← Back

Digital agency · cash forecasting

A cash forecast built on how clients pay.

Thirteen weeks of operating cash, rebuilt from the ledger and payroll every time I press refresh. Client ad spend is kept out of it entirely, because that money is only passing through and was never the agency's to spend.

The part that makes it work isn't the model. It's that collections are driven by a pay curve measured from 748 real invoices, rather than by the terms printed on them.

Every forecast I'd used before was built on net 30. Net 30 is a sentence in a contract, not a description of behaviour. The question a forecast has to answer is not when a client has agreed to pay — it's when clients like this one, with an invoice like this one, have actually paid before.

0.6% off, across six billing runs

Back-tested against what actually came in. The number is honest, and the next section is about why it flatters the model.

Horizon 13 weeks of operating cash
Curve 748 invoices measured

Why terms don't forecast

A forecast built on payment terms produces a tidy picture in which every invoice lands thirty days after it goes out. Nothing lands thirty days after it goes out. Some clients pay in a week, some pay the day their own client pays them, and some pay when chased — which means the real driver is a distribution, not a date.

So the curve is built from history: for every invoice, how long it actually took to turn into money. Run that across 748 of them and you get the shape of how this client base behaves — a fat early bulge, a long thin tail, and a centre of mass that sits nowhere near day thirty.

A forecast that assumes terms isn't wrong by a little. It is wrong in a specific, repeatable direction: too optimistic early, and silent about the tail that actually shapes the cash position.

What the back-test actually says

the caveat is the finding

Across six billing runs the forecast landed within 0.6% of what came in. That is the headline, and on its own it would be misleading, so here is the rest of it.

Individual billing runs ranged from roughly a fifth of the prediction to double it. The curve does not forecast a single run. It forecasts the aggregate, and the errors cancel because they are errors of timing rather than of amount — the money arrives, just not in the week the curve expected.

Which makes the honest claim narrower than the number suggests: this tells me what the quarter looks like. It does not tell me what next Tuesday looks like, and I don't use it that way.

That distinction matters for the decision the forecast actually supports. For "can we commit to this hire" it is the right instrument. For "do we move money before Friday" it is not, and the answer there still comes from the bank balance and the payment run.

The past-due fix

The first version applied the same curve to past-due balances, which produced something obviously false: more than half the entire past-due balance scheduled into a single week. The thin tail of the curve, applied to invoices that had already outlived it, stacked them all onto one date.

Past-due balances now use measured weekly pay rates instead — of the invoices already this late, what share of them historically turns into cash in a given week. It spreads the balance the way it actually arrives, and it stopped the forecast from showing a cliff that was a modelling artefact rather than a real event.

An invoice that is ninety days late is not a normal invoice running late. It is a different population, and it needs its own rate.

How it runs

It lives as a page rather than a workbook. Pressing refresh rebuilds it from the ledger and from payroll, so it is never a stale file someone forwards with last month's numbers in it. Ad spend is excluded on both sides of the ledger — the client money coming in and the media bills going out — so what's left is the operating cash the agency actually controls.

It's one of eight live pages the skills and agents report to, and the only one that is a model rather than a worklist. The agents that feed the others →

Built with Claude Code. One of several things built for the agency.