Grounded in published research, not folklore.
We did not invent the idea that customer payment history predicts late payment, or that early, plain reminders get paid. We productised the studies that showed it.
Adding aggregated customer payment history to invoice fields lifts late-payment prediction accuracy by 9 to 27 percentage points.Zeng et al., KDD 2008
Prioritising collectors by value × probability of lateness saved a multinational bank about $1.75M a month in simulation.Zeng et al., KDD 2008
Invoices chased in the first week overdue are about 2.5 times more likely to be paid than those first chased after 30 days.Atradius Payment Practices Barometer
Invoice-level prediction with customer history
Zeng, Melville, Lang, Boier-Martin and Murphy (IBM Research, KDD 2008) framed collections as a supervised learning problem: for each open invoice, predict whether it will be paid late and by how much. Their key finding was that features aggregated over the customer's payment history, such as the share of past invoices paid late and the average days to pay, matter far more than the fields on the invoice itself. Appel, Malfatti, Cunha, Lima and de Paula (2020) reproduced the result with gradient-boosted trees on a different corporate ledger.
CashPilot follows that recipe. A HistGradientBoosting classifier and regressor are trained per organisation on 24 features spanning customer history, invoice attributes, seasonality, disputes and reminder activity, with a time-ordered split so the model is only ever evaluated on invoices newer than the ones it learned from. On the demo ledger it reaches 79% accuracy against a 52% majority-class baseline with an AUC of 0.83. Every organisation can see its own calibration, feature importance and backtest on the model page and retrain in one click.
The ranking rule
The same paper showed that ordering collectors' work by expected loss, balance multiplied by the probability of late payment, beats ordering by balance or by age. It is why the biggest overdue invoice is often not at the top of a CashPilot worklist: a large invoice from a customer who always pays in week two is less urgent than a medium one from a customer who has slipped three times this year.
The reminder cadence
Atradius's Payment Practices Barometer finds that the probability of collecting an invoice falls steeply with time overdue, and that an invoice chased in its first week overdue is about 2.5 times more likely to be paid than one first chased after 30 days. CashPilot's default sequence therefore contacts at −3 (courtesy), +1 and +7 days before escalating at +14, +30, +45 and +60, with call tasks where the evidence says phone beats email.
Behavioural field experiments on payment reminders, including work with tax authorities and utilities, consistently find that plain, personalised messages stating the amount, the due date and how to pay outperform elaborate designs, while social-norm appeals ("most customers pay on time") do not move payment. CashPilot's templates are short, specific and attach the invoice, the statement and a pay link.
The metrics
DSO, Best Possible DSO, Average Days Delinquent and the Collection Effectiveness Index are computed to the Credit Research Foundation's published definitions, so the numbers on the dashboard are comparable with the benchmarks your board already reads. Smirnov (Tartu, 2016) informs the survival-style haircut applied to balances past 60 days in the cash forecast.
References
- Zeng, S., Melville, P., Lang, C., Boier-Martin, I., Murphy, C. Using predictive analysis to improve invoice-to-cash collection. KDD 2008.
- Appel, A. P., Malfatti, G. L., Cunha, R. L. F., Lima, B., de Paula, R. Predicting Account Receivables with Machine Learning. 2020.
- Smirnov, J. Modelling late invoice payment times using survival analysis and random forests. University of Tartu, 2016.
- Atradius. B2B Payment Practices Barometer, 2025 editions for Western Europe and the UK.
- Intuit QuickBooks. Small Business Late Payments Report, 2025.
- Credit Research Foundation. Performance measures: DSO, BPDSO, ADD, CEI.
Stop financing your customers’ cash flow.
Import a CSV and see your own ranked worklist and 13-week forecast in about ten minutes. Or click around the live demo workspace with 18 months of ledger history.