Outlier Detection Calculator — Advanced

❓Outlier 🏠Home 🧮Basic Video
📘 How to use this — copy and paste CSV, TSV, or spreadsheet data

What this page adds. The basic calculator scores up to three columns in your browser. This page takes any number of columns, adds two formal tests - Grubbs' test and the generalized ESD test. It flags points and never changes them: what to do with an outlier stays your call. To do that it sends your table to the server, computes there, and sends the numbers back. Nothing is stored, but the data does leave your machine. If that matters for your data, use the basic page. The outlier detection article explains each method.

A flag is a question, not a verdict: look at each flagged point before you remove it. Grubbs and generalized ESD assume normally distributed data; a long tail (response times, sizes, concentrations) is flagged by every method here - take the logarithm first.

  1. Paste your data into the Input grid — straight from a spreadsheet, or as CSV/TSV text — or import a file. First column is t (time, or a row number), then one column per series: y1, y2, … Rows are sorted by t. Each column is judged on its own, and needs at least 3 numbers.
  2. Pick a method.
    • z-score, modified z-score, IQR, Hampel: a score per point, flagged above the threshold - as on the basic page.
    • Grubbs: tests the point farthest from the mean; if it is an outlier, removes it and tests again.
    • generalized ESD: runs the same test up to max outliers times without stopping, then counts the outliers - not fooled when outliers hide each other.
  3. Set the threshold (score methods): 3 for the z-score, 3.5 for the modified z-score, 1.5 for the IQR, 3 for Hampel - filled in when you pick the method. Window (Hampel): points on each side.
  4. Set alpha (tests): the chance of flagging a point in clean data, per test - 0.05 by default. Max outliers (generalized ESD): an upper bound, below half the column.
  5. Real time range — try settings live on the number boxes.
    • ▲▼ buttons or the mouse wheel over a box: the value moves and the result is recomputed - one request a moment after the last step.
    • Digit wheel: over the threshold and alpha boxes, the wheel steps the digit just left of the mouse pointer.
    • Limits: threshold 0.1 to 20, window 1 to 500, alpha 0.001 to 0.2, max outliers 1 to 100; a typed value outside is set to the nearest limit.
    • Typing or the arrow keys change the box only; press detect to recompute.
  6. Press detect. The server's time for the detection appears beside the button. The upper chart shows your data, with each column's threshold as dotted upper and lower lines in a lighter shade of its color, in the data's own units - for a test, mean ± the critical value of its last step times s of the points kept. A point beyond them is an outlier; for Hampel the lines follow the window. The lower chart shows each score, with the outliers marked in a darker shade. The status line lists the flagged t for each column that has any.
  7. Read the result. The Output grid lists t, then per column its score and a flag (1 = outlier, 0 = not). Export saves it as a CSV/TSV file.

Input

Output

Signal

Score