Cash Flow Forecasting: Three Horizons and a Test for Each

Cash flow forecasting estimates the cash a company will receive and pay in each coming period, so that decisions on spending, hiring, borrowing and raising come before the balance forces them. It runs on three horizons with two methods. The weekly and monthly forecasts earn their detail only where they beat a naive forecast, the benchmark forecasting research recommends.

Treasurers rank the work as their hardest. In the Association for Financial Professionals’ 2026 benchmarking survey, 49% of 425 treasury practitioners named cash and liquidity forecasting their most challenging activity (AFP, 2026). Yet the obvious way to score that work, a percentage gap, breaks when a balance nears zero. The sounder test compares a forecast with the simplest alternative: next week will look like last week.

Cash flow forecasting runs on three clocks, and each answers a different question

	13 weeks	12 months	2 to 5 years
Question	Can the company pay what falls due each week?	Does the plan stay funded through the year?	How much capital to raise, and when?
Time step	Week	Month	Quarter or year
Method	Direct: each receipt and payment on its expected date	Indirect: profit adjusted for working capital	Driver-based scenarios
Built from	Bank, receivables and payables ledgers, payroll	Operating plan, collection and payment days, hiring plan	Growth, margin and burn assumptions
Decision	Payment timing, credit-line draws	Hiring pace, budget, covenant headroom	Round size and timing
Accuracy test	Error by line against last week’s actual (naive)	Error and bias against the same month last year (seasonal naive)	Whether outcomes fall inside the stated range
Who checks it	Treasury and lenders	Boards and auditors (going concern looks at least 12 months ahead)	Investors

Source: CX Cash synthesis; IAS 7, paragraphs 18 to 20; IAS 1, paragraph 26; Hyndman and Athanasopoulos, Forecasting: Principles and Practice, sections 5.2 and 5.8.

Cash flow forecasting exists because profit and cash arrive at different times

IAS 7, the international standard on cash flow statements, gives the reason in its objective. Users must judge a company’s ability to generate cash and “the timing and certainty of their generation” (IAS 7, as endorsed in the EU). The standard also fixes what is being forecast. Cash is cash on hand and demand deposits; a cash equivalent must be “readily convertible to known amounts of cash”, which normally means a maturity of 3 months or less. A forecast that counts a six-month deposit, or a customer’s promise, as cash is forecasting something else.

The three horizons in the table above follow from the decisions they serve. The 13-week forecast covers a fiscal quarter week by week, close enough to date each receipt and payment, and a separate piece on the 13-week cash flow forecast covers its mechanics. The 12-month model matches the going-concern test, under which management considers information about the future that is “at least, but is not limited to, twelve months from the end of the reporting period” (IAS 1, paragraph 26). The multi-year view sizes the next round, a decision that has to be taken well before the money is needed. A company that runs one clock makes the other two decisions blind, and the bill can arrive as a bridge round, a covenant breach or a hiring freeze ordered after the hires have started.

The direct method sees the weeks; the indirect method depends on working capital

IAS 7 allows two ways to present operating cash flow. The direct method shows “major classes of gross cash receipts and gross cash payments”; the indirect method starts from profit and adjusts for non-cash items, accruals, deferrals and changes in receivables, payables and inventory. The standard states a preference and its reason: “The direct method provides information which may be useful in estimating future cash flows and which is not available under the indirect method” (IAS 7, paragraph 19). A separate piece on direct vs indirect cash flow forecasting sets out the mechanics of each.

Research on published accounts backs the reason, at least for outsiders forecasting from them. Steven Orpurt and Yoonseok Zang found that direct-method line items, such as cash received from customers, “are not reliably estimable using income statements and either balance sheets or indirect method statements of cash flows” (Orpurt & Zang, 2009). Adding the gap between the reported items and their estimates to forecasting models made cash flow and earnings forecasts significantly better. A study of pre-IFRS Australian companies reached the same result: splitting operating cash flow into receipts from customers and payments to suppliers and employees improved forecasts of future cash flow (Farshadfar & Monem, 2013). Both studies concern investors working from published statements rather than a finance team forecasting its own cash, and both findings come from the papers’ abstracts, so their samples are not reported here.

