In Stock Market Prediction Based on Generative Adversarial Network by K. Zhang et. al, the authors feed financial data (X0...Xt) into an LSTM to predict Xt+1.
Then, they evaluate whether the series (X0...Xt+1) is real or not (with as Xt+1 either the predicted or the one appearing in the data).
This is done in a GAN system, with the predictor trained to create an Xt+1 so that the series would fool the discriminator.
What is the benefit of training the predictor as a generator in GAN, with the discriminator in providing the loss function, when it would be simpler to use a loss based on |predXt+1 - Xt+1| ?
After all, the aim is to predict the target values - so why not train using the fit of the prediction directly?