Analyzing prices of cryptocurrencies is a challenging task. Cryptocurrencies are not correlated directly with the movement of the traditional financial market instruments. Cryptocurrencies cannot be valued based on companies' fundamentals. Hence, machine learning can help to detect trends and movements of the market of cryptocurrencies. Data enrichment can also be attempted for the sake of quick evaluation of which economical forces are connected to the movement of the cryptocurrencies' price level. This analysis is an example of the application of machine learning to unconventional financial instruments and demonstrates how AI and advanced computing can be applied in similar cases.
Results in this article were obtained using the following machine learning python models:
• Linear Regression, Support vector machines, Stochastic Gradient Descent, Gradient Boosting and Neural Networks.
Cryptocurrencies included in the analysis are the following:
The dataset has 28,944records from April2013 to December 2019.
Currencies and respective daily trading volume (original data):
Bitcoin is an outlier in the data. As a consequence, in this article an emphasis is placed on the cryptocurrencies' market as a whole.
Currencies and respective daily trading volumes after outliers are treated:
After cleaning the data from outliers, the number of observations becomes 25,324. For some algorithms and visualizations, this means that a random sample must be drawn to complete all calculations because this is a considerably large dataset.
After pre-processing, the data is 'zoomed' to all currencies except for Bitcoin and Bitcoin Cash which have extreme price values and are treated as outliers. The focus of this analysis is shifted intentionally towards the rest 10 currencies and Bitcoin and Bitcoin Cash are recommended for individual analysis (outside of the scope of this work).
Modeling of the data and error estimates are obtained using Linear regression, Support Vector Machines, Stochastic Gradient Descent, Gradient Boosting and Neural Networks.
Summaries of results, including regression plots and models' performance measures:
Contrary to expectations, the more complex models such as neural networks and SVMs do not perform well.
In conclusion, the following observations can be made:
Each dataset has its specifics and not all algorithms/models perform well on a target dataset even if they are superior in complexity.
The cryptocurrencies high daily price can be modeled using supervised learning algorithms such as linear regression and gradient descent and boosting.
A separate analysis is required to find the best algorithm that can model Bitcoin pricing.
Data enrichment, such as the addition of S&P500 data or other market data, can be applied to this analysis to enhance results.
Comentarios