Neural Networks for time series

Question

I have an understanding problem. I am a beginner in machine learning and have also a little experience in modelling NNs but not for time series.

But I cannot imagine how to use Neiural Networks for time series.

So if I want to train a Multilayer NN what is the input?

I read several papers about such ANN to predict time series. But they did not explain how actually.

Can I imagine it as follows: Is every input node a specific time stamp?

If I want to predict the value for time t+1, and I use 15 input neurons, so I use the values at each time stamp from t-15 upt to t?

How do I train such a NN?

I am a little confused.

Answer 1

The type of neural networks you are looking for to predicting timeseries are called recurrent networks; the previous neuron states in the network affect the new neuron states. Examples of such recurrent networks are LSTM, GRU and NARX.

If you have the data from t₀ to t₁₀, and you want to predict t₁₁, then you need to input the values from t₀- t₁₀ one by one into the network. You have to train the network so that the output of t_n is t_n+1. So after you have inputted t₁₀, the output will be the values predicted for t₁₁.

Real life example: I want a network to keep on decreasing it's output until it has reached 1, and then make it output 0 to start over again:

in: 0.0, out: 0.2
in: 0.2, out: 0.4
in: 0.4, out: 0.6
in: 0.6, out: 0.8
in: 0.8, out: 1.0
in: 1.0, out: 0.0

And that is what your training data should look like

If it is still to complicated, I recommend you to play around with this front-end neural network library for the browser: Neataptic. It is very easy to fiddle around with. Example