Predictive model. The first thing that we should note is that these are fat arrays — that is, they have too many variables and very few observations. This is by itself a challenge for most machine learning techniques unless an evolutionary algorithm is used.
When we start learning ranking models, we interpret returns as an indication of value, and the demand signals from traffic sources as the drivers of that value. Since we are constructing a ranking system, we are interested in the demand relationships that optimize BTC_returns > BTC_returns.