A linear model on a heretofore unknown predictor is basically how all hedge funds make money.

Coming up with the predictor is often the hard part. For example, take the tweets of a (sane) president and run sentiment analysis on it. If it is positively correlated with mentioning an equity, the sentiment of the tweet might be a good linear predictor of the stock price.

The math is simple once the feature is well defined.

Feature development is the current frontier, as I understand it.

That's my point. They hire the smart people so they can really abuse the simple stuff.