The researchers used Bayesian regression – a family of mathematical techniques for making optimal predictions based on incomplete input information – for predicting the price of Bitcoin. Based on this price prediction method, they devised a simple strategy for trading Bitcoin. The strategy can nearly double the investment in less than 60 day period when run against real data trace.
Similar techniques have been previously used to forecast which news topics on Twitter become trends, and detected such trending topics in advance of Twitter 79% of the time.
The information in the historical data related to the price of Bitcoin that can help predict future price variation in the Bitcoin and thus help develop profitable quantitative trading strategies. Technical analysis assumes that price movements follow a set of patterns and one can use past price movements to predict future returns to some extent. For example, previous studies found that some simple geometric patterns empirically developed by analysts, such as heads-and-shoulders, triangle and double- top-and-bottom, recur frequently in price charts and can be used to predict future price changes. Other patterns emerge from a theoretical model involving two distinct groups of traders with different assessments of valuation.
The researchers used Bayesian regression inspired by the so-called Latent Source Model to model underlying patterns for the purpose of better prediction, without explicitly finding them. They simulated the detailed trading strategy described in the paper with historical pricing data from Okcoin.com, one of the largest exchanges operating in China, with over 200 million data points between February 2014 to July 2014. The results show that the automated trading method yielded an 89% return in 50 days.
The figure below plots two time-series – the cumulative profit of the strategy starting May 6, 2014 and the price of Bitcoin (in blue). The scale of the Y-axis on the left corresponds to the price while the scale of the Y-axis on the right corresponds to the cumulative profit.
With a typical academic sobriety the researchers admit that the simulation only used token investments and state that more research is needed to find out whether similar results would be obtained with higher, real-life investments. On the other hand, the results could be improved upon by analyzing larger amounts of historical data that, of course, would be more computationally intensive, but not prohibitively so.
It seems therefore that the age of algorithmic trading, performed at lightning speed by sophisticated machines of vast computing power, is dawning on the world of Bitcoin.
If these techniques become common, and all existing algorithms have a similar performance, the profitability of algorithmic trading would decrease. But there is always the possibility that an algorithm with much higher performance is found and kept secret, and then it would be the only player in the top league. Existing algorithms are too “simple” to qualify as Artificial Intelligence, but future versions may be based on ground-breaking results of AI research.
What do you think? Comment below!
Images from Shutterstock.