Manaz Afzal

Accounting & Finance · Cash-flow forecasting

A cash forecast built on when customers pay, not when invoices fall due

Tested on 98,631 public payment reports from 9,388 large companies, for suppliers whose cash depends on a few big accounts.

4.5 daysaverage error in days to pay; 30-day terms 13.3
2.3 ptserror in first-month cash for a 12-customer ledger; terms assumption 40
75%of heavy late payers stay late the next period

Executive summary

A large customer’s last published payment report forecasts when it will pay within 4.5 days on average, against 13.3 days for the due-date assumption most cash forecasts use.

4.5 daysaverage error in days to pay, against 13.3 for 30-day terms
2.3 ptserror in first-month cash for a 12-customer ledger, against 40
75%of heavy late payers stay late the next period

Situation

A supplier with a few large customers forecasts cash on due dates; large customers pay early, late or very late, and the forecast misses in the weeks that matter.

Approach

Each customer’s published payment reports used to forecast its next days to pay with an 80% range, tested on 24,692 later reports from 6,618 companies against the due-date assumption and a boosted model.

Result

The last report is within 4.5 days on average and its ±20% range held 81% of outcomes. The boosted model adds 0.06 days, so the simpler method goes live.

The problem

A supplier whose revenue comes from a few large customers forecasts cash by assuming each invoice is paid on its terms. Large customers often pay later than that. Australia’s Payment Times Reporting Scheme shows large businesses take about four weeks on average to pay small suppliers, and the slowest invoices take more than twice the agreed term. For a business with three or four big accounts, a forecast built on terms is wrong in the weeks that matter.

I built a forecast that places each open invoice in the week the customer is expected to pay, from that customer’s own published payment record, and measured how far it lands from the actual payment date against the due date.

Team
Treasury or finance manager, credit control team, financial planning and analysis
Customers
Large corporate and government customers carrying most of revenue
Systems
Xero or MYOB receivables, the Payment Times Reports Register, a cash-flow tool such as Float or a spreadsheet
Measure
Error in forecast days to pay; error in cash expected within 30 days

What I built

A payment-timing table for each large customer, refreshed when the customer files its twice-yearly payment report. It gives the expected days to pay with an 80% range and flags customers likely to pay many invoices late. The cash forecast places each open invoice at its expected payment date instead of its due date. Once the business has a few months of its own receipts from a customer, those take over.

Public reportCustomer’s latestpayment reportTiming tableExpected daysand 80% rangeCash forecastOpen invoices placedat expected dateOwn historyOwn receiptstake over
The same table feeds credit terms for new contracts and the weekly receivables review.

Results

For 24,692 payment reports from 6,618 large companies covering periods ending July 2024 to October 2026, never used to fit anything, the customer’s previous report forecasts its next average days to pay within 4.5 days on average. Assuming 30-day terms is out by 13.3 days, and the customer’s own stated standard terms by 17.1.

Assume 30-day termswhat most cash forecasts do13.3 daysCustomer’s stated standard termsfrom its own report17.1 daysCustomer’s last published reportaverage days to pay4.5 daysBoosted model on report historygradient boosting4.5 days
The boosted model on the full report history is 0.06 days better than the last report alone (95% range 0.00 to 0.12 days). That gain is too small to justify the added complexity.
Assume 30-day terms29%Customer’s last published report78%Boosted model on report history74%
The last report lands within five days of the actual figure for 78% of customers, against 29% for the 30-day assumption.
Forecast of days to payAverage errorWithin five days80% range held
Assume 30-day terms13.3 days29.1%–
Customer’s stated standard terms17.1 days––
Last published report, with ±20% range4.5 days78.0%81.2%
Boosted model, with conformal range4.5 days74.4%80.5%

Ranges that hold

A range of ±20% around the last report’s days to pay, set on 2023 reports, held the actual figure for 81.2% of later reports, close to the 80% it was set for. The median range is 13 days wide. The cash forecast shows a likely and a slow case from it.

What it means for a ledger

For 5,000 random ledgers of 12 large customers each, the median share of invoices actually paid within 30 days is 60%. A forecast that treats every invoice as paid on 30-day terms overstates cash in the first month by 40 points. Using each customer’s last report, the error is 2.3 points on average and under 5 points for nine ledgers in ten.

Assume every invoice paid on 30-day terms40.0 ptsUse each customer’s last published report2.3 pts

Late payers stay late

Late payment is persistent. A customer that paid more than 40% of its invoices late in one period did so again in the next 75% of the time. One that paid 20% or fewer late crossed that line 1.7% of the time. These customers get shorter credit limits, deposits or milestone billing written into the contract before work starts.

