Figure 7.

Median absolute and relative estimation error in Φ for each model implementation, over one thousand shorter (10-day) bivariate time series, with no minimum interval length. In this graph, each stacked bar represents a model implementation and each color represents a different element of Φ. In terms of the total amount of absolute estimation error, MI 2 and MI 4 are hard to distinguish and only MI 1 is clearly more biased than the rest; but when we consider the relative estimation error, it is clear (from the length of the stacked bars) that MI 3 and MI 4 outperform the other two MIs. MI 2 stands out because it is the only MI that overestimates the autocorrelations.