Multi-currency cash forecasting: what machine learning really adds
Vendors promise "AI-driven" cash forecasts. The major forecasting competitions show where machine learning wins — many similar series, plenty of history — and where it loses. Applied to treasury: which flows benefit, which do not, and how to tell whether a tool beats a simple method.
The 13-week cash forecast is a finance department's most useful steering tool, and one of the most laborious to maintain. The promise of a model that would feed it on its own, from bank statements and the ERP, is therefore appealing. It deserves to be examined in the light of what is known about forecasting methods.
What the major forecasting competitions say
Since the 1980s, the competitions run by Spyros Makridakis have pitted forecasting methods against each other on thousands of real series. The last two shed direct light on the question.
| Competition | Data | What won |
|---|---|---|
| M4 (results published in 2020) | 100,000 series of very diverse kinds | Pure machine-learning methods did worse than a combination of three simple statistical methods; the winner was a hybrid of exponential smoothing and a neural network |
| M5 (results published in 2022) | 42,840 hierarchical sales series from one retailer | Machine-learning methods, notably tree-based ones, dominated |
The lesson for treasury is direct. A machine-learning model shines when it can learn from thousands of comparable cases; it adds nothing, or even worsens results, on a few short, heterogeneous series.
Cash flows are difficult series
A study of companies' daily cash flows shows that the usual assumptions of statistical models — normal distribution, no correlation, stability over time — rarely hold, and that non-linear effects often matter for forecasting (Salas-Molina et al., 2018).
That is an argument for more flexible methods, hence for machine learning — provided there is enough data for it to learn something other than noise.
Which flows benefit, which do not
| Type of flow | Machine-learning contribution | Why |
|---|---|---|
| Numerous customer receipts | High | Thousands of invoices, each customer's own payment behaviour: this is the M5 case |
| Recurring supplier payments | Medium | Enough volume, but dates are largely decided by the company |
| Payroll, rent, taxes, loan repayments | None | The calendar is known: a rule is enough |
| FX hedge deliveries | None | Amounts and dates are contractual |
| Intragroup flows | Low | They follow internal conventions, not behaviours |
| Exceptional flows: acquisition, down payment, dividend | None | Rare, large, decided: a matter of judgement |
The benchmark rule
The reference forecasting textbook, by Rob Hyndman and George Athanasopoulos, sets a simple rule: any sophisticated method must be compared with naive methods — repeating the last observed value, or the value of the same period last year. If it does not do better, it is not worth using.
Applied to treasury, the rule becomes a test every tool must pass:
- Choose a past period the tool did not see while being tuned.
- Compare, currency by currency and horizon by horizon (4, 8, 13 weeks), the tool's error with that of a naive method.
- Keep the tool only for the flows and horizons where it does clearly better.
A serious vendor accepts this test on your own data. An accuracy rate announced without a benchmark says nothing.
What you need before you start
- A long history: for annual seasonality to be learned, at least two years of data are needed.
- Flows classified consistently, by category and currency: a model learns nothing from a bank description that changes form every month.
- Contractual maturities separated from behavioural flows: hedges, loans and payroll do not need to be predicted.
- An archive of forecast versus actual: without it, nothing can be measured.
The link with FX hedging
This is where multi-currency forecasting shows its full value. The forecast error measured by currency and horizon tells you how much of the exposure is certain enough to be hedged. If four-week dollar receipts are forecast within a few per cent, they can be hedged broadly; if the 13-week error is large, only a base layer deserves hedging.
A machine-learning model that reduces the error on customer receipts therefore does more than improve the forecast: it increases the share of exposure the company can hedge without risking over-hedging.
Where to start
- 1Separate the certain from the behavioural
Take out of the forecast everything that is contractual or calendar-based: it is carried over, not forecast.
- 2Measure today's error
Before any tool, archive a few weeks of forecast and actual, by currency, to know the starting point.
- 3Test on customer receipts
That is the flow where a model is most likely to add something. Compare it with a naive method over a past period.
- 4Extend only what wins
Keep the tool where it beats the benchmark, and judgement where it does not.
Sources
- Makridakis, S., Spiliotis, E. and Assimakopoulos, V. (2020), "The M4 Competition: 100,000 time series and 61 forecasting methods", International Journal of Forecasting, 36(1), pp. 54-74.
- Smyl, S. (2020), "A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting", International Journal of Forecasting, 36(1), pp. 75-85.
- Makridakis, S., Spiliotis, E. and Assimakopoulos, V. (2022), "M5 accuracy competition: Results, findings, and conclusions", International Journal of Forecasting, 38(4), pp. 1346-1364.
- Salas-Molina, F. et al. (2018), "Empirical analysis of daily cash flow time-series and its implications for forecasting", SORT, 42(1), pp. 73-98.
- Hyndman, R. J. and Athanasopoulos, G. (2021), Forecasting: Principles and Practice, 3rd edition, OTexts, chapter "Some simple forecasting methods".
Your hedges in the forecast, without re-keying
ELEGG FX keeps the schedule of your hedges and their market value: contractual currency deliveries flow into the forecast as they are.
Discover ELEGG FXNot ready to open the pricer yet?
Get the one-page FX policy template — hedge ratios by horizon, tolerance band, delegation table. The one your auditor will ask for.
- One field only
- One-click unsubscribe
- No data resold
Frequently asked questions
The questions finance departments ask most often on this topic
Does AI improve cash forecasting?
On some flows, yes. Machine learning mainly helps on numerous, behavioural flows such as customer receipts. It adds nothing on contractual or calendar flows — payroll, loan repayments, hedges — nor on exceptional flows, which are a matter of judgement.
How can I tell whether an AI forecasting tool performs well?
By comparing its error with that of a naive method — repeating the last value or the same period last year — over a past period it has not seen, currency by currency and horizon by horizon. If it does not do clearly better, it is not worth using.
How much history is needed to forecast cash with machine learning?
At least two years for annual seasonality to be learned, with flows consistently classified by category and currency, and an archive of forecast versus actual.
What is the link between cash forecasting and FX hedging?
The forecast error by currency and horizon indicates how much of the exposure is certain enough to be hedged. Reducing that error lets the company hedge a larger share without risking over-hedging.
Read next
Multi-currency cash forecasting: building a reliable 13-week forecast
15 min read
Hedging when you do not know your volumes: the layered method
9 min read
Hedge ratio: what share of your exposure should you really hedge?
9 min read
AI and currency risk: what it does well, what it does badly, and where it becomes dangerous
9 min read
Choosing treasury software: the decision grid for an SME or mid-sized company
10 min read