CryptAI

From the CryptAI page, access cryptocurrency trends, the CryptRisk Factor, and investment recommendations.
When tested against simulated short-term investments, CryptAI outperformed the baseline by 85.4%.
CryptAI achieved high accuracy with our benchmarks.

Inspiration

Cryptocurrency interest has piqued in recent years, and along with it, a lot of fear and uncertainty. Few actually know what drives cryptocurrency markets and research in the area is lacking. Many people are turned off by the high volatility and risk involved with investing in cryptocurrencies, especially when compared to their stock market counterparts.

We sought to make the mysterious predictable with CryptAI, a model that predicts cryptocurrency price trajectories and, ultimately, drives highly profitable trades.

What it does and how we built it

CryptAI puts deep learning-powered investment recommendations at your fingertips. Using Google Cloud’s Natural Language API, we integrate public opinion with domestic (NASDAQ, S&P500) and international market data to predict the cryptocurrency price trajectories. Our algorithm uses a deep neural network implemented with TensorFlow and Keras. Our probabilistic CryptRisk Factor evaluates risk (0 being low, 10 being high) associated with investing at any given time.

Using CryptAI, we were able to generate a short-term investment strategy for Ethereum with a substantially higher ROI than simulated, uninformed investments.

Challenges we ran into

The existing cryptocurrency models we encountered rely on limited feature sets, so we knew this was a gap CryptAI had to fill. Therefore, the most time-consuming (yet salient) challenge we encountered was scraping data from the web. We dealt with several server timeouts, missing data, and formatting issues that took us hours to sort out, but in the end, we wanted our training dataset to be as comprehensive as possible to give us the best predictions.

Accomplishments that we're proud of

We’re really proud of our ROI metrics, and honestly did not expect our model to perform as well as it did. We were also uncertain if we’d have enough data for our model (since crypto is not as heavily researched as stock market data), but we managed to parse enough to successfully build and tune its hyperparameters. We’re also really happy with the way our web portal turned out.

What we learned

We learned which loss and optimization functions perform best for a given model. We learned how to make requests to the Google Cloud API, and use the Keras framework for both TensorFlow and Theano architectures. And of course, we got to exercise our creative side with some web development at the end.

What's next for CryptAI

We’re hoping to build up the capabilities of our predictive pipeline with larger training sets and a chatbot. We're excited for people interested in investing in cryptocurrencies to use CryptAI as their decision-making platform.