Skip to content
PlaidCloud

Recursive Allocations

Recursive allocations handle the case where a target of one allocation becomes a source of the next. Common when modeling shared services that consume each other — IT serves HR, but HR also serves IT.

When You Need Recursion

Section titled “When You Need Recursion”
  • Reciprocal services — two cost pools that consume each other.
  • Layered spreads — divisional costs cascade through a hierarchy and the lower levels need to absorb the upper levels before re-allocating.
  • Iterative balance — the model needs to converge after multiple passes (cost pool A allocates some to B, B allocates some back, repeat until stable).

How Recursion Works in PlaidCloud

Section titled “How Recursion Works in PlaidCloud”

Configure the allocation step with:

  • Source table — where the values start
  • Driver table — the basis for spreading
  • Recursion mode — direct (one pass), reciprocal (resolve mutual dependencies), or iterative (loop until convergence)
  • Convergence tolerance — for iterative mode, how close residuals must be to zero before the loop stops
  • Maximum iterations — safety cap so a non-converging model doesn’t loop forever

The output table includes a generation or iteration column so downstream consumers can see which pass each row came from.

Reciprocal vs Iterative

Section titled “Reciprocal vs Iterative”
  • Reciprocal — solves a simultaneous equation in one mathematical pass. Use when relationships are well-defined and finite.
  • Iterative — runs allocations repeatedly until each round produces a residual smaller than your tolerance. Use when you want explicit control over how many passes happen, or when the relationship isn’t easily inverted.

Tips

Section titled “Tips”
  • Start with direct (single-pass) allocations and confirm the simple model behaves as expected before introducing recursion.
  • For iterative models, log the residual at each pass while tuning. If residuals don’t shrink, the model has a circular dependency the engine can’t resolve cleanly.
  • Recursive allocations can be expensive on large datasets. Test on a slice before running across the full source.
Section titled “Related”