Data Department built a collections forecasting model for a finance company's treasury. The treasury needed to know how much cash would come in from customer repayments. By using customer-level risk scores, the new model reduced forecast error, measured as MAPE (mean absolute percentage error), from 32% to 9.5%.
What was the decision?
A treasury team plans liquidity. It decides how much cash to keep available, how much to borrow and when, and how much can be put to work. All of these decisions depend on one input: how much money customers will actually pay back, and when.
When the collections forecast is wrong by about a third, the treasury has two poor options. It can hold large cash buffers, which cost money. Or it can run tighter and risk a shortfall. A forecast error of 32% meant the treasury was planning around a wide band of doubt.
The question was not "can we build a more complex model?" It was "how much will come in, so the treasury can plan with confidence?"
What did we build?
The key change was to build the forecast on customer-level risk scores: an estimate, for each customer, of how likely they are to pay as agreed. The forecast then reflects who owes the money, not only how much is owed.
| Measure | Before | After |
|---|---|---|
| Forecast error (MAPE) | 32% | 9.5% |
| What the treasury gets | A forecast with a wide error band | A forecast it can plan liquidity around |
Why customer-level risk matters. Total collections are the sum of many individual payment behaviours. A portfolio where risky customers make up a growing share will collect less than its history suggests. A model that sees the risk of each customer can pick up that shift; a model that sees only totals cannot.
In general terms, projects of this kind combine two steps: score the payment risk of each customer, then roll those scores up into an expected cash flow for the treasury. We describe the general method here; the client's data and models stay confidential.
Results
- Forecast error (MAPE) reduced from 32% to 9.5%.
- Customer-level risk scores as the core input of the collections forecast model.
What made it work?
We started from the treasury's decision. Liquidity planning needed a reliable view of incoming cash. That set the target: a forecast error the treasury could plan around.
Customer-level detail carried the gain. The improvement came from looking at customers one by one, through their risk scores.
Risk models and forecasting models can share work. In many finance companies, credit and payment-risk scores already exist for other decisions. In general, reusing that view of the customer for treasury forecasting is a practical way to get more value from the same data.
This work sits within our demand forecasting service and our work for financial services and insurance. A related example is our insurance renewal case.
Frequently asked questions
What is MAPE?
MAPE (mean absolute percentage error) is the average size of forecast errors, expressed as a percentage of the actual values. A MAPE of 9.5% means the forecast was, on average, 9.5% away from what actually happened.
Why did customer-level risk scores improve the forecast?
Collections are the sum of many individual payment behaviours. Risk scores show how likely each customer is to pay, so the forecast reflects changes in the mix of customers, not only past totals.
Is collections forecasting only useful for finance companies?
No. Any business with many customers paying over time, such as leasing, subscription or business-to-business trade on credit terms, needs to forecast incoming cash. The same approach applies.