Source: Deep Learning on Medium
Looking into the future is always exciting, especially if it’s the future of markets…
Is It Possible to Predict the Crypto Asset Price?
This story started about a year ago in an ancient European city Lviv. Two old friends were meeting on weekends discussing various global topics. One of them was running a software company and the other one was a data scientist and a university professor. Once they had an intense discussion about the future of crypto and its price predictability. That was a day when everything started…
If it’s possible to predict crypto asset price with some threshold precision — then it’s basically a Philosopher`s Stone people were searching. I wasn’t a big believer that this is possible. I was doing this research more to extend my Data Science expertise in an area of Algo Trading, since unlike some other machine learning areas it can get you a good job and a small chance to win it all at once.
I will not go much into detail of how many thousands of experiments I failed and how many times I was thinking that this is not going to ever happen. However, my friend seemed to believe in success much more than me and his attitude made me come back to this project over and over again.
So one day I landed with a deep learning model that was showing about 500% profit over 2 month period. We were happy that day, «California dreaming» and all that stuff, you know:)
First Successful DeepCrypto Models
Since accuracy isn’t much of a metric that can help with evaluating trading model I had to develop a reasonable estimator. So I took all the predictions from the model and emulated trading.
It works like this: if the model predicted «buy» I just assumed that I buy Bitcoin at «Close» price with 0.1% trading fee. In real life that deal would be closed 10–15 seconds later, but our estimation still can be pretty good for trading amounts under 5 Bitcoin. Similarly, when the model predicted «sell» we took all our Bitcoin balance and converted it into USDT balance.
On our architecture, we were able to build a lot of models which performed excellently in the period of October — December 2018. The best of them was this one: Link to file