Integrate AI model
#1
I am looking into whether or not I can use Gekko as part of my neural network project and Gekko looks like it could really help me.

My project is looking into whether or not I can build a model that can do a certain amount of forecasting. Gekko does a lot of stuff that I would need to do as in retrieving and organising data from exchanges, preprocessing, strats on buying and selling etc.

What is the best way for me to inject my ANN forecaster into Gekko (of course assuming I have taken care of all the learning parts etc)? Do I create a new custom indicator? You have a lot of info on creating custom "strategies" however from what I can gather thats not what I need to do?
  Reply
#2
It would be super awesome if you could use Gekko for your project! Before we can figure out how to integrate an AI model into Gekko let's have a look at what Gekko is: Gekko is designed to automate strategies and automatically test and run them on crypto markets. In order to do this Gekko consists of a few different parts, when you run Gekko against a live market these are the parts:

[Image: architecture.jpg]
The trading advisor (bottom right) is responsible for taking candles as input and pumping out "advice" signals as output. The trading plugins take this advice and do something with it: using Gekko you can trade by actually creating real orders OR simulating all orders and calculating what would have happened. You can read more on general architecture in the docs.

The trading advisor is the single box within Gekko that does all the decision making, right now it's a simple wrapper around a strategy (that the user can script). I am no expert in AI, but I am understanding that you first need to build a model and after that feed input into this model and get a some kind of a result (a signal or advice if you will). If you could explain a little bit more about how this usually works (assuming you have a model you think is successful), we can figure out how to combine that with Gekko.
  Reply
#3
Thanks for coming back to me. Your understanding of AI is broadly correct. I will be aiming to use genetic algorithms to train a neural network eg AI -> Machine Learning -> ANN; AI -> Evolutionary Computation -> Genetic Algorithms in a process that will happen outside of Gekko. I then plan to insert the trained model into the Gekko pipeline to make use of existing strategies etc

From your description would I wrap my model within a candle creator?

NB: I have some finance knowledge however my background is more in Space and Defence, AI and robotics.
  Reply
#4
I am trying to do something similar by using a neural network to predict the cryptocurrency rates. However I also want to include other custom signals which are not supplied by the crypto exchanges.

Is it possible for Gekko to read these signals from custom data sources, e.g. UNIX or TCP sockets and then to pass these signals to the custom strategy code together with the information from the crypto exchange?
  Reply
#5
What's everyone's thoughts on trying particle swarm optimization?
https://en.wikipedia.org/wiki/Particle_s...timization

https://www.npmjs.com/package/inveniemus
  Reply
#6
(01-28-2018, 03:37 AM)md3inaustin Wrote: What's everyone's thoughts on trying particle swarm optimization?
https://en.wikipedia.org/wiki/Particle_s...timization

https://www.npmjs.com/package/inveniemus

Have you looked at this project? http://gekkowarez.com/gekko-genetic-algorithm/
  Reply
#7
I have never used Particle Swarm Optimization. Are you planning to use it as a replacement for the Gradient Descent to find the neural network's weights and biases?

I think the most important improvement would be to allow external signals to the custom strategy code, e.g. through sockets, pipes or even external files. Without this kind of external input any custom strategy will be limited to doing some kind of technical analysis.
  Reply


Forum Jump:


Users browsing this thread: