Artificial intelligence (AI) has long been used to predict various events. Algorithms based on historical data have successfully predicted Oscar winners, stock prices, and even the spread of COVID-19 pandemics.
In the last five years, there have been many attempts to predict the price of Bitcoin. Startups, research groups and enthusiasts are trying to create an algorithm that can predict the price of the biggest cryptocurrency in the short and long term.
In this article, we will explore how Bitcoin price prediction AI works and what are the main challenges it faces.
Developers of machine learning algorithms use different approaches in creating predictive tools. The most popular ones are the recurrent neural network and the long short-term memory model (LSTM).
Recurrent neural network: A type of neural network where connections between elements form a directed sequence. This allows the algorithm to process a series of events over time.
The LSTM model is a type of recurrent neural network that can remember long-term sequences of data. Just as humans use previous experiences to predict future events, a neural network is able to remember information over long periods and quickly find behavioral patterns.
The first attempts to predict the price of Bitcoin were made during the 2017 crypto boom.
Serbian entrepreneur and developer Ognjen Gatalo created a Twitter bot that predicted the Bitcoin price using historical data. Every two hours, the bot published the predictions for the next few days. Ognjen used the blockchain.info API as a source of information. His algorithm collected price data for the last two months and tried to guess future market fluctuations by searching for nearest neighbors.
The developer’s attempt was unsuccessful. Why? The main reason is the lack of information outside of the market. The algorithm took into account only historical price data, while many other factors that influence the value were left aside.
Another attempt was made two years later, in 2019, by an American developer of machine learning algorithms Marco Santos. Marco used the LSTM model for his forecasting tool and chose Yahoo Finance as the source of data. According to the original concept, the algorithm analyzed the Bitcoin price fluctuations for the last 30 days and made the predictions for the 10 upcoming days.
As in the previous example, the algorithm relied only on historical data. The main difference is that Marco used the LSTM model to train the AI. The recurrent neural network model allowed the developer to achieve more accurate predictions, which were still far away from being perfect.
Marco explains that his algorithm may have errors because nobody — neither humans nor algorithms — can reliably predict the future. The neural network can only display trends formed by past experience.
Another noteworthy example is the algorithm of the American entrepreneur Frederic Riveroll, who has also developed an artificial intelligence solution for Bitcoin price prediction. The main distinctive feature of Frederic’s neural network is that it analyzed not only the historical price data but also the news headlines.
Frederic used two datasets for his algorithm – Bitcoin/USD price charts and Fox Business News. The data from both sources were parsed and matched by date. As a result, the neural network revealed a number of patterns between keywords in news headlines and the price of the main cryptocurrency.
For example, when the media discussed President Trump’s impeachment, the price of Bitcoin went up. But when the media wrote about Netflix the price went down. Based on those correlations the AI made further predictions.
As a result, Frederic’s algorithm performed with an accuracy of 64.7%. It is better than the previously mentioned AI’s, but still not enough for confident prediction. Nevertheless, the experiment proved the fact that the more factors are taken into account the more accurate predictions neural networks can give.
Unlike the stock market, which has almost 100 years of data to study, the cryptocurrency market has only 10 years of history.
In addition, financial data alone is not enough. As proved by the examples above, other factors must be taken into account to create algorithms that can perform with acceptable accuracy.
Nowadays all the best engineering is still created by humans. This means that just like people, trading tools are subject to the biases and limitations of the average person. If we don’t fully understand how the market works, how can we create an AI-based algorithm that exceeds our capabilities?
Sophisticated machine learning algorithms require enormous computing power. For example, DeepMind required 1202 CPUs and 176 GPUs to run the AlphaGo algorithm.
Large companies can easily overcome these barriers, while such computing power is still out of reach for startups or independent researchers.
Although AI has proven effective in many industries, it is still unreliable for the high-risk investment market.
Traders who will implement neural networks as a forecasting tool need to work hard to prove to the public that it is highly error-free for long-term predictions.
AI-based tools offer new opportunities for development and growth.
Predictive algorithms can recognize patterns in new goods and phenomenons. Cryptocurrencies are not the only new and volatile investment asset. Learning the patterns of volatility and learning from them can be a shortcut to faster success.
Sooner or later the uncertainty among cryptocurrency traders and investors will decrease, leading to a more sustainable market. Stabilization and predictability of digital asset behavior will create more accurate analytics and forecasting tools in the future.
So far, all attempts to create a reliable tool based on artificial intelligence have failed.
Yes, in practice the AI developers have managed to create relatively accurate algorithms, but they are still unsuitable for long-term forecasting. High volatility, unpredictability, and lack of data complicate the task.
However, everything may change very soon. As practice shows, recurrent neural networks have proven themselves in forecasting tools in other industries. The more researchers put effort into collecting data and better understanding the market, the sooner the day will come when there will be sustainable AI tools for cryptocurrency investors.