Key Takeaways
- CashPulse from Transformance reaches 90-95% forecast accuracy (100 minus MAPE) out to 90 days (Transformance customer data, 2026).
- MAPE punishes over-forecasts harder: in the example below, the same EUR 0.5m miss scores 5.00% too low and 5.26% too high.
- In the same four-week example, one forecast reads 94.3% (100 minus MAPE), 94.4% (100 minus WAPE) and hits a 5% band in 2 of 4 weeks.
In This Article
- Key Takeaways
- How accurate is AI cash flow forecasting?
- What does 100 minus MAPE actually measure?
- Why does MAPE treat a forecast that is too high differently from one that is too low?
- Worked example: how is 100 minus MAPE calculated?
- What changes the accuracy figure besides the model?
- How do you test an AI forecast’s accuracy on your own data?
- What accuracy should you expect from your own data?
How accurate is AI cash flow forecasting?
CashPulse from Transformance reaches 90-95% forecast accuracy (100 minus MAPE) out to 90 days, and accuracy falls as the horizon lengthens, so every figure needs its horizon.
Forecast accuracy is a score that compares what a frozen forecast said with what actually happened, at a stated horizon. Ours comes from Transformance customer data (2026). CashPulse forecasts on 30-, 90- and 365-day horizons, and error grows with distance.
As a baseline, forecast variance in spreadsheet-led treasury environments “often exceeds 20%” (EY India, 2026).
No universal benchmark ranks accuracy figures, ours included, because each rests on its own metric, horizon, level and test period. How AI improves cash flow forecast accuracy covers the methods; this guide covers how the number is computed and tested.
What does 100 minus MAPE actually measure?
100 minus MAPE subtracts the average percentage miss per forecast period from 100, so a 6% MAPE at a 90-day horizon reads as 94% accuracy at that horizon.
MAPE is the mean absolute percentage error. Hyndman and Athanasopoulos define the percentage error as p = 100 e / y, where e is the error (actual minus forecast) and y the actual, and MAPE as the mean of the absolute values of p (Forecasting: Principles and Practice, 2021). Three things follow:
- Every period counts the same. A EUR 2m week and a EUR 20m week each add one percentage to the average.
- The actual sits in the denominator. Each miss is scaled by what happened, not by what you planned.
- Only frozen forecasts count. A rolling forecast overwrites each run with the next, so save every run with its date and score that version.
Why does MAPE treat a forecast that is too high differently from one that is too low?
MAPE divides each miss by the actual, so percentage errors put a heavier penalty on forecasts that land too high than on forecasts that land too low.
Percentage errors “put a heavier penalty on negative errors than on positive errors” (Hyndman and Athanasopoulos, 2021). A negative error means the forecast was above the actual.
Two mechanics produce it. An under-forecast can never score worse than 100%, the score of a forecast of zero. An over-forecast has no ceiling: forecast EUR 3m for a week that brings EUR 1m and the week scores 200%. And when you over-forecast, the actual is the smaller number, the one MAPE divides by.
Kolassa and Martin showed “how you can be led astray if you use the MAPE” to select a forecasting method or to reward forecast accuracy (Foresight, 2011). Minimise MAPE and you reward forecasts that lean low. The line decides whether that hurts:
- Receipts: cash arrives above plan, the cheaper mistake, though it ties up buffer you did not need.
- Disbursements: outflows arrive above plan, the expensive mistake.
- Net cash: the wrong place for MAPE, because the series crosses zero, where a percentage error is undefined. The accuracy guide explains what to use instead.
Worked example: how is 100 minus MAPE calculated?
On an illustrative four-week receipts forecast, the weekly misses average 5.69%, so 100 minus MAPE reads 94.3%, while 100 minus WAPE reads 94.4% on the same data.
One entity, four weeks of receipts in EUR m, each week scored against the point forecast frozen one week ahead. Illustrative figures, not customer data.
- MAPE: (5.00 + 5.26 + 5.00 + 7.50) / 4 = 5.69%, so 100 minus MAPE is 94.3%.
- WAPE: (0.5 + 0.5 + 0.6 + 0.6) / 39.5 = 5.57%, so 100 minus WAPE is 94.4%.
- Bias: the errors sum to zero, so bias is 0.0%.
- Hit rate within +/-5%: weeks 1 and 3, so 2 of 4 (50%).
Weeks 1 and 2 use the same two numbers, swapped: the EUR 0.5m under-forecast scores 5.00%, the EUR 0.5m over-forecast 5.26%. Week 4 is the worst at 7.50%, although its EUR 0.6m miss equals week 3’s: it is the smallest week and an over-forecast.
MAPE and WAPE land 0.1 points apart because the weeks are of similar size. A wide gap means small weeks drive MAPE: in the accuracy guide’s five-week example, one bank holiday week opens a gap of almost 7 points.
