How to obtain the autocorrelation of an unevenly sampled time series (containing longer gaps)?

194 Views Asked by At

Problem: Obtain the autocorrelation of unevenly(!) sampled time series.

Data (Time series) description: ~$10^6$ unevenly sampled time series, each containing ~$10^4$ data points at irregular intervals.

Distribution of the interval (gap) length:

  • 12 seconds : most common (minimum)
  • 12-18 hours : very common
  • 1-5 days : occasionally
  • 1 week - 3 months : rare

The data is periodic (but the amplitudes are not constant) with periods from 1 hour - 100 days and thus on the same order as many of the gaps. The shape of the signals is not always sinusoidal.

Standard approaches (that do not work):

  • Lamb-Scargle Fourier Transform: This method does not work since the signals are not sinusoidal, but can contain sharp peaks etc.

  • (Linear) interpolation (then use standard time series analysis): This method does not work because the gaps are too large and too common, since whole oscillations are missing.

  • Gaussian process regression/ Kernel methods: They can be made to work - however they are unfortunately much too slow to compute the autocorrelation of all ~$10^6$ time series.

EDIT: I am at this moment unable to publish any sample data sets publicly.