Do not minmax scale the data. New data outside ranges seen before will be chopped off. Do make relative (high/low), so patterns you find generalize to other domains. Don't fit complex LSTM models on tiny datasets. Somehow this is the shiny thing all starters jump on, while the most simple models are the most proven and robust. Don't just share such a model without giving a disclaimer ("I never tried this out on real data, and would not invest my own money in this solution"). This is real money, real economies, and you could horribly crash these. That is also why you don't write with authority about how to build a bridge (while you never crossed a bridge you've build before). The only reason for the "clearly showing how powerful LSTM models are" is it responding to test set samples it got wrong, and correcting with a lag. How about some evaluation, preferably on another stock than trained on?