Your Cash Forecast Is Only as Good as the Dates in It

The forecast said the account would not go below four thousand. On the eighteenth it went below zero, and nothing in the forecast had been wrong about amounts. Every invoice was the right size. One customer simply paid on day forty one instead of day thirty, and a supplier direct debit that always lands on the twenty fifth landed on the twenty second because the twenty fifth was a Sunday.

This is what cash flow forecasting is actually about, and it is not what the tools are sold on. Prediction accuracy gets the marketing. Timing produces the overdraft.

What does a cash forecast actually get wrong?

The dates, almost always. Amounts are knowable because most of them are already contracted, invoiced or recurring, while dates depend on other people's behaviour.

Break a single week down and the asymmetry is obvious. The committed payments are near certain in both size and timing. The confirmed receipts are certain in size and uncertain in date. And the discretionary spending is the part you control, which means it is the lever you will actually pull when the forecast turns out wrong.

Notice that only one of those four boxes is a forecasting problem in the statistical sense. Committed payments are a list. Discretionary spend is a decision. Confirmed receipts are known amounts. Timing risk is the only genuine uncertainty, and it is driven by customer payment behaviour, bank processing days and your own invoicing discipline. A model trained on your sales history has very little to say about any of those.

What a Walmart competition tells a corner shop

The largest open test of retail forecasting ever run produced a result that should change how a small business shops for these tools, and it is not the result the tools quote.

The M5 competition used real Walmart sales data at daily frequency across a hierarchy of products and stores. The introduction to the M5 special issue records nearly 6,400 teams across both tracks, 5,507 of them in the accuracy track, competing for 100,000 dollars. Tree based methods dominated the top of the table, with gradient boosting used by practically every leading entry.

Then comes the line that matters. Only 7.5 percent of teams managed to beat an extremely simple benchmark. Thousands of skilled people, a large prize, clean data from one of the most sophisticated retailers on earth, and fewer than one in thirteen improved on a basic method. That is not an argument against forecasting. It is an argument against assuming a sophisticated model is an upgrade.

There is a second finding underneath it. The competition deliberately included intermittent series, meaning products that sell in ones and twos with long