I will tell you this. What you are trying to do is a bad idea almost universally if you feed only stock time series data into a model what you will get out is something resembling a moving average. The reason for this is that the stock market is a complex adaptive system and any strategy used will become less effective the more its used and because every single person who studies machine learning has thought of this i.e. it's a canonical idea it basically will not work at all. (low loss and low percentage error are misleading if you really want an accurate measure of performance check what percentage of the time its right about the direction of movement)