Using Bitcoin Pricing Data to Create a Profitable Algorithmic Trading Strategy

Bitcoin, especially as of late, has been an incredibly high rising, although incredibly volatile, currency. Because of this, an algorithmic trading model that can make accurate predictions of short-term market trends can take take advantage of the spikes in the bitcoin market while avoiding the sharp decreases, allowing for substantial gains. In this project, we created an algorithm that would predict the price of bitcoin in x minutes relative to the current price, where x ∈ {5, 10, 20}. Our investment strategy would then be to repeatedly invest for x minutes if the algorithm stated that the price increased, and do nothing otherwise. We modeled this problem as a classification problem, where the two categories were based on whether the price increased or decreased. Three models were used: a simple logistic regression, a logistic regression after Principal Component Analysis, and a neural network with one hidden layer with a ReLU activation function. In all three cases, our loss function was a weighted logistic loss function. We found that each model did better than the previous one, where our metric was the expected amount of gains we would make in x minutes following our strategy. All three models significantly beat the baseline of the average increase in bitcoin price in x minutes, suggesting that their implementation might lead to a profitable trading algorithm.