Years ago I started working on some deep learning experiments centered around Automated Trading. I have varying luck with those experiments as I ran out of time to optimize the models and input data before being dragged away. The biggest part of the problem is pre-processing of data for IID format and stability of the model during training. I don't have a lot of time to focus on these activities but see the TensorFlow (tools are getting better which may open them up to experimentation by members of this group. I was using a Ubuntu box for running the older versions of TensorFlow and at the time, exported data from NinjaTrader. I didn't get time to produce another webinar for this as I hoped due to time as well. I wish someone can come up with the 48 hour day so I have more time to tackle some of these interests I have outside of what demands my time.
I have placed this thread in the Automated Trading rather than NinjaTrader specific as this thread should focus more on the Learning Models, Input, Output, and topology of the models. TensorFlow in the 2.0Alpha has started to integrate Agents as well as probabilistic learning. The Agents will help with Reinforcement Learning and the probabilistic learning can be helpful to understand a level of confidence of a specific output. Probabilistic learning would help with risk management (51% sure we should go long vs 75% sure we should go long for example).
I have some older Python libraries which might need updating and some might be able to be replaced with newer TensorFlow API calls. Most of this code is in INVESTIGATION state, so not well documented and have lots of un-needed stuff. I also have some NinjaTrader code (Indicator) which I used for export of training data for TensorFlow (which is really crude).
I would like to help support any interested community members to explore this space of deep learning but need help on the back end getting the tools ready for a broader community access.
I do not want to post the raw code here until it starts to get cleaned up, and the NT Indicator might need to become an ADD-ON due to its nature. I think it just needs a better approach than what I have done as an Indicator. If there are any interested active NT coders and AI interested members, please drop me a note and I can start to get you some of my tools for cleanup. I can probably help answer questions and work to get the code cleaned up and maybe documented for others to use. We will need a base of exported raw data for now, we can use the TensorFlow scripts to predict profitability for now using a simulator. Later we can connect an Inference of models back into NinjaTrader and other platforms.