This module takes a time series and returns: (a) the piecewise underlying linear trend, and (b) the times where there is a change in the underlying trend
We use a Kalman Filter to find the piecewise underlying linear trend, with a state space representation of seasonality and an underlying linear trend, plus a random walk. We initialise the parameters of the Kalman filter using a Least Squares estimator out of "statsmodels'
For the change detector on the underlying trend we use another Kalman filter, this time a single period autoregression. We then consider the Malhalanobis distance between that filter output, using a Gumbel distribution to decide where the increase in distance likely indicates a change in trend. This implements the approach described by Lee & Roberts at https://www.robots.ox.ac.uk/~sjrob/Pubs/LeeRoberts_EVT.pdf
This module is in PyPi, install via:
pip install ChangePointDetectorEVT
In order to get a Kalman filter of a time series, plus change points in the trend, do the following:
- from ChangePointDetector import ChangePointDetector
- Prepare your time series as data plus Panda dates
- Create the necessary Kalman representation by creating a "session" object by calling the ChangePoint class, e.g.:
Session=ChangePointDetector.ChangePointDetectorSession(data,dates).
- 'SeasonalityPeriods' is an optional input, e.g. if your data are sequentialmonths, "SeasonalityPeriods=12" indicates calendar month seasonality
- 'ForecastPeriods' is another optional input, indicating how many periods to forecast. Default = 3
- Determine the changepoints by running the ChangePointDetectorFunction on your "session", e.g. Results=Session.ChangePointDetectorFunction()
- This will return a "Results" object that contains the following:
- ChangePoints. This is a list of 0s and 1s the length of the data, where 1s represent changepoints
- Prediction. This is the Kalman smoothed actuals, plus a 3 period forecast. Note no forecast will be made if there is a changepoint in the last 3 dates
- PredictionVariance
- ExtendedDates. These are the original dates plus 3 exta for the forecast (if a forecast has been made)
- Trend. This is the linear change factor
- TrendVariance. Variance of the trend
See "Trend.png" for an illustration of this module https://github.com/mhaupt63/ChangePointDetectorEVT/blob/main/Trend.png
