We use a Recurrent Neural Network (RNN) using Google's Tensorflow tool to make our AI predictions, while these are made by blending the best performing more traditional predictive models. Is there anything in particular you were curious in hearing more about?
You are viewing a single comment's thread from:
AI is a hobby of mine, I love it! I also hate it! The secret to any predictive model is dealing with random walk. Filter out the random events to uncover the patterns in the data. It's been statistically proven that free and open markets are random, not just exhibiting randomness, but really random. An RNN or any other kind of Neural Network shouldn't be able to predict it because what they do is predict the future based on patterns from the past and a random system doesn't have a pattern. Once you filter out the randomness there isn't anything left to latch onto.
Anyone that cracks that particular nut has a Nobel prize waiting for them and probably a few billion dollars in licensing agreements coming.
Obviously you can. This means that the crypto markets are not free and open, but have a driver of some kind. I'd be interested in learning what the driver is or if you uncovered several drivers, or is it some pile on effect from algorithmic traders that might be negated if better tools come to the fore.
For the record, I really hate dealing with TensorFlow, I'm impressed you could wrangle it. You're getting an upvote from me just for that!
Thank you for the reply! We do not believe these markets can be predicted, but we also think that there are ways to use predictions on days with very high confidence to produce some positive long-term results. We have a very heavy focus on our data collection, you can view an example of the data that we collect daily here. We also believe that it is theoretically possible to use statistics to have a very good chance of winning even with super simple strategies, but in practical terms that is much much more difficult if not impossible. The reason we show both more traditional predictions as well as AI predictions, is that the ones using the RNN have more bias to them and make decisions that are more differentiated between each other. We believe that the way a neural network learns will work very well (assigning decision weight to individual neurons and learning which ones to listen to) in the long-term because of that bias assuming the dataset was made very well.
Another thing you mentioned is the fact that crypto is a random system and having data that is backward-looking cannot predict the future, and we agree with you so we actually try to track other things that are forward looking as well (like Technical Analysis recommended actions).
Excellent answer. I was expecting a handwavy answer but you got to the meat of it without buzzword soup. That tells me you know what you're talking about, which surprisingly few people actually do. Many people wave AI around like a magic wand, but in many cases it can create more problems than it solves. You seem extremely knowledgeable, I'm glad to know about this project.
We totally agree with you! Thank you so much for the kind words!!