Resistant Smoothing of Sequences (3RSSH, Hanning, Reroughing)

Description

Split a sequence into smooth and rough using repeated running medians of 3, splitting and hanning, so outliers fall into the rough and stay visible.

Tukey's resistant smoothers separate a series into a smooth and a rough (data minus smooth). Running medians of 3 are repeated until nothing changes (3R), so isolated spikes cannot drag the curve. End values are set with the end-value rule: the median of the end value, the next smoothed value and a one-step extrapolation. Two-point flat peaks and valleys are split and resmoothed (S), and hanning, a 1/4-1/2-1/4 running mean (H), softens the result; the usual recipe is 3RSSH. Fences on the rough flag outside points. Reroughing, or twicing, smooths the rough again and adds it back when the smooth must follow the data closely. Across a genuine break, stop at 3RSS or smooth each side separately, because smoothing across it hides the change.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Hours

Steps

  1. Choose the expression of the values (raw, logs or roots) so that variation is roughly even.
  2. Take running medians of 3 and repeat until there is no change (3R).
  3. Set end values with the end-value rule: median of the end value, the next smoothed value and a one-step extrapolation.
  4. Split two-point flat peaks and valleys and resmooth twice (SS).
  5. Apply hanning (1/4, 1/2, 1/4) if a smoother curve is wanted (H).
  6. Compute the rough, set fences on it and flag outside points for investigation.
  7. If the smooth must follow the data closely, smooth the rough and add it back (reroughing or twicing); do not smooth across a known break.

Sources

  • Exploratory Data Analysis (Tukey, 1977)
  • Order Statistics (David & Nagaraja, 2003)

Also known as

  • 3RSSH
  • Running medians
  • Median smoothing
  • Hanning
  • Twicing
  • Reroughing
  • Tukey smoother

Categories

  • Analytics & Behaviour
  • Statistics & Modelling