this is garbage. the partial autocorrelation clearly shows that successive values are highly correlated, and the model was trained to predict the price at time t using the price at time t-1: the model is basically learning to add a constant to its only input value. just look at the predictions on the test set
As can be seen from PAC, lookback should be 1, larger values give worst result, but perhaps the model remains more stable to overfittting. There is no constant term.
Predicting markets with machine learning is always dangerous, because unlike most types of statistics, markets do not evidently have a single representation that is slowly uncovered; "gravity can flip upside down overnight".
You are absolutely right. Our main idea was to show how RNNs could be used for time series forecasting. We have chosen public bitcoin price data because this problem is well-known for general public. There is no doubt, that such difficult problem cound be solved using only previous prices. But this article may be a good starting point to understand potential methods to approach such problems.
