time series
Suppose I have some time series. I already removed seasonality, so it consists of trend, some events and error.
I need to find the events, their length in periods,and their strength they had on the trend.
My idea is to divide data into several "windows", and calculate the length and strength
Do you know of a similar technique?
There is a function borrowed from the TA community, called a ZigZag indicator (which can be adjusted for noise thresholds). It is available in the TTR package of R.
It might be possible to do the transformation(s) using classical time-series decomposition methods, but I don't know of a simple way to adjust the noise thresholding as in the zig-zag function.
Some Googling for specifically neural networks and seasonality leads to this paper, Neural network forecasting for seasonal and trend time series, Zhang and Qi, European Journal of Operational Research, V.160, 2, 16 January 2005, 501–514. In this paper the authors sought to compare the Box-Jenkins approach with a neural network approach. From the abstract:
We find that neural networks are not able to capture seasonal or trend variations effectively with the unpreprocessed raw data and either detrending or deseasonalization can dramatically reduce forecasting errors. Moreover, a combined detrending and deseasonalization is found to be the most effective data preprocessing approach.
They conclude that accounting for trend and seasonality in your preprocessing steps is a good idea.
Best Answer
Ivan,
No. I will point you to the work of Tsay www.unc.edu/~jbhill/tsay.pdf
You might not have removed the seasonality properly. Did you look for seasonal pulses?
As for your outliers, there are outliers, there are level shifts, there are changes in trend as well as seasonal pulses. You need to run a tournament of models with different approaches where you might need to look for the outliers first and then seasonality. There could also be changes in parameters and variance that need to be tested for. Consider many many options.....post one data set here and perhaps we can do a deep dive?