NR Dynamic Baselining Algos

gistfile1.txt

New Relic automatically discovers seasonality and evaluates the use of 4 algorithms for unique time series slices. Further, we've built an extensible ensemble algorithm that we continue to add to. In regards to seasonality, we've applied a technique common in signal processing, called Fast Fourier Transforms (FFT’s). FFT’s can be used to identify the underlying frequency in a time series. We use FFT’s to sniff out good candidates for 'seasons', then try the candidates on the baseline bootstrapping data and see if works better than the default. Currently in our ensemble approach we evaluate four options: triple exponential smoothing with the discovered seasonality, triple exponential smoothing with the default seasonality (Holt-Winters), double exponential smoothing (i.e. look only at recency and trend factors) and single exponential smoothing (just look at recency). Lastly, end users do not currently have access to modify these dynamic baselining algorithms.