Bitcoin Returns Prediction via Machine Learning
Updated: Aug 5, 2020
This new post is a quick summary of the most recent research that the quant and asset management team of 4E Capital has been working in the past weeks. We will briefly present our machine learning model for Bitcoin returns prediction and go through some key elements within the model.
Since the inception of Bitcoin in 2009, this cryptocurrency has experienced very volatile movements with some famous rallies. One of the most famous bullish runs ended some days before Christmas 2017 with its price reaching almost 20 thousand dollars and subsequently crashing. Predicting the returns of any asset is one of the most challenging tasks that any asset management team could face. At 4E Capital, we designed a model that measures different types of signals and assess the current state of the crypto market in order to achieve an adequate prediction for Bitcoin in the future.
First, we draw some general observations and found out that overall Bitcoin tends to behave bullishly whenever the VIX index is constant. On the contrary, Bitcoin has experienced sudden crashes at times of increasing VIX. The next image depicts this observation of the negative correlation between Bitcoin's price and the VIX index. The windows highlighted in green show the effect of rising Bitcoin price and constant VIX index. The red windows highlight the decrease in the Bitcoin price mixed with an increase in the VIX index. A possible explanation about this effect is that risky investors look for alternative investments during times of low volatility expectation in the S&P 500 market. On the other hand, those risky investors might go away from a risky Bitcoin investment if the stock market volatility is expected to rise.
With the last observation serving as a starting motivation to understand better the cryptocurrency market, we included additional signals in our model to have a better feeling of how Bitcoin's price behaves. Subsequently, we included a list of technical indicators to see if there is a pattern to recognize among those indicators and the Bitcoin returns. These technical indicators can be grouped as:
With this list of technical indicators, we were able to design a machine learning model in order to predict the returns one week, two weeks, and one month ahead. This was a challenging task since we tried different types of models starting from simple linear regressions, elastic nets, decision trees, all the way to neural networks. We will leave as a mystery which model is our final one as a motivation for the reader to try out these novel techniques and discover which model is the best for you. This is a fun task that requires choosing a model, setting the labels, picking the correct training and testing size, calibrating the model, and discovering the most important variables driving the Bitcoin market. As a hint, we do not believe that the Bitcoin market is explained purely by technical indicators. Rather, you can have a broader view of the crypto market with the inclusion of macroeconomics variables, blockchain-related data, FOREX, sentiment information with searching engines, and more data.
An interesting example of Bitcoin sentiment is the following image that illustrates the relative number of Google searches of the word "Bitcoin halving" together with its price. It is clear from the image that during April and the beginning of May of 2020 the number of people anticipating a big movement increased exponentially until the last Bitcoin halving took place on the 11th of May 2020.
After implementing, calibrating our model, and identifying the key elements explaining the Bitcoin market, we achieved an excellent in-sample mean return of Bitcoin and solid out-of-sample testing results. The next figure compares the historical Bitcoin's returns (blue) and our in-sample results (green). These results serve as a training set for our machine learning model. This example shows a histogram of the bi-weekly returns.
Next, the out-of-sample results represent the back-testing of our machine learning model. The curve in blue shows the historical returns of Bitcoin and on green our past predictions. The graph describes the return in the percentage of Bitcoin 14 days ahead. For example, the point on the graph on the15th of February represents the returns 14 days after that day (29th of February) which speaks for a negative return if we compare the closing price of Bitcoin on the 15th ($9,889.42) and the closing price of Bitcoin on the 29th of February ($8,599.51). It is a major achievement that our model was able to predict the negative returns during the beginning of the coronavirus selloff in March and the following trends in April and May. Nevertheless, there is room for improvement in the magnitude of the returns. This task will be the scope in our new research.
To conclude, this post is brief summary of our newest model to predict Bitcoin's returns. We are currently working on other machine learning prediction models for other cryptocurrencies and different asset classes. Stay tuned for the next update which will show a novel idea to construct a cryptocurrency-blockchain portfolio based on machine learning predictions and quantitative methods. As a sneak preview, the following figure shows different cryptocurrencies with their real returns and our passed predictions.
We want to stress out the fact that cryptocurrencies are a very risky investment and this post should not be seen as investment advice. If you are interested to learn more about machine learning models and cryptocurrencies we recommend the following links to have a general idea of what machine learning and Bitcoin are and their potential.
Deep learning video explanation: https://www.youtube.com/watch?v=aircAruvnKk
Bitcoin technology: https://www.youtube.com/watch?v=bBC-nXj3Ng4
Overview of machine learning techniques and cases: https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/an-executives-guide-to-ai
If you are curious to know more about cryptocurrencies, blockchain technology, machine learning techniques, and quantitative finance you can contact:
Daniel Partida - Quant & Associate
For questions about traditional investments in asset management such as bonds and stocks, you can contact our experts.
Javier Lamelas - Chief Investment Officer & Partner
Phlipp Leibundgut - Chief Eexecutive Officer & Partner
Or you can visit our webpage www.4e-capital.ch you meet the team and discover our offers.