What we’ve learned so far about AI Crypto Trading

Objective:

The objective of this project is to train a neural network to trade Bitcoin and other cryptocurrencies. We will use reinforcement learning to train a neural network, using machine learning libraries such as TensorFlow.

We will train the network on historical cryptocurrency price data. Once trained the network will be evaluated on its trading during a period that was not part of the training data.

In addition to creating the neural network itself we will connect it with the Binance Exchange API and show that it can execute trades in real-time on a paper trading account.

The Challenge:

The problem of predictive trading of cryptocurrencies can be classified as a subset of time-series forecasting. This is a well-researched field that impacts multiple industries such as stock markets, weather forecasting, pandemic monitoring, and business strategizing.

In terms of designing a neural network, the defining characteristic of our time-series data is that order of inputs matters. This is in contrast to say a neural network that is trained for basic object recognition – where the order of the images should not have an impact on the model’s performance. With time-series data the order of data is critical and provides further selection guidance for both traditional forecasting techniques as well as neural network design.

Our group will be selecting a Recurrent Neural Network model for our implementation.

Recurrent Neural Networks

“Recurrent neural networks – specifically long short-term memory (LSTM) networks and gated recurrent units (GRUs_- are good at extracting patterns in input data that span over relatively long sequences.” [1]. What makes the recurrent neural network architecture unique is that inputs can map not only to an output, but also map a hidden state that feeds back into the network. [1] In simpler terms, this provides a memory-like quality to the network to allow mapping of input data over time to an output – which is ideal for time series data.

 Fig 1: Representation of a recurrent neural network. [1]

In our project, we will explore both LSTM and GRU networks if time allows to see which produces a better model for our application.

Citations

[1] Lazzeri, F. (2020). Machine learning for time series forecasting with Python. John Wiley & Sons.