The indirect method’s weak point is the working capital it must forecast. Dan Givoly, Carla Hayn and Reuven Lehavy found that analysts’ cash flow forecasts were less accurate than their earnings forecasts and improved more slowly through the year. They also looked like an extension of the earnings forecast, “thus providing limited information on expected changes in working capital” (Givoly, Hayn & Lehavy, 2009). Analysts work from outside, but a company’s own 12-month model, built from profit, can carry the same blind spot. For a hypothetical software company with $6M of annual revenue, customers moving from paying in 45 days to 60 leave about $247k more cash in receivables, with no change in profit. The receivables and payables lines carry the cash, and forecasting them is a discipline of its own.

Where the two forecasts overlap, they should agree. Weeks 1 to 13 of the direct forecast cover months 1 to 3 of the indirect one, and a gap between them often points to a wrong working-capital assumption on one side. Mentor Graphics’ treasury team described this practice to the AFP: it merged a direct forecast built from its subsidiaries’ cash with the company’s indirect forecast to produce quarterly projections (AFP). For software companies, a driver-based model of bookings, billings and collections can bridge the two.

Error grows with the horizon, and the balance compounds it

Accuracy falls as a forecast reaches further out, and the Congressional Budget Office has measured how fast. Reviewing the baseline projections it has made each spring since 1984, CBO found an average absolute error in its deficit projections of 1% of GDP for the budget year and 2% for the sixth year. For federal debt the figures were 1.7% and 7.1%, “because errors in projections of debt held by the public compound each year” (CBO, 2019). Between the budget year and the sixth year, the flow error doubled and the stock error grew about 4.2 times (our arithmetic).

A cash balance is a stock in the same sense as federal debt, so its forecast error accumulates by the same mechanism: week 13’s balance carries every earlier week’s misses on receipts and payments. Only the mechanism transfers. CBO’s ratios describe a federal budget moved by the economy and by Congress, and they say nothing about how fast a company’s error grows; misses can also offset, as CBO’s revenue and outlay errors partly did on average. So measure the flows line by line, and expect the balance to drift further than any single line.

Averages also understate bad years. CBO’s deficit projection for 2023 was $1.0 trillion, and the actual deficit was $1.0 trillion larger. The error, 3.9% of GDP, was more than 3.5 times its long-run average of 1.1% (CBO, 2023).

The Bank of England met the same problem in drawing its uncertainty. Until early 1996 it published its inflation projection with a shaded band: the central path plus or minus “the average absolute value of past forecast errors”. The Bank expected the outcome to fall inside it “just over half of the time”. Readers took the band for “upper and lower bounds for the forecast”. From February 1996 the Bank published a fan chart, a probability distribution that widens with the horizon, meant to convey its view “without suggesting a degree of precision that would be spurious” (Britton, Fisher & Whitley, 1998). The arithmetic explains the misreading. With normally distributed errors, 1 average error either side covers about 57.5% of outcomes, and an 80% range needs about 1.6 average errors either side (our arithmetic).

For the 12-month and multi-year views, the number that matters is often a date: the week the balance crosses a floor. Forecasters have a measure for that, the timing error, “the difference in time between when the outcome crosses the threshold and when the forecast does so” (Forecast error, Wikipedia).

The US Treasury ran a forecast of this kind in public in May 2023, with its borrowing capped by the debt limit. Its estimate of the day cash would fall short moved from early June, possibly “as early as June 1”, to June 5, after more than $130 billion of scheduled payments on June 1 and 2 (NPR, 2023). Congress passed the suspension on June 1 (US Treasury, 2023). A company’s cash-out date deserves the same scrutiny; in our view, how far it moves between forecasts says more about the forecast than any single balance.

