I wrote my thesis last year comparing different RNNs against each other using this exact paper as baseline and basically concluded that you would be better off predicting the price yesterday than using their results. Authors did not respond when prompted for implementation details or comments.

Overall, concluded that amongst RNNs the GRU architecture proved most favorable but still would not outperform simple stochastic models of the financial industry toolbox.

You can check it out here https://github.com/jensgrud/financial-forecasting-lstm/tree/...