Skip to content

Bindings & the graph

A binding is a labelled group of cells that represents one variable in the model — a whole column ("Policies In Force", one value per projection month), a single input cell ("Discount Rate"), or a block of assumptions. Extraction groups the raw cell-level graph into bindings so that auditing a model means reasoning about its variables, not walking thousands of individual cell references one at a time.

A binding is always a contiguous rectangular range on a single sheet, and it is one of two kinds:

  • formula — the cells share a repeated formula pattern (e.g. the same formula, position-adjusted, down every row of a column).
  • constant — the cells hold input values with no formula.

A tiny worked example

Take a two-column projection sheet, one row per month:

Cell Formula Meaning
B2 =B1-C1 Policies in force this month = last month's in-force − last month's lapses
B3 =B2-C2 (same pattern, next row)
⋮ ⋮ ⋮
C2 =B2*0.02 Lapses this month = 2% of this month's in-force
C3 =B3*0.02 (same pattern, next row)
⋮ ⋮ ⋮

Column B collapses into one binding — call it pols_if — and column C collapses into another, pols_lapse. Instead of a graph with one node per cell, you now have a graph with two nodes: a variable, not a spreadsheet coordinate.

How bindings relate: binding edges

Bindings are connected by edges derived from the underlying cell references: if any cell in one binding refers to any cell in another, an edge connects the two bindings (with a count of how many underlying cell references it represents). Extraction also recognizes VBA-driven dependencies the same way — a macro that paste-as-values a template row into an output block produces a binding edge too, tagged with the procedure responsible; see VBA extraction.

Binding graphs can be cyclic. In the example above, pols_if this month depends on pols_lapse last month, and pols_lapse this month depends on pols_if this month — cell by cell, that's a strict DAG (nothing depends on its own future), but collapsed to bindings, pols_if and pols_lapse depend on each other. This is normal for any recursive projection model, not a defect: extraction handles cyclic binding graphs correctly and always produces a stable ordering over them, so tooling built on top (documentation, lineage walks) doesn't need to special-case recursive models.

Why bindings are the useful audit unit

A model with 10,000 policy rows and 50 calculated columns has 500,000 formula cells but on the order of 50 model variables. Auditing at the cell level means drowning in near-duplicate edges; auditing at the binding level means answering the question an actuary or analyst actually asks — "what does Net Cashflow depend on, and what depends on it?" — with a handful of named variables instead of a wall of cell coordinates.

Next