Score a cash forecast on its flows and against a naive forecast

Many companies score their forecasts by hand or not at all. In Strategic Treasurer’s 2021 survey of nearly 250 treasury professionals, 57% ran variance analysis fully by hand and one-fifth ran none (treasuryXL, 2022). A software vendor sponsored the survey, and its respondents chose to take part. The same firm’s surveys show the share of treasurers calling forecasting difficult rising from 39% in 2018 to 53% in 2025 (CTMfile, 2025).

IAS 7 already assumes the check, describing historical cash flows as useful “in checking the accuracy of past assessments of future cash flows” (IAS 7, paragraph 5). The obvious way to run it is a percentage gap per line against a variance target such as 5%, and for a balance that is the wrong unit.

Rob Hyndman and Anne Koehler showed that percentage error measures are “infinite or undefined” when an actual value is zero and extremely skewed near it. The M3 forecasting competition sidestepped the problem by a route no treasurer can take: “Excessively large (or infinite) MAPEs were avoided in the M3-competition by only including data that were positive” (Hyndman & Koehler, 2006).

CBO takes the same precaution with deficits, scaling its errors by GDP because “relatively small differences between projected and actual amounts in years with small deficits or surpluses would result in large projection errors” (CBO, 2023). In a hypothetical company, a $32k miss is 400% of an $8k closing balance and 12.8% of $250k of weekly receipts (our arithmetic). Percentage error is undefined on the day the balance reaches zero, which is the day the forecast was for.

The benchmark is the naive forecast: next period equals the latest actual or, for seasonal lines, the same month last year. Hyndman and George Athanasopoulos’s textbook says the naive method works well “for many economic and financial time series”, and sets the rule for anything more elaborate: “If not, the new method is not worth considering” (Hyndman & Athanasopoulos). The test is a ratio. Divide the forecast’s mean absolute error (MAE) by the naive forecast’s over the same weeks; below 1, the forecast adds information, and above 1 it does worse than the naive guess (Hyndman & Koehler, 2006). CBO’s evaluation adds two tests a board can ask for: centeredness, whether errors lean one way, and dispersion, how widely they scatter (CBO, 2019).

A hypothetical receipts forecast beats the naive forecast for about 8 weeks, then loses

Horizon	Forecast MAE ($k)	Naive MAE ($k)	Relative MAE	Reading
1 week	22	48	0.46	Adds information
4 weeks	41	55	0.75	Adds information
8 weeks	58	61	0.95	Close to break-even
13 weeks	78	66	1.18	Worse than carrying the last week forward

Hypothetical company with weekly customer receipts of about $250k, scored over 26 weeks of saved forecasts. Naive forecast: the latest actual week carried forward. Relative MAE = forecast MAE ÷ naive MAE (Hyndman and Koehler, 2006). Ratios are our arithmetic.

This company’s receipts forecast stops beating the naive one at about 8 weeks. Beyond that point, the line-by-line receipts forecast should give way to the run rate or to a range, and the hours belong to the lines that still win. Each line crosses over at its own distance.

An average of three simple forecasts beat most entrants in the M4 contest

Contracts fix payroll, rent and debt service, so those lines need a schedule more than a technique; receipts are where method matters. The M4 forecasting competition of 2018 tested methods on a scale no company could: 61 of them on 100,000 series, about a quarter financial, with every score published (M4 results).

How a plain average of three smoothing methods scored in M4

Method (naive forecast = 1.000)	All series	Daily series	Finance series
Winner: statistical and machine-learning hybrid (Uber)	0.821	1.046	0.822
Average of three exponential-smoothing forecasts	0.898	0.978	0.874
Best pure machine-learning entry	0.915	1.071	0.900
Naive forecast, seasonally adjusted	1.000	1.000	1.000

Source: M4 organisers, Evaluation and Ranks (2018). Scores are OWA: sMAPE and MASE, each divided by the naive forecast’s, then averaged; lower is better. The average is of simple, Holt and damped-trend exponential smoothing.

