Lost when trying to get good time series prediction results (regression problem) even after trying many things


I’m not able to get good results after a long time testing when using TensorFlow to predict time series data (regression problem). I don’t know if the problem is with the data (little quantity and/or low quality) or with the model (or if there is a problem with both) (although I have tried several and tested various combinations of hyperparameters). I’m really tired and I don’t know what else to do. What would you do if you were in my place? Where could I find someone who can help me with this problem? Thanks in advance for any help you can give me.

What type of model are you using ?

Hi. Currently ConvLSTM. Thanks for your answer.

Ah, okay. So not a ‘straight’ regression then.

I mean ConvLSTM is most often used for videos/imagery-- So you’d have to spend some time experimenting with it.

It also depends on what your dataset is. But you might have better luck with a ‘traditional’ timeseries model (i.e. ARIMA, SARIMA, VAR, BSTS, etc) [in particular because there is a direct ‘time’ component in the model which you won’t directly find in a Conv Net).

Yes, I am trying to predict videos, the data are space weather related. I don’t know if it’s relevant, but I use a sliding window to adjust the data, is this something conventional that everyone does when adjusting the data? In another forum, a professor told me that as my data is related to the Sun’s solar cycle, and I only have data from 2008 to 2018 (I use part of 2018 for the test set), I should have much more data to get better results. But as obtaining more data is not such a simple task and will take me a lot of time, I wanted to confirm if this is really necessary. A few days ago I discovered, via histrogram plot, that the beginning of my data, more or less between 2008 and 2010, which takes part of the solar minimum (in the same way that 2018 is also part of the solar minimum), had considerably less data than the rest of the data, which hindered the prediction, since this is an important part of the data. So now I’m thinking what’s the next step I should take. It’s interesting that you comment on these simpler techniques, I take this opportunity to comment that I already tried using MLP a while ago too, but at most it comes close to the best results I get using ConvLSTM. I’ll see if I can test at least some of these techniques. Thanks.

Hello @marcocintra

Was this choice for test data done randomly?

Also can we know the ratio between the datasets.

You could also share some screenshots of the results you are not happy with!!!

For us to know if the above is the issue, we need glimpse into your dataset, model architecture, how you used your dataset and ofcourse what fine tuning measure you took.