# Is there any time series model which handles data at variable frequencies.?

Goal: Predict the yellow points.(yellow events appear at varying frequencies)

But I'm struggling to find a good model to fit this use case. Most of the time series algorithms are handling data which are at same frequencies(like per day/every 10 secs). I tried a lot of stuff but may be completely on the wrong approach.

Thanks a lot for any hint!!!

• Do you mean that your data is irregularly sampled in time? Commented Dec 21, 2022 at 16:45

It looks like stock price data. You can use fbprophet(python),BATS(R),TBATS(R) to predict the data. You said your model contains different frequencies which mean it contains more than one seasonality. To find seasonality(frequencies) you can use Fourier transform. You should have significant data to get a good prediction

• Hi, there thanks a lot for your comment! Yes, there are some similarities with stock data. Although in this case, the data is much more irregular. I already tried fbprophet but wasn't satisfied with the results. Will try TBATS. Regarding the frequency, I mean that the data points arent distributed homogeneously. So if Y_t are the yellow points, then t != t+1. So in some cases, you have to wait a lot for new data in others you get clusters. Commented Jun 13, 2019 at 10:59
• Ya. It is not distributed homogeneously so it is non stationary data . You should find the lag and season correctly to get perfect prediction. One more things make sure it has collinearity ( a test for time series) Commented Jun 13, 2019 at 11:04
• gist.github.com/tartakynov/83f3cd8f44208a1856ce @MichaelRazum Look into this problem. Similar to yours highly dynamic data. They predict using fft but result may not be quite satisfy you Commented Jun 14, 2019 at 5:32

Little more insight than just a graph is needed. Like what all features along with the type of data. If it's just this graph then

1. Data might be coming from a scientific instrument as per y-axis reading.
2. Try identifying the trend and seasonality and noise by first decomposing it.
3. You can go for LSTM as they are pretty good at handling time series.
• Thanks. Decomposing might be a good Idea. I will try that. I just thought that sharing a graph is the easiest way to get a feeling about which model could handle it. Commented Jun 13, 2019 at 11:08