Last report saidNext report: more than 40% of invoices paid late
More than 40% paid late74.8%
20% or fewer paid late1.7%
All large companies17.3%
0%5%10%030609012030-day termsAverage days to pay (capped at 120)
Most large companies pay in 20 to 45 days; 22% average more than 45 days and 7% more than 60.
0%10%20%30%2019202120232025Invoices paid late
Late payment has fallen since reporting began in 2017. The median share of invoices paid after agreed terms was 25% in late 2018 and 15% in the first half of 2026.

Data and method

Data
UK Payment Practices and Performance reports, published by the Department for Business and Trade under the Reporting on Payment Practices and Performance Regulations 2017 (Open Government Licence): 98,631 usable reports from 9,388 large companies, each covering six months
Target
The company’s reported average days to pay in its next period, and whether more than 40% of invoices were paid after agreed terms
Split
Fitted on periods ending before January 2024; ranges set on periods ending in 2023; tested on periods ending July 2024 to October 2026
Methods
Flat 30-day terms; the company’s stated shortest standard payment period; its last reported days to pay; gradient boosting on the last three reports, trends and stated terms
Measures
Average absolute error in days; share within five days; coverage of 80% ranges; company-level bootstrap 95% range on the model’s gain
Ledger test
5,000 random sets of 12 test companies; share of invoices paid within 30 days, forecast against reported

Limitations

  • The reports are averages across all of a company’s suppliers. A supplier can be paid faster or slower than average, so the business’s own receipts replace the public figure once a few months of history exist.
  • Reports cover UK companies above the size threshold. In Australia the Payment Times Reporting Scheme publishes similar figures for large businesses, and the same method applies to that register.
  • Reported figures are self-assessed and some contain obvious errors; reports with more than 365 average days were excluded.
  • Days to pay is measured from invoice receipt. An invoice that sits unapproved because of a missing purchase order number is late for reasons the report cannot show.

How I would take it live

  • Export aged receivables and two years of receipts from the ledger, and list the large customers with their published reports
  • Build the timing table and compare its forecast against the last six months of actual receipts
  • Load expected payment dates into the cash forecast in Anaplan, Fathom or Excel, with a likely and a slow case
  • Refresh the table when customers file their reports and review the late-payer flags before each new contract

Financial impact

About A$1,508k less cash held back against forecast error each month

The gain is a smaller safety buffer in the overdraft or the supplier run, because first-month receipts are forecast closer to what arrives. Invoice values and the facility rate are assumptions.

InputValueSource
Invoices to large customers a monthA$4,000,000 mid-market supplier to large customers
Error in cash paid within 30 days, due-date forecast40.0 ptsMeasured 5,000 test ledgers
Error with last published reports2.3 ptsMeasured
Overdraft or facility rate9%

Calculation

  1. First-month forecast error, due datesA$1,600,171
  2. First-month forecast error, published reportsA$92,368
  3. Buffer no longer neededA$1,507,803
  4. Annual interest on that bufferA$135,702

Annual interest savedA$135,702

Better credit terms for persistent late payers and fewer surprises in the supplier run are further gains and are not counted here.

Implementation and rollout

Four weeks from receivables export to a cash forecast on expected dates

  1. Week 1

    Export and list

    Aged receivables and two years of receipts from the ledger; large customers matched to their published payment reports.

    Finance manager, Manaz

  2. Week 2

    Timing table

    Expected days and 80% range for each customer, checked against the last six months of actual receipts.

    Manaz

  3. Week 3

    Cash forecast

    Open invoices placed at expected dates in Anaplan, Fathom or Excel, with a likely and a slow case.

    Finance manager, accountant

  4. Week 4 onward

    Refresh and review

    Table refreshed as customers file reports; late-payer flags reviewed before each new contract.

    Credit controller, Manaz on support

Tools

NetSuite, Dynamics 365, MYOB Advanced or Xero receivables, Anaplan, Fathom or an Excel cash forecast, Published payment-times registers (UK Payment Practices, Australian Payment Times Reporting Scheme), Python and scikit-learn

Team

  • Treasury or finance manager owns the cash forecast and credit terms
  • Credit control team reviews late-payer flags and chases
  • Manaz timing table, ranges and forecast integration

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
A customer pays one supplier differently from its averageThe business’s own receipts replace the public figure once a few months of history exist.
Self-reported figures contain errorsImplausible reports are excluded and each customer’s figure is checked against its own receipts.
A customer stops filingThe last figure is kept with a wider range and flagged for review.
Invoices held for missing PO numbersInvoices are checked for PO and approver details before sending, which the report cannot show.

Key takeaways

Key takeaways

  • Due-date forecasts miss days to pay by 13.3 days on average; the customer’s last published report by 4.5.
  • A boosted model on the full history adds a fraction of a day, so the simpler method is the one that goes live.
  • Heavy late payers stay late 75% of the time, which belongs in credit terms before the contract starts.
  • The business’s own receipts take over once they exist.