LSTM for time series prediction / noisy dataset

[encomes]

Hello,

I am trying to implement a RNN using LSTM to predict values in a time series.

This is, I have around 1400 samples (one sample per each day), and I would like to predict the value in a day given a window of previous values.

When plotted, the dataset looks like this:

Therefore, it is very noisy. I tried using different methods of LSTM (memory between batches, time steps, etc) and have got quite bad results, so I need advice on how to tackle this problem.

I've done other problems where the dataset looks somewhat seasonal or stationary, but those methods can't handle this case, at least from what I have tried.

Any help is highly appreciated.

[Thomas Johnson]

The dataset actually looks quite stationary, it's just that as you pointed out it has high noise. Honestly your best bet here might be a simple moving average. LSTMs aren't magic. Garbage in; garbage out.

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/7ef1a512-2a7f-4f42-bcb7-9f31e8458f97%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[encomes]

Looks way better than before, but I am afraid I might be losing way too much info. Is there any rule of thumb about how big the window for the simple moving average should be, or any relationship with that and the window of values used by the RNN to predict the t+1 value?

[Thomas Johnson]

You can use regular cross-validation to try different window sizes. You might also try an ARIMA model (you can google for tutorial and packages that implement ARIMA modeling)

To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/6ccef5ad-ae8a-4d02-ae05-42c75355fc63%40googlegroups.com.