The price of Bitcoin is notoriously volatile. In the past week alone it’s dropped as much as $40. In the week before that, it fell as much as $70.
But an MIT professor says it is possible to predict the cryptocurrency’s price movements.
In a new paper, Devavrat Shah, along with recent MIT grad student Kang Zhang, write that over 50 days, their 2,872 trades on their model gave them an 89 percent return on investment.
To do so, they needed lots of computing power to do it.
They worked by using Bayesian regression analysis, a form of probability testing that finds patterns in datasets from fields that are subject to seemingly random variables. Shah has previously used this method to track trends on Twitter, successfully predicting trending topics with 95 percent accuracy an hour and a half before the social media site labeled them as such.
In this case, Shah and Zhang collected 200 million datapoints from Okcoin.com, a large exchange operating in China, between February 2014 and July 2014, taking snapshots of how prices moved in given time intervals.
Surprisingly, they found the price movements often resembled patterns commonly found in stock price movements: triangle patterns, and head-and-shoulders patterns:
“There are only a few ways in which price movements of Bitcoin happen, and the model captures most of the changes and uses them to predict them,” he said.
Shah said he used Bitcoin because the data was free.
Despite having full faith in the model, Shah said he was surprised that it produced such robust results.
“I was expecting it to be successful, but not that much,” he said.
Shah said he’s gotten inquiries from finance types about working as an investment consultant, but has declined so far.
“Right now I’m not running a hedge fund,” Shah said.
Rob covers business, economics and the environment for Fusion. He previously worked at Business Insider. He grew up in Chicago.