Detecting trends in time series

In the previous recipe, we covered changepoint detection. Another class of algorithms can be used for trend detection, that is, identifying significant and prolonged changes in time series.

The kats library offers a trend detection algorithm based on the non-parametric Mann-Kendall (MK) test. The algorithm iteratively conducts the MK test on windows of a specified size and returns the starting points of each window for which this test turned out to be statistically significant.

To detect whether there is a significant trend in the window, the test inspects the monotonicity of the increases/decreases in the time series rather than the magnitude of the change in values. The MK test uses a test statistic called Kendall's Tau, and it ranges from -1 to 1. We can interpret the values as follows:

• -1 indicates a perfectly monotonic decline
• 1 indicates a perfectly monotonic increase
• 0 indicates that there is no directional trend in the series

By default,...