Analysis

The first thing that we need to understand is the meaning of the models obtained. A trivial observation would be of the kind that one could find in the financial press, for instance, that because an increasing share of Google searches brings people to Bitcoin-related sites, then the market might be validating the positive sentiment expressed by higher prices. Rather, what we would like to understand from the shape of the data is what are those factors which variability has a noticeable impact on actual investor expectations changes, as measured by price returns--even if those sources are not among the largest traffic contributors to the crypto economy. This is because prices act similarly as a confounding factor (prices are tracked by both actual investors and enthusiasts, they may increase because there is more demand of informational resources and actual transaction activity, but because there is more transactional activity there might be more demand of informational resources as well).  Instead, price returns are more likely to be used by professional investors as a success metric.
We would also prefer to focus on the models in the knee of the Pareto front since those represent the better trade-off between complexity and accuracy.
For Bitcoin, the model with complexity equal to 22 becomes informative. It contains a metavariable (laser.online * vKontakte) that appears in 6.6% of the models, and one of the variables from that specific metavariable construct (laser.online) appears in some form in 4 of the 6 finalist models.  This is notable because while the other two variables in the model are a proxy for demand (the largest social network and search engine in Russia), usage of laser.online actually has investment implications -- that service was a famous bitcoin scam and Ponzi scheme, where BTC holders actually invested and lost funds \cite{online}.  The p-value for the metavariables considered in the analysis is under 0.03, as shown in Figure 9 and Figure 10.