I am trying to use time series neural network to predict future values. I have time series data from 2010-2014 and I need to predict the values from 2015-2020 using time series neural network. I am using encog library in java. I read the workbench time-series example, but I am a little confused. I think I should have both the training and evaluation datasets to use the neural network, but in my case I don't have the evaluation dataset because these values will occur in the future. How I can do this?
$\begingroup$
$\endgroup$
5
-
$\begingroup$ Guys I really need your help if its not clear please let me know $\endgroup$– toztoz toztoztozJul 22, 2015 at 15:44
-
$\begingroup$ If your question is how to build a testing or validation set, you can simulate the future by splitting the data based on time, rather than taking a random subset. Is there something else to your question? $\endgroup$– Sean EasterJul 23, 2015 at 0:09
-
$\begingroup$ @SeanEaster thank for reply but my question is how to continue the predictions beyond the current data $\endgroup$– toztoz toztoztozJul 23, 2015 at 5:08
-
$\begingroup$ Ah, I see. You may have more luck attracting answers if you specify more about your data and objective. It sounds like you have a single time series variable and you want to train a neural network in a sort of auto-regressive fashion, but that is not clear from your question. Links to the resources you're using and how they're falling short may also help other CV members understand your question. $\endgroup$– Sean EasterJul 23, 2015 at 16:03
-
$\begingroup$ It's not clear what your model is. One option would be to take (e.g.) 3 years as input and predict the 4th as output. E.g. [2010 2011 2012] -> [2013]. Then for future data, at each iteration simply take the last predicted value as one of the inputs. For the first few samples, you have known targets, so you can use them for evaluation. $\endgroup$– dcorneyJul 24, 2015 at 15:01
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) can be used to learn time-series data. Using Mean Square Error as a loss function it is possible to predict the future values of a time series. For a python implementation using Keras have a look at the following ipython notebook.
