On fitting and forecasting the log-returns of cryptocurrency exchange rates using a new logistic model and machine learning algorithms