Recommended data cleaning techniques for multivariate time series prediction?

Question

I have to predict the next step(s) in a multivariate time series with about 30 features and 50.000 samples. I am thinking of using LSTM. Which techniques are usually recommended for cleaning the data when using LSTM?

Does it make sense to transform the data into a stationary time series when using LSTM? Should the data be normally distributed when you are using PCA?

There is also a very large amount of missing timestamps. Does it make sense to impute/fill (by forward filling or something else) big gaps or is it just better in that case to ignore the missing data completely?

Answer 1

3 questions:

1. Does it make sense to transform the data into a stationary time series when using LSTM? Always Stationarity is always desired property and data should be transformed (read more)

2. Should the data be normally distributed when you are using PCA? No There are multiple assumptions around PCA IF you use some matrix factorization techniques. For example if using SVD you should make sure that your matrix is of full rang.

3. Does it make sense to impute/fill (by forward filling or something else) big gaps or is it just better in that case to ignore the missing data completely? If for certain features a lot of data is missing you should drop them all together. Dont try to impute it, you will add false information.

Answer 2

1. Not sure. I think LSTM dont need the stationarity assumption as traditional stat method. I have tried to use it to predict on a dataset with complex temporal and spatial correlation and it works fine. But you should consider normalize the data, such as MaxMinNormalization. Not sure whether the diff will improve the results.
2. Not sure what you are asking. Do you want to use PCA for preprocessing? PCA itself doesn't need normal distribution assumption. But I dont think using PCA then make prediction would be a good idea.
3. No. If you have a few gaps, you can impute them as many papers do. But if there are a very large amount of missing timestamps, drop them.