
Forecasting collections without a promise to pay
The forecast says the money arrives on the due date. Your own ledger says this customer has never paid on a due date in four years.
- Forecasts built on invoice due dates assume behaviour that the ledger contradicts for most customers.
- Behavioural forecasting uses each customer's own payment pattern, which is available and rarely used.
- Promises to pay are the highest quality input and they have to be recorded as structured data to be usable.
- Forecast accuracy should be measured by period and by customer segment, otherwise errors cannot be corrected.
- Disputes and unapplied cash distort the forecast at the same time as they distort the ageing.
Short term cash forecasting in most companies works from the receivables ageing and the due dates on it. Everything due next week is assumed to arrive next week.
Every collections professional knows that is wrong, and knows precisely which customers make it wrong, because they deal with them every month.
That knowledge is rarely in the forecast, because it lives in the heads of the collections team rather than in a model.
Putting it into the model is not difficult. Each customer's payment history is a better predictor of their next payment than the due date printed on the invoice.
What the ledger already tells you
For every customer you have a history of invoice dates, due dates and payment dates, which yields an average and a distribution of days beyond terms.
That distribution is usually stable. Customers pay the way they have always paid, because it reflects their own internal payment run and approval process rather than an attitude to your invoice.
Applying each customer's own distribution to their open invoices produces a forecast that is substantially more accurate than one based on due dates.
It also produces a range rather than a point, which is more honest and considerably more useful to treasury. A range makes the facility decision explicit, since the question becomes how much headroom is needed to cover the spread rather than whether a single number is right. That framing is easier to defend to a lender and easier to review internally, which is the property that any figure you have to defend needs, and it depends on the underlying data being reconciled rather than assembled, as routine reconciliation makes possible, in the way quantifying the value of a day makes the analysis concrete.

Four inputs in order of quality
A good forecast uses the best available input per invoice rather than one method for everything.
Promises to pay deserve a field
A promise to pay is a customer stating a date. It is the single most predictive piece of information in collections and it is almost always recorded as free text in a call note.
Stored that way it cannot be aggregated into a forecast, measured for reliability or reported on, so its value is limited to the individual who took the call.
Stored as structured data with a date and an amount, it becomes both a forecast input and a measure: the proportion of promises kept, by customer.
That measure is also a credit signal, because a customer whose promises stop being reliable is telling you something before their ageing does, which connects to behavioural credit review.
The most predictive data in collections is stored as a note. Give it a field.
Measuring forecast accuracy
Three cuts, because a single accuracy percentage tells you nothing about where to improve.
| Cut | What it shows | Typical finding |
|---|---|---|
| By week | Whether the timing is right even where the total is. | Cash forecast correctly for the month and wrongly within it, which matters for facility usage. |
| By customer segment | Which segments the model handles badly. | Large customers with payment runs are predictable; small ones are not, and are treated identically. |
| By input type | Whether promises, behaviour or due dates are driving the error. | Promises are usually the most accurate, which justifies capturing them properly. |
Forecasting approaches should be consistent with your treasury policy and any covenant reporting requirements.
The due date is a contractual fact. It is not a forecast.
ChatFinBuilding a behavioural forecast
Five steps, each using data you already hold, producing something materially better than a due date forecast.
Calculate the distribution per customer
Days beyond terms over a rolling period, as a distribution rather than an average, so the forecast can carry a range.
Give promises a structured field
Date and amount, captured at the point of the call rather than in a note.
Exclude disputes and unapplied cash
Both distort the open balance, which is the same prerequisite described in separating disputes.
Apply the best available input per invoice
Promise first, then behaviour, then segment, then due date.
Measure accuracy weekly and correct
By segment and input type, so the model improves rather than being rebuilt annually.

Two reasons forecasts stay bad
Both are organisational rather than technical.
Accuracy is never measured
A forecast that is never compared to actuals cannot improve, and in many companies last month's forecast is not retained in a form that allows the comparison.
Collections knowledge stays verbal
The people who know how each customer pays are not asked to encode it, so the model competes with their knowledge rather than using it, which is the pattern in preparing the analysis systematically.
How ChatFin supports collections forecasting
The behavioural analysis is repetitive calculation over ledger data, which is straightforward to prepare weekly rather than occasionally.
Treasury decisions, facility management and any commitment made to a lender remain matters for your team.
AI Cost Estimator for the CFO Office
Four inputs. A first year range, a timeline, and three delivery models compared. No login required.
Frequently Asked Questions
How much more accurate is behavioural forecasting?
It varies by customer mix, and the improvement is largest where customers are consistent but not on terms, which describes most business to business books. The bigger gain is usually in weekly timing rather than in the monthly total, and weekly timing is what drives facility usage.
What is a promise to pay worth?
It is the most predictive input available, and only if it is captured as structured data with a date and an amount. Recorded as a call note it helps the individual who took the call and contributes nothing to the forecast or to any measure of reliability.
Should disputes be in the forecast?
Not at their invoice due dates, since they will not pay until resolved. They can be forecast separately using historical resolution times, which is more honest than either assuming they pay on terms or excluding them entirely.
How often should the forecast be rebuilt?
Weekly, with accuracy measured against the prior week. The distributions themselves change slowly, so the recalculation is cheap. What matters is that the comparison happens and that errors by segment are visible and corrected.
The due date is the weakest input in your forecast.
Calculating payment distributions per customer, capturing promises as data and measuring accuracy weekly produces a forecast treasury can act on.
If you want a first year range for behavioural collections forecasting, the estimator will produce one before any scoping.
Open the estimatorBook a Demo
See what this looks like in your ERP.
We will walk through the workflow behind this article on the system you already run, using your own data, and show what a first phase would actually cover.
This article is general information for finance and accounting teams. It is not legal, tax, accounting or investment advice. Dates and thresholds are given as published at the time of writing and several of these timelines have moved before, so confirm the current position with your advisers and with the relevant authority before acting.