ntpstats.analysis

Descriptive statistics, detrending and outlier handling for offset series.

detrend(t, x, kind='linear')

Remove a least-squares polynomial trend.

linear removes a constant frequency offset; quadratic also removes a linear frequency drift (aging). Time is centred first so the fit stays well conditioned with POSIX timestamps.

linear_fit(t, x)

Least-squares slope (s/s) and intercept of offset vs time.

theil_sen_slope(t, x, max_pairs=200000, seed=0)

Robust (median of pairwise slopes) frequency estimate, s/s.

Insensitive to up to ~29 % outliers, unlike least squares. Pairs are randomly subsampled for long series.

mad_outliers(x, k=5.0, t=None)

Boolean mask of outliers: |x - median| > k * 1.4826 * MAD.

If t is given the test is applied to linearly detrended data so a frequency offset is not mistaken for outliers.

summary(series)

Summary statistics in the spirit of NTPsec's ntpviz report.

format_seconds(v)

Pretty-print a time value with an SI prefix (e.g. 12.3 µs).

compare(estimate, reference)

Score estimate against a reference (e.g. simulated truth, a GNSS/PPS-disciplined clock or a better server).

The reference is linearly interpolated onto the estimate's times over their common span. Returns error statistics of estimate - reference.