Highspeed algoritmic trading is already running most of the tradevolume. I believe it will be very hard to use machine learning on this data unless you sit in the nexus between trades and can algotrade in the interim between when a buyer sends his order and the sellers accepts. There are «clearinghouses» in between who has better access to realtime data than ordinary endusers, so you will always loose. They make their money on both ends of the trade, from the seller and the buyer, plus can do multiple trades in between.

But perhaps take a look at sentiment analysis. A few years back someone made code that based their automatic trades on Trump tweets.

World events / news can cause massive shifts in stock value, which might be tradeable.