ARNNResult.r_squared

property ARNNResult.r_squared: float

Share of the effective sample’s variance the fitted mean explains.

The headline number for this family, since there is no coefficient table and the log-likelihood is not a maximized one. Computed against the effective sample’s own total sum of squares, so it is comparable across specifications only when they trim the same number of leading observations – which is why compare() refuses mismatched samples.

Note this is an in-sample figure for a flexible learner, so it rewards capacity: a network with enough hidden units will drive it toward one without having learned anything that generalizes. Read it next to parameters_per_observation.