Research Map

Source: Deep Learning on Medium


Go to the profile of Myron Leskiv

Looking into the future is always exciting, especially if it’s the future of markets…

So Shortly About Research Opportunities…

Each currency pair has 9-time intervals (we don’t take into account time intervals longer than 1 day ). There are about 25 currency pairs on Binance to USDT.

There 3 main Neural Network architectures that we can use: Feed forward, Recurrent and Convolutional (we exclude ensembles of Neural Networks at this point for simplicity :). Also Reinforcement Learning is definetely worth attention.

Each network has hyperparameters like a number of epochs and class weights, which by the way seem to have a significant influence on model profitability. We also use seeds to be able to get reproducible results on each experiment. Otherwise Neural Network will each time start with random initial weights. Training on candles has a very large amount of local optimas and it often converges with significantly different profits. One possible option could be to save each model, but since we run about 10, 000 different experiments a day it can soon lead us to significant storage problems.

Data for Analysis

Ideally one day we want to run a lot of data analyzing streams: few servers running experiments on building technical analysis models like testing the predictability of different technical indicators and their combinations. This alone has billions of possible experiments. Also, we need to test how those indicators work in different neural network architectures, also we need to perform exhaustive grid search of neural network hyperparameter tuning for each architecture for each set of indicators. This requires a lot of time and computational power.

The second data stream is text. This includes developing NLP models that evaluate both tweets and news on the respective asset.

The third data stream is a blockchain activity. All the transactions inside each blockchain could also yield a lot of valuable information about possible price changes.

After we collect all those data streams and build all of those models we will very likely to have enough predictable power to make stable and significant profits on crypto trading.

So total number of possible combinations is really big. We need need to build our software in a way that it can make hundreds of thousands of experiments a day. Ok, lets do it…