Can Machine Learning Predict the Future of Cryptocurrency?
Cryptocurrency markets are far better understood now than they were a few years ago.
Nevertheless, they still represent a new asset class for investors, and it’s fair to say that we’re still trying to figure them out. Part of the process of doing so now involves exploring whether or not machine learning might help predict crypto movements.
In some respects, we have actually seen machine learning put to use in the crypto sector — specifically with regard to security and fairness. Our article on how AI enforces the crypto market covered this idea, and revealed that AI is being “employed to detect trading irregularities” within the market.
The article noted that the Ronan AI program detected trading irregularities by analyzing data around the time when a Goldman Sachs announcement sparked a 10% decrease in heavily traded cryptos. While similar programs may not be policing the crypto market specifically, they are keeping an eye out for problematic patterns.
The reason that this has become a popular question, beyond the fact that crypto markets are still new, is that cryptos are extraordinarily volatile. A high-value cryptocurrency such as bitcoin can easily drop $2,000 in two days.
But at the same time, a report showed a 16% jump in bitcoin’s value in 24 hours just last month (with cryptos overall rising $23.8 billion over the same period of time). These wild swings make many traders nervous, and naturally make the idea of AI-driven predictions appealing.
CFDs are essentially contracts in which a trader attempts to forecast whether a given cryptocurrency will gain or lose value in relation to another currency (either fiat or crypto). It’s still tricky, but in a way it’s less precise than price trading. In theory, if machine learning can predict even general trends in a market, something like CFD trading could become very appealing.
The article first mentioned time series forecasting methods (such as ARIMA, DeepAR+, and Facebook’s Prophet), which it called “easy to implement but not very resilient.” Basically, the market variations in cryptocurrency overwhelmed these methods.
Next, the article covered traditional machine learning methods (such as linear regression and decision trees), which, when put into action, had a hard time generalizing knowledge.
While there’s a great deal of data at hand to support these methods, cryptos behave too unusually for that data to be compiled in a way that is necessarily meaningful — at least at this stage.
The analysis found that deep learning models had the most potential, as they achieved “decent levels of performance” regarding predictions. This idea, though, is still being explored. Perhaps the most prominent public example of the concept in action was written up by Towards Data Science.
The analysis there discussed in great detail the use of LSTM neural networks to facilitate a model predicting bitcoin prices in real time — a complicated and imperfect idea, but certainly an intriguing one.
Ultimately, right now, we can’t say definitively that machine learning can reliably predict cryptocurrency prices. Though as some of these examples indicate, different methods are being explored, and some have had degrees of success.
It may not be long before reasonably predictive analysis is more accessible, and traders can take advantage of better data.
Written by Barbara Rhodes & Edited by Michael Ding & Alexander Fleiss