The plain average, which a spreadsheet can reproduce, beat 32 of the 49 submitted methods. None of the six pure machine-learning methods beat it, and five did worse than the naive forecast (Makridakis Competitions, Wikipedia). On the 4,227 daily series, only 13 of 49 submissions beat the naive forecast, and the winner was not one of them.

Averaging pays when forecasts miss in different directions. Across 30 comparisons reviewed by J. Scott Armstrong, an equal-weighted average cut error by about 12.5% against the typical component (Armstrong, 2001). His rule is to combine methods that “differ substantially and draw from different sources of information”. For receipts, those are the ledger’s due dates, a trailing average and the same month last year. A bias the components share survives: if every method assumes customers pay on time, so does the average.

The organisers concluded that M4 “proved that complex methods provided accuracies that were well above that of simple ones” (Makridakis & Petropoulos, 2020). In M5, on 42,840 series of Walmart unit sales, the leading entries were pure machine-learning methods and beat every statistical benchmark (Wikipedia). On the official M5 scores, though, the winner beat the best benchmark by 43.2% on total sales and by only 2.3% on the 30,490 single product-store series. A startup has one receipts series, a few dozen months long. A machine-learning forecast of it faces the same test as any other: beat the naive forecast, then the plain average.

Can a forecast be scored when acting on it changes the outcome?

The best case against scoring cash flow forecasting comes from the textbook behind the naive benchmark. Hyndman and Athanasopoulos list the conditions for predictability, and the last is “whether the forecasts can affect the thing we are trying to forecast” (Hyndman & Athanasopoulos). A cash forecast exists to do exactly that. When it shows a shortfall in week 9 and the company chases two invoices and defers a supplier, the shortfall never arrives, and the forecast scores badly for having worked. The textbook also warns that a forecast error “does not mean a mistake, it means the unpredictable part of an observation” (Hyndman & Athanasopoulos).

The textbook’s condition holds on two counts: the forecast that gets scored must be the one made before anyone acted, and a forecast used well will often look wrong. It does not follow that scoring is impossible. CBO faces the same problem, since Congress keeps changing the laws its projections assume. It handles the problem by design: its baselines assume current law, and its evaluation “focuses on differences between projected and actual deficits that result from nonlegislative factors”, with the effects of legislation analysed separately (CBO, 2019). A company can copy the design. Freeze each forecast before management acts, log every action taken because of it with its expected cash effect, and score the forecast on everything else.

Freeze each forecast, then score it line by line

1. Run three clocks: a direct 13-week forecast, an indirect 12-month model and a multi-year range for the next raise. Reconcile weeks 1 to 13 with months 1 to 3 every month.
2. Save every forecast before acting on it, and log the actions it prompted.
3. Score receipts and payments by line, at 1, 4 and 13 weeks, against last week’s actual and the same month last year. Report relative error, never a percentage of the closing balance.
4. Report the bias on its own line: receipts that arrive late every month call for a corrected forecast, whatever the average error.
5. Find, for each line, the horizon at which the forecast stops beating the naive one. Beyond it, show a range: about 1.6 average past errors either side gives an 80% band if errors are roughly normal.
6. Where a line fails the naive test, try the plain average of two or three simple forecasts before anything elaborate. Log every manual override with its sign. In one study of more than 60,000 supply-chain forecasts, upward adjustments were “much less likely to improve accuracy than negative adjustments” (Fildes et al., 2009).
7. Track the cash-out date across forecasts, since its movement is the timing error, and show the board how far it moved over the last three. The free 13-week cash flow model below is a starting point; add a column for the naive forecast.
8. Treat a single-line five-year cash forecast as a plan until it comes with its range and the past errors it was built from.

