Or values are adjusted by placing external influences such as news, events, updates, hardforks, and other effects then determining the proportion of each effect in the prediction.
I'd be surprised if they did that. It means they already have the capabilities to allow a machine to judge the value of each news. Maybe it's not that far, I suspect they'll keywords and then categorize each keyword into a value. For example, "broke the channel up", "optimist", and so on into 1, and bear-related keywords like "dumped" and so on to negative on, and so on.
That would be sick.