Time-Series In Crypto Machine Learning

Our team member’s contribution to understanding the application of time-series and contextual information in the market analysis:

Price changes in time of cryptocurrency trading can be seen as a time-series. Artificial intelligence and especially machine learning methods have proven to be of essential asset of finding repeating patterns and predicting the future course and the trend of the time series.

Machine learning methods usually use numerous hand-crafted features, some of them representing the expert knowledge and some of them are extracted using some statistics on the time-series itself (mean, standard deviation, mean crossing rate). This way, the machine learning model is able to analyse thousands and sometimes millions of feature vectors in order to learn some patterns and better predict the future trends.

In our approach, we additionally include contextual information about the current value, which is not present in the time series data itself, but it is known and available for human experts. This contextual information can be anything that has the ability to cause the major price movements: lock up periods (tracking ICO crowd sale addresses), movements in the known major institutional addresses (Satoshi, Kraken etc.), or approaching a psychological barriers with the value, the trend of other cryptocurrencies, amount/volume of trades, etc.

In this way, our approach will automatically 24/7 try to simulate the behaviour and thinking of an expert trading person with access to crucial information which in financial world are private (insider), but in cryptoworld part of a public ledger. Machine then includes this additional contextual information when making the analysis and the prediction of the cryptocurrency value.