I am currently working on a market forecasting project based on the current news feed. Thus, for every windowframe i extract articles that were released in that time and I detect a number of topics that these articles relate to. Therefore, i end up with for instance 300 numerical features representing the topics discussed and 10 related to the market. Then, on top of that i would like to train an LSTM network to predict whether price in an hour goes up or down.
However, i have noticed that there is no clear approach of feature selection for such memory based classifier. As far as i understand, to make a decision the network not only makes use of current windowframe but also the information about past windows stored in the network. Therefore, most of the feature selection approaches (correlation based, information gain etc) cannot be applied, because the final decision is not based on a single sample and windowframe from past week might actually have a high influence on the current prediction. Moreover, i suppose the approach to apply in that case would be testing the accuracy of the classifier on the validation set for different sets of features, however, this approach requires the validation set.
My question is, Is there some smart way to perform feature selection for LSTM like classifiers, which does not require testing the classifier for different sets of features, but rather can tell how useful is specific feature for a memory based classifier? If there are some approaches i would appreciate some references to the literature.