Collections forecasting: how to predict what you will collect

Collections forecasting predicts how much cash you will collect from open receivables over a coming period, and when each payment will land. It takes the open AR ledger and turns it into a week-by-week expectation of incoming cash you can plan around.

Done well, it answers the question treasury actually asks: how much will we have to spend in week three, and how sure are we? A good collections forecast is specific by week and honest about its own confidence. A bad one is a single number that turns out wrong.

What collections forecasting is, and is not

Collections forecasting is narrower than a full cash flow forecast. It covers one inflow, the cash you expect to collect from invoices already issued. It does not cover new sales you have not booked, financing, or payables. That focus is its strength: it is the most controllable and most forecastable slice of cash, because the invoices already exist and you know who owes what.

It is also distinct from a revenue forecast. Revenue is what you billed. Collections is what you will actually receive, and when. The gap between the two is timing and risk, and forecasting collections is the discipline of estimating that gap accurately.

Methods and models

There are three common approaches, in rising order of accuracy.

Terms-based. Assume every invoice pays on its due date. Simple, and consistently optimistic, because customers pay late. Useful only as a rough ceiling.

Roll-rate. Use historical aging-bucket behavior to estimate what share of each bucket converts to cash each period. If 70 percent of your 1-30 bucket historically collects within the next month, you forecast 70 percent of today's 1-30 balance as next month's cash. Better than terms-based, but it treats every customer in a bucket the same.

Behavioral, invoice-level. Forecast each invoice at the date its specific customer tends to pay, then sum. This is the most accurate, because it captures that one customer reliably pays 5 days late and another routinely runs 25 days over. It needs more data, but it is the method that actually holds up week to week.

Most teams start with roll-rate because it works off an aging report. The accuracy ceiling is the invoice-level method, which is also what makes a continuously updated forecast possible.

A quick worked example shows the gap. Say you have 500,000 in the 1-30 bucket. A roll-rate model that has seen 70 percent of that bucket collect within the month forecasts 350,000 next month, full stop. The invoice-level method looks inside that 500,000: a 200,000 invoice from a customer who promised next Tuesday, a 150,000 invoice from a reliable payer due in two weeks, and 150,000 spread across erratic accounts. It forecasts the first two with high confidence and specific dates, and discounts the third. Same starting balance, but one method gives you a number and the other gives you a number, a date, and a confidence level.

Data you need to forecast well

A forecast is built from inputs, and weak inputs cap accuracy no matter how good the model is. The essentials:

• Open invoices with due dates. Each invoice, amount, and contractual due date. Not a single AR total.
• Per-customer payment history. How many days each customer typically pays past due, and how consistent they are.
• Promises to pay. Specific commitments a customer has made, by date.
• Open disputes and deductions. Invoices that will not pay clean or on time until resolved.
• Payment terms by customer. So the baseline due dates are right before behavioral adjustment.

Pull these together and reading the open ledger as a future cash timeline becomes straightforward. The same source data feeds your AR aging analysis, which is the static cousin of the forecast: aging tells you where receivables sit today, the forecast tells you when they will turn into cash.

Promise to pay as a forecast signal

The single most valuable input is the promise to pay. When a customer commits to paying a specific invoice by a specific date, that beats any statistical estimate, because it is stated intent for this exact balance.

A forecast that captures promises is dramatically sharper than one built on terms or averages alone. A 50,000 invoice that is 40 days late looks like a low-probability collection in a roll-rate model. If the customer promised Friday, it is high-probability cash this week. The model only knows that if the promise was captured and fed in.

The catch is that promises decay and break. A promise made three weeks ago and already missed is not a forecast signal, it is a flag. A useful forecast weights promises by how fresh they are and how reliably that customer has honored past commitments. A customer who keeps 90 percent of promises gives you near-certain cash. One who keeps half gives you a coin flip you should model as such. Tracking promise to pay conversion per customer turns a soft signal into a calibrated one.

Improving accuracy over time

A collections forecast should get more accurate the more you run it, because every period gives you actuals to check against. Compare what you forecast to collect against what landed, find where the misses came from, and adjust.

The common failure points are predictable. Promises that broke, so a planned payment never arrived. Disputes that surfaced after the forecast was built. A reliable customer who slipped for the first time. Each miss is a lesson the model should absorb, so a customer who starts paying later gets forecast later next time. Measuring forecast variance every period is how you close that loop, which is the heart of cash flow forecasting accuracy. Without the comparison, the forecast never learns and quietly drifts.

The other accuracy killer is stale data. A forecast built on a two-week-old AR export is forecasting a ledger that no longer exists. Invoices have aged, cash has landed, promises have been made. The fresher the underlying data, the tighter the near-term weeks, which are the ones treasury cares about most.

Segment your accuracy review, too. A forecast that is reliable on the long tail of small invoices but keeps missing on the top accounts has a concentration problem, and the fix is to model those large customers individually rather than pull them into an average. Reviewing variance by customer segment, not just in aggregate, tells you exactly where the error lives instead of leaving you to adjust a single blended number and hope.

Forecasting with an AI AR agent

Behavioral, invoice-level forecasting with live promise data is the accurate method, and it is also the hardest to sustain by hand. Keeping per-customer payment behavior, fresh promises, and current disputes synced into a model every week is more than most teams can do on top of actually collecting.

An agentic AR system removes that constraint, because it is already doing the work the forecast depends on. Rex tracks every invoice, captures and follows up on every promise to pay, and flags every dispute as part of running collections across the whole ledger. The same signals that decide its outreach feed the forecast directly, so the forecast reflects what each customer is actually likely to do, not what the terms imply. Because Rex acts on accounts continuously, the forecast updates as promises are made and cash applies, and it sharpens as the agent's read on each account improves. Finance sets the assumptions and reviews the variance; the agent keeps the underlying behavior current and escalates the accounts where a human call changes the number.

The result is a forecast by week, with a confidence you can defend, that holds up because it is built from how customers really pay.

See how Rex turns live promise-to-pay data into a collections forecast you can plan against.