal and trend part. But unlike the mentioned test, STL and ESDTEST are performed piecewise. Plus, the trend is approximated with the piecewise median instead of the trend part of the STL decomposition.

A LIST of ticks (or a LIST of LIST of ticks), corresponding to the outliers, is pushed back onto the stack.

This technique was first developped at Twitter.

This function only applies to bucketized GTS of type DOUBLE.

References

  • Owen Vallis, Jordan Hochenbaum, Arun Kejariwal. “A Novel Technique for Long-Term Anomaly Detection in the Cloud”, Twitter Inc (2014).
  • Cleveland, Robert B., et al. “STL: A seasonal-trend decomposition procedure based on loess.” Journal of Official Statistics 6.1 (1990): 3-73.
  • Rosner, Bernard (May 1983), “Percentage Points for a Generalized ESD Many-Outlier Procedure”,Technometrics, 25(2), pp. 165-172.

Signature

Examples

// Macro used to generate an approximately normal distribution (using central limit theorem) <% RAND RAND RAND RAND RAND RAND + + + + + 3.0 - %> 'normal' STORE // we generate a GTS with an approximately normal distribution [ NEWGTS 1 100 <% NaN NaN NaN @normal ADDVALUE %> FOR // we add outliers (> 3.0 in absolute value) // Note that we do this before adding seasonal and trend components 25 NaN NaN NaN -3.9 ADDVALUE 36 NaN NaN NaN 3.8 ADDVALUE 71 NaN NaN NaN 4.01 ADDVALUE 82 NaN NaN NaN -3.21 ADDVALUE DEDUP // we generate a periodic GTS of mean 0 NEWGTS 1 100 <% NaN NaN NaN 4 PICK 10 % 4.5 - ADDVALUE %> FOR // we generate a trend GTS (piecewise raise) NEWGTS 1 100 <% NaN NaN NaN 4 PICK 50 / 1 + ADDVALUE %> FOR ] // we sum up the 3 components: remainder, seasonal and trend [ SWAP [] reducer.sum ] REDUCE 'sum' RENAME // bucketize [ SWAP bucketizer.first 0 1 100 ] BUCKETIZE 0 GET // we call HYBRIDTEST 10 5 2 HYBRIDTEST