Why ML Attempts to Predict the Stock Market Are Meaningless

Most of the articles about AI predicting the stock markets focus on a more or less complicated model which tries to predict the move of the next timestep. And almost all of them are just focussing on architecture and layers like LSTM or CNN. But then they are just using the Mean Squared Error (MSE) as a loss function -- and this is a problem and this article shows you why. Assume we have the following Deep Neural Network (DNN) implemented in PyTorch. Note that we use log returns because these should follow more closely a Normal Distribution which is an important detail for this article, as you will find out in a minute.