I understand what you say is true for you. On the surface yes it’s only the change in price that has the possibility of profit. With that said, it is also true volume leads price. It’s certainly a distillation and simplistic reduction of a complex concept. What’s makes it more difficult to discern is that is operates on negative logic with price permitting or killing the measuring of volume. Price and volume form a relationship and do not exist in liquid continuous dual auction markets without each other. If volume doesn’t mean anything, then your prolly killin’ it in the forex markets where no accurate reflection of market volume exists.
Firstly, pick the appropriate development language. For automated tasks or web services or scripts, Python is a 'great' choice, however it is not suitable for enterprise application requirements. This means bigger projects / services / application frameworks which may have multiple modules, specific disparate hosting requirements and library support etc - which in turn need some robustness around abstracted logic's/deployment/maintenance etc, and where the money/time spent in it should be optimal. With that constraints, I highly recommend switching from Python to C# asap. Whilst you will lose a lot of learning and training material reference samples along the way - these can easily be converted to C#, and if you know python the move will not be too painful, so best to get it done early. I'm not a python 'hater' and on paper it has some advantages which make it 'easier' for many people to 'quickly work with or hashout and test a basic function'. This comes at a big cost however imo in terms of overall capability and ability to handle complex internal application 'routing'. What 'tree' (unsupervised, supervised, reinforcement learning) do you wish to explore? All can be used to built an ATS, but will produce different style of useful trading outcomes and have different data preparation requirements. ML is not an 'exact science', and as such experimentation with data and run time variables is the key to success - just don't expect it to be the holy grail. Checkout https://bhrnjica.net/ for some samples in c#. One last tip. Your models (neural networks) should be per instrument (don't try to build a one size fits all model), and persistent (keep your brain state 'alive' 24/7 365 days a year).