I have worked with several machine learning methods in the past (mostly with financial data) and my immediate gut reaction is that the training/validatin datase

I have worked with several machine learning methods in the past (mostly with financial data) and my immediate gut reaction is that the training/validatin dataset is far too small to make conclusions that the predictions are accurate. I would guess that the projection the neural network gives will differ enormously based on the position they split the in- and there is no clear description on how the projection model has been trained/validated ie. has some of the out-sample data been used in verifying training results (ie. curve fitting) Financial research is full of examples on overfitting the data to get some kind of result, and this is in most cases not even intentional. Disclaimer: I have not read the paper discussed in detail so I don't know how they have eliminated the overfitting angle.