Profit Overflow
Project Description

This was a fun project that aimed to use machine learning to model and predict the (highly-volatile) trends in popular cryptocurrencies such as Bitcoin or Ethereum. We gave the bot $10,000 to start with and optimized it to maximize earnings on value-spikes and minimize loss on sudden drops. After training our high-frequencing trading algorithm, we were surprised to find that the bot made a significant profit (>10%) in the few hours we allowed it to run. (sadly, the money was only simulated)

01. The Goal

Inspired by the surging value in cryptocurrencies, my team and I wanted to develop a bot that could accurately model and predict trends in stock or coin prices, and buy/sell with the optimal strategy. Because of the potential complexity of finding this strategy, we decided a machine learning model would best suit the problem. Another goal for the project was to make the bot trade with extremely high frequency, which would allow it to keep up with the rapidly fluctuating values that are especially present in cryptocurrencies.

02. Problems & Solutions

Our initial model was a convolutional neural network (CNN). The model recorded the value of the cryptocurrency over 10 1-minute intervals and used these as inputs into the network. We did this to give the model the ability to learn from information over a period of time so that it could (hopefully) predict whether the price would rise or fall over the next interval. This model struggled a lot, and didn't perform well in predicting trends. Instead, we looked to a long-short term memory (LSTM) technique that would learn about trends continuously as time went on. With this method, we had a much greater model for predicting the optimal investment strategy.

03. Final Product

Although the trading strategy could have been further optimized, we focused mostly on the machine learning aspect of the project. This resulted in a model that clearly showed the ability to predict trends in the stock price, and taught us a ton of stuff about machine learning that we never expected to learn! By the end of the hackathon, we were able to accurately predict minute-by-minute trends in the price of Bitcoin and make buy/sell decisions accordingly. Given just over an hour, the bot was able to earn over 10% in profit. We were thrilled to have successfully implemented the LSTM model, and even more excited to receive first place overall!

Interested? Read more about it on Devpost!