Detail pays only as far out as a forecast beats the naive one

Cash flow forecasting is three forecasts sharing a spreadsheet: a dated view of the weeks, a working-capital model of the year and a range for the next raise. Score the flows rather than the balance, compare every horizon with the naive forecast, and draw anything beyond that crossover as a widening range, as the Bank of England began doing in 1996. A forecast that cannot beat last week’s actual adds work and no information, however detailed or sophisticated it is. In the hypothetical receipts forecast above the crossover came at about 8 weeks, and each line of a real forecast will reach it at its own distance.

Notes and sources

• Armstrong, J. S. (2001). Combining forecasts. In J. S. Armstrong (ed.), Principles of Forecasting: A Handbook for Researchers and Practitioners. Norwell, MA: Kluwer.
• Association for Financial Professionals (2026). AFP Survey: AI Priorities Rise Across Treasury Teams While AI-Related Challenges Grow (press release on the 2026 AFP Treasury Benchmarking Survey).
• Association for Financial Professionals. Top Treasury Tasks: Cash Management and Forecasting (Treasury in Practice Guide excerpt).
• Britton, E., Fisher, P. and Whitley, J. (1998). The Inflation Report projections: understanding the fan chart. Bank of England Quarterly Bulletin, February.
• Congressional Budget Office (2019). An Evaluation of CBO’s Past Deficit and Debt Projections.
• Congressional Budget Office (2023). The Accuracy of CBO’s Budget Projections for Fiscal Year 2023.
• CTMfile (2025). Treasury cash forecasting: rising expectations, growing complexity, AI’s promise (summary of Strategic Treasurer’s 2025 Cash Forecasting & Visibility Survey).
• Farshadfar, S. and Monem, R. (2013). Further evidence on the usefulness of direct method cash flow components for forecasting future cash flows. The International Journal of Accounting 48(1), 111–133.
• Fildes, R., Goodwin, P., Lawrence, M. and Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planning. International Journal of Forecasting 25(1), 3–23.
• Givoly, D., Hayn, C. K. and Lehavy, R. (2009). The quality of analysts’ cash flow forecasts. The Accounting Review 84(6), 1877–1911.
• Hyndman, R. J. and Athanasopoulos, G. Forecasting: Principles and Practice, 3rd ed., OTexts, sections 1.1, 5.2 and 5.8.
• Hyndman, R. J. and Koehler, A. B. (2006). Another look at measures of forecast accuracy. International Journal of Forecasting 22(4), 679–688.
• IASB. IAS 1 Presentation of Financial Statements and IAS 7 Statement of Cash Flows, as endorsed in the EU by Commission Regulation (EU) 2023/1803.
• M4 Competition organisers (2018). Evaluation and Ranks, Benchmarks and Evaluation code and M4-info dataset file. Mcompetitions/M4-methods, GitHub.
• M5 Competition organisers (2020). Scores and Ranks and M5 Competitors’ Guide. Mcompetitions/M5-methods, GitHub.
• Makridakis, S. and Petropoulos, F. (2020). The M4 competition: conclusions. International Journal of Forecasting 36(1), 224–227.
• 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), 54–74.
• NPR (2023). Yellen sets new deadline for Congress to raise debt limit: June 5. 26 May.
• Orpurt, S. F. and Zang, Y. (2009). Do direct cash flow disclosures help predict future operating cash flows and earnings? The Accounting Review 84(3), 893–935.
• treasuryXL (2022). Survey says: treasurers want more accurate cash forecasting (summary of Strategic Treasurer’s 2021 Cash Forecasting & Visibility Survey).
• US Department of the Treasury (2023). Statement by Secretary of the Treasury Janet L. Yellen on the bipartisan passage to suspend the debt limit, 1 June.
• Wikipedia. Forecast error; Makridakis Competitions.

The figures come from the standards, reports and papers above; the findings of Orpurt and Zang, Farshadfar and Monem, Givoly, Hayn and Lehavy, and Fildes and colleagues are taken from their abstracts. Competition results come from the official M4 and M5 score files. Numbers that carry no citation, including the counts of M4 entries and the M5 gaps, are derived from the figures above, and the $6M company and the company whose receipts forecast is scored against the naive one are hypothetical.