The problem I see with ML on stock prices is that stock price is not a function of the various available numerical indicators. Markets are non-deterministic and chaotic systems. Contrast that with driving, where the decision to turn left, right, throttle, etc. is largely a function of the features (surroundings, current speed, etc.).

That said, people are using ML to predict future prices, although no hedge fund claiming this has published their returns.

With regards to reinforcement learning, why not just try it? Numer.ai is an ML competition that gives you free obfuscated stock data and rewards you if your model is successful.