Great article, in my experience using LSTMs is not as efficient as feature engineering data into "memory" manually. LSTMs don't have a very clear definition of what memory they retain, whereas it's quite easy to give a financial model memory by using Moving Averages of prices, volume, or other indicators. Then, more traditional and explainable ML methods can be used on each "stationary" data point (but which itself contains the memory of whatever previous data is relevant from the MA values)