The Three-Statement Model That Cannot Go Out of Balance

Change one assumption in a spreadsheet and you’re trusting yourself to have wired every downstream cell correctly. Miss one reference and the balance sheet quietly stops balancing, and nobody notices until an auditor, or a board member doing the arithmetic by hand, does. Rexfin’s modeling engine is built so that failure mode can’t happen: it recomputes all three statements from a set of typed drivers, and it enforces the accounting identities as a rule the model cannot violate, not as a checklist someone runs afterward.

What “deterministic” actually buys you

The engine is not a smarter spreadsheet. It’s a calculation layer underneath the model that does one job: given a set of driver values (growth rate, gross margin, days sales outstanding, capex as a percent of revenue, whatever the model needs) it recomputes every dependent line and hands back a statement set that ties. Change the growth driver and the engine walks the dependency graph, revenue then cost lines then working capital then cash, touching only the cells that actually depend on what changed. That’s what keeps a full recompute fast enough to feel like a live edit instead of a batch job someone waits on.

The identities aren’t a display convention. Assets equal liabilities plus equity, and the cash line ties to the cash-flow statement, because the engine won’t produce an output that fails those checks. If your driver logic is wrong, you’ll get a wrong number (the engine doesn’t fix bad assumptions), but you will never get a silently wrong number where the sheet still looks fine.

Real filings don’t sum as cleanly as you’d like

The hard part shows up the moment you point this at an actual annual report instead of a clean textbook model. Extracted line items don’t always add up to the subtotal a company printed. A chunk usually goes missing (right-of-use assets, deferred tax, a reserve buried in a footnote) because the filing rounds, nets, or discloses it differently than the granular lines suggest.

Rexfin doesn’t drop that remainder and quietly force the totals to match. The gap between what the extracted lines sum to and what the company printed becomes its own line: an explicit bridge, labeled “Other (held constant)” and projected flat, rather than smeared invisibly across other accounts. You can see it, question it, and override it. Cash is handled the same way: derived from the cash-flow logic rather than plugged after the fact to make the balance sheet close.

Scenarios without rebuilding the model

Because the engine works off drivers, a named scenario is just a sparse set of driver changes layered on top of the base case, not a separate model maintained in parallel. That’s what makes driver-based what-if analysis fast: flexing a churn assumption or a rate change re-runs the same deterministic logic against the same reconciled base, instead of forcing someone to rebuild a parallel tab by hand. Sensitivity ranges and driver-by-driver comparisons work the same way, because they’re all reruns of one engine rather than separate calculations someone has to keep in sync.

Historical periods are seeded as actuals and stay actuals: the engine projects forward from them but does not silently overwrite what already happened, which matters the first time someone asks why last quarter’s number changed after a routine refresh.

Who this is for

This is the layer underneath everything else in the model. It’s what three-statement projections back the board pack, what the reconciliation against printed totals checks against, and what gives the AI something deterministic to narrate instead of estimate. If your model only ever faces an internal spreadsheet review, you may not need this level of rigor. If it has to survive a board question, an auditor, or a lender’s diligence team, this is the difference between a projection you can defend line by line and one you hope nobody checks.

See how the layers fit together on the Rexfin product tour.