Calibrate

Calibrate runs a correlation test — Pearson r or Spearman ρ, chosen automatically per pair by a distribution test — across every connected pair of nodes in the tree, at lags of 0 to 6 months. A Detrending choice sits above the results: Raw, Linear, MoM % or log Δ, which changes what the correlation is measured on rather than the raw series.

Every relationship lands in exactly one bucket. Actionable — statistically supported and worth keeping. Tautology — a relationship that's true by definition and tells you nothing new, such as a subtotal correlating with its own parent. Weak — tested, and not strong enough to act on. Then three ways a pair can be blocked before it's even scored: no target data, no source data, or fewer than 8 months of overlap between the two series.

The threshold for Actionable is stated outright: a 95% confidence interval (Fisher-z) of at least 0.30, and an effect size that clears |r| ≥ 0.20. Both conditions have to hold — a relationship can be statistically significant and still too weak in effect size to count, or vice versa.

Results carry a high / medium / low confidence tag, filterable by chip, and a table of Connection, Weight, Lag, Effect, N and Conf. per pair. Each row also carries its own lag profile and diagnostics, with warning chips for conditions like a small sample size or a sign flip between detrending modes — worth reading before ticking a row, since a technically-Actionable pair built on very little data or an unstable sign is a weaker basis for a decision than the label alone suggests. You tick the rows you want and Apply them — Calibrate proposes, it doesn't rewrite the tree on its own.

A tree with plenty of nodes but not much history will typically come back with few or no Actionable relationships and most pairs blocked for missing data — and that's a genuinely informative result, not a failure of the feature. It says the tree has a lot of structure but very little of it is yet backed by enough real history to trust. The exact counts on any given tree will drift over time as more history accumulates; treat the pattern (mostly blocked, few Actionable) as the thing to recognise, not any specific number.