One forecast, four ways to state its accuracy
100 minus MAPE fits when every week matters equally and no week is near zero. 100 minus WAPE answers how much of the cash you missed; where weekly volumes swing, it is the more defensible headline, which is why the accuracy guide leads with it. Bias belongs next to either.
What changes the accuracy figure besides the model?
Five choices can move an accuracy figure as much as the model does: the horizon, the level, the line scored, the test period and whether forecasts were frozen.
Horizon. 30, 90 and 365 days are three separate claims. A strong week-two score says nothing about month three.
Level. Group totals score better than entities because misses offset; invoice-level payment timing is the hardest test. See multi-entity cash flow forecasting for how levels roll up.
Line. Receipts, disbursements and net cash need separate scores.
Test period. 12 months or more, out of sample, so quarter ends and holiday weeks are in the test. Four calm weeks prove little.
Frozen forecasts. A forecast refitted after the fact scores better than the one you planned with.
CashPulse runs 20+ countries on one model with 100+ drivers, so one test design covers every entity. Ask for entity scores next to the group figure: the entity is where funding decisions are made.
How do you test an AI forecast’s accuracy on your own data?
Agree the metric in writing, freeze every forecast, score 30- and 90-day horizons separately on your own actuals, and report bias and WAPE next to 100 minus MAPE.
Six things to agree in writing before an accuracy test:
- The metric and its formula, written out so both sides compute the same number.
- The horizons, scored separately: 30 and 90 days, with 365 days as a planning range.
- The level and the line: invoice, entity or group; receipts and disbursements separately, never net cash on MAPE.
- The test period, and the rule that every forecast is saved with its date before actuals are known.
- The comparison: your current forecast and a naive due-date forecast on the same weeks.
- What sits next to the headline: WAPE, bias and the five largest misses with their causes.
Your first 13-week forecast from CashPulse is live in 2-4 weeks, and a proof of value runs 4-5 weeks on one entity. That window scores the near weeks. A forecast frozen on day one reaches its 90-day mark in month three, so agree a backtest on your history for that horizon. A 13-week cash flow forecast built from due dates makes a fair naive benchmark.
What accuracy should you expect from your own data?
Expect the highest accuracy where receipts are frequent and payment history is long, and lower accuracy for new customers, lumpy project payments, disputes and one-off items.
A model predicts payment timing from how each customer has paid before: three years of monthly invoices give it a pattern, a new customer one data point. Your AR data moves the score too:
- Unapplied cash distorts the actuals. A receipt not yet matched leaves its invoice looking open. Touchless cash application cleans the actuals you score against.
- Disputes delay or shrink receipts, so a forecast that does not know about one over-forecasts the week. See AI dispute management.
- Promises to pay are dated commitments a due-date forecast cannot see; read how AI tracks promise to pay.
In Transformance, payment history, promises to pay, disputes and deductions sit in Customer 360, one graph per customer that Vero, the AI agent, works on. When you review the five largest misses each month, the cause of many (a slipped promise, an open dispute, an unapplied payment) sits in one record. More causes: why most cash flow forecasts are inaccurate.
One honest limit: CashPulse forecasts out to 365 days, but the 90-95% (100 minus MAPE) is stated out to 90 days. Hold us to that.
Conclusion: an accuracy figure is only as good as its method
Ask for the metric, horizon, level, test period and bias with every accuracy figure, and test it on your own actuals before you rely on it.
Ours, stated that way: CashPulse reaches 90-95% forecast accuracy, measured as 100 minus MAPE, out to 90 days (Transformance customer data, 2026). Test it on the six points above in a proof of value.
Frequently Asked Questions
How accurate is AI cash flow forecasting?
AI cash flow forecasting is most accurate at short horizons: CashPulse from Transformance reaches 90-95% forecast accuracy (100 minus MAPE) out to 90 days. Any figure should state its horizon, metric and test period.
What does 90% forecast accuracy mean?
On a 100 minus MAPE reading, 90% means the forecast missed by 10% of the actual in an average period at the stated horizon. On 100 minus WAPE, the total miss was 10% of the cash that moved.
How is CashPulse forecast accuracy measured?
CashPulse states its accuracy as 100 minus MAPE: 90-95% out to 90 days (Transformance customer data, 2026). In a proof of value, recompute it on your own frozen forecasts and actuals.
Why does MAPE penalise over-forecasts more than under-forecasts?
MAPE divides each miss by the actual, so an under-forecast can never score worse than 100% while an over-forecast has no upper limit. A forecast tuned to minimise MAPE tends to lean low.
How long does it take to know whether an AI cash forecast is accurate?
A few weeks for the near horizons, about three months for live 90-day scores and a year for the 365-day view; a backtest on history bridges the gap.
Why can’t you compare accuracy figures from different forecasts directly?
Each figure rests on its own metric, horizon, level, line and test period, and each choice moves the number, as the worked example shows.
