Neural Networks and machine learning is just starting out but there are lots of code projects to use on time series data (stock data). I have seen neural network code run in a variety of scenarios and pattern recognition is a strong area but in absolutely no way are they a crystal ball into the future. My opinion is that there are about a dozen implementations for time series data that can work. They all have distinctly different approaches. My implementation uses a method called LSTM (Long Short Term Memory) to try to predict a forward looking daily trend to forecast 2 or 3 days into the future. My implementation of LSTM is providing a second opinion about the direction of the short term trend. By using Machine Learning as a final check in executing a trade, the overall system sees a 3% to 9% increase in success.
The Neural Networks in my implementation is an application that I run after the system has generated a trade idea. I run it this way because it's a time consuming process and it's an expensive process with regards to computer resources. Machine learning is an exciting topic because it's evolving fast and it has such huge developer base.