MIT Computer Scientists Can Predict the Price of Bitcoin

Devavrat Shah and recent graduate Kang Zhang, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and the Laboratory for Information and Decision Systems, claim to have developed a “machine learning algorithm” that enabled the team to nearly double the amount of Bitcoin that they used in the study.
Devavrat Shah and recent graduate Kang Zhang, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and the Laboratory for Information and Decision Systems, claim to have developed a “machine learning algorithm” that enabled the team to nearly double the amount of Bitcoin that they used in the study.

Two computer scientists from MIT claim to have developed a “machine learning algorithm” that can potentially be used to predict the price of Bitcoin.

Stock Market analysis software tools are nothing new to savvy investors. There are in fact dozens of excellent tools online and some of them are even free. It is no surprise that someone thought to apply many of the same principles to Bitcoin, especially considering the wild fluctuations in the Bitcoin market over the last several months.

Devavrat Shah and recent graduate Kang Zhang, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory and the Laboratory for Information and Decision Systems, claim to have developed a “machine learning algorithm” that enabled the team to nearly double the amount of Bitcoin that they used in the study.

MIT

Shah and Zhang used basically the same procedure that was used by Shah in 2012 when published his Twitter study and the team decided that it would work just as well with Bitcoin predictions. After compiling an enormous amount of data from Bitcoin exchanges, collecting more than 200 million data points, they used a technique known as Bayesian regression to “train” an algorithm to locate patterns from the data. They then used the patterns to predict price shifts in Bitcoin over a period of time.

Specifically, over a set period of 50 days, they made a total of 2,872 trades using one Bitcoin for each trade, predicting the average price movement every two seconds over a 10 second period. The test could be considered highly successful because it gave them an 89% return and a risk ratio (return relative to risk) of 4.1.

After setting a baseline, they set the algorithm to buy if the price rose above it and sell if it dropped below. If movement was slight, the program did nothing. The paper, detailing the study, was published this month at the Allerton Conference on Communication, Control, and Computing. Shah explains the method used to predict price:

“We developed this method of latent-source modeling, which hinges on the notion that things only happen in a few different ways. Instead of making subjective assumptions about the shape of patterns, we simply take the historical data and plug it into our predictive model to see what emerges.”

The team used Bitcoin because there is a great deal of free data to draw from and there are many high-frequency traders. “We needed publicly available data, in large quantities and at an extremely fine scale,” continued Shah. “We were also intrigued by the challenge of predicting a currency that has seen its prices see-saw regularly in the last few years.”

Cointelegraph spoke with an experienced commodity trader who preferred to not share his name and while he did find the MIT project interesting, he did have a couple of caveats:

  • Their returns are from a very short time period (50 days). Perhaps this is because that was the only data they had, but I would like to see how robust the model over a larger data set. This could also be made more robust by seeing how the model performs during periods of high volatility (The window they chose was one of the more stable trading periods for Bitcoin)
  • They tested their strategy using 1 Bitcoin of notional.  They acknowledge that further research is required to see if the model can scale. I believe that deploying this strategy in a more meaningful size would greatly reduce returns. There's more impact cost the larger you are.

However, Shah was perfectly clear in noting that there were intangibles, which could not be considered and that the study only looked at the numbers. Unpredictable and spontaneous global events in the real world can certainly impact trading such as new regulations, large scale thefts, potential government bans, and sudden widespread adoption.

These events, which have been giving ulcers to experienced traders since the inception of the stock market, are impossible to predict. Nevertheless, the algorithm could have its uses even for traders in the real world, but depending on its accuracy exclusively might also be a good way to go broke in a hurry.


Did you enjoy this article? You may also be interested in reading these ones: