v0.05 extends v0.04 with level shift (LS) detection, and also fully makes use of apriori hydrostatic pressure information. The latter allows for dynamic threshold determination based on any specified time-interval.
A timeseries always starts with a green color. Levelshifts are represented with colors: red if the LS is up, and blue if it is down. The start and endpoints are represented with dotted vertical lines. Outliers are represented with red dots.
Here, beside the LS, there are also 2 outliers.
For inbo data, varying time-differences play a major role:
In the last LS, the timeinterval is a couple of months, and based on apriori information, the difference is deemed a LS, whereas here it is not:
An other clear example:
There are other cases which are falsely detected as LS. Some are TC, and since v0,05 has no degrees of freedom for TC, they are falsely considered LS. Others seem borderline cases or bugs and still need to be analysed.
Some advantages vs. tsoutliers:
speed
no need for data aggregation
apriori knowledge
varying time-intervals possible
Disadvantages:
not "off-the-shelf": needs development and testing
ad-hoc optimization: maybe better or worse, depending on case
For modeling and optimization, v0,05 is largely inspired by Chen, C., e.a., Joint Estimation of Model Parameters and Outlier Effects in Time Series, 1993.
v0.05 extends v0.04 with level shift (LS) detection, and also fully makes use of apriori hydrostatic pressure information. The latter allows for dynamic threshold determination based on any specified time-interval.
Results: inbo and geotech
A comparison with
tsoutliers: geotech and inbo.Some examples.
A timeseries always starts with a green color. Levelshifts are represented with colors: red if the LS is up, and blue if it is down. The start and endpoints are represented with dotted vertical lines. Outliers are represented with red dots.
Here, beside the LS, there are also 2 outliers.
For inbo data, varying time-differences play a major role:
In the last LS, the timeinterval is a couple of months, and based on apriori information, the difference is deemed a LS, whereas here it is not:
An other clear example:
There are other cases which are falsely detected as LS. Some are TC, and since v0,05 has no degrees of freedom for TC, they are falsely considered LS. Others seem borderline cases or bugs and still need to be analysed.
Some advantages vs.
tsoutliers:Disadvantages:
For modeling and optimization, v0,05 is largely inspired by Chen, C., e.a., Joint Estimation of Model Parameters and Outlier Effects in Time Series, 1993.