In my problem I have a longer period of historical data of a time series. I need to predict for some specific points in time in the future. For these points in time five previous values are also available.

So far my approach was to use a sliding window of size five, use lag features and apply machine learning methods.

However I have a feeling that when doing like this I am not exploiting the historical data to the full extent. (The methods see only one sliding window at a time.)

I am now looking for some method (or ideas to design my own) which takes as an input historical data and measurements just before the time point I need to predict for.

Thank you!

  • $\begingroup$ Do you have the same measurements available for the future point in time, for which you already have the 5 mentioned values? $\endgroup$ – geekoverdose Jun 9 '16 at 13:09
  • $\begingroup$ @geekoverdose Yes, I have exact same data for historical and future points in time. $\endgroup$ – Beginner Jun 9 '16 at 15:22

As you have the same information available in your past data as for your future data (both in the form of time series), you could use a number of sequence data/time series based model types to predict your target variable. The core difference is that those models "remember" information between samples, so don't only derive the output on from the $N$ input features of the current input, but also from input they've seen before. Such models range from e.g. Hidden Markov Models to Recurrent Neural Networks - but there are many more suitable model types, with different advantages/disadvantages depending on the details of your problem.

PS: have a look at e.g. this answer or this slideset for some more details and further references.

  • $\begingroup$ The problem is: I have a big gap in between. The data looks like this: 5 years of historical data, then gap, then 3 month, let's say May, June, July and I need to predict for August. I haven't seen models which can swallow such a big gap in between. $\endgroup$ – Beginner Jun 9 '16 at 19:37
  • $\begingroup$ I see, your assumption is that the relation of month $m_1$, $m_2$, $m_3$ to $m_4$ is the same for historical data (5 years ago) and current data. If this assumption is correct, then the model should learn to predict month $m_4$ from $m_1$ to $m_3$ with both historical and current data. I don't see why this should not conceptually be possible using sequence based models (you don't use the 5 years gap explicitly). $\endgroup$ – geekoverdose Jun 10 '16 at 12:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.