I am working on a market prediction project based on the timeseries data. Essentially i am trying to windowframe the data and predict whether the price goes up or down in an hour, 30 minutes and 15 minutes using LSTM networks (each of these cases is separate problem).
However, I am trying to figure out how to perform the windowframing of the data. Typically, if i try to predict for the next 30 minutes i would windowframe the current market data in 30 minutes window and predict for the next 30 minutes, and then the next windowframe would start 30 minutes later. Yet, i am using quite a complex network with many features and with such approach i get not that many training and test windowframes for 1 hour windows.
I saw some guides for making such prediction that such window should not move by the full window length, but for instance 30 minute window predicting for the next 30 minutes moves by 1 minute, and such network makes 30 prediction every half hour. This way we get much more data, but i am not sure whether lstm is not biased towards current situation and the first 30 predictions on the test set are easier for the network since the training data includes the labels of market value for these frames.
Therefore, my question is what would be the correct way to perform windowframing for the timeseries prediction? Should i shift the window by its full length or is scientifically approved to move it by shorter length and what is then the implication on the end results? Is there some method to estimate how much such window should move for given data? (I would appreciate references to the literature giving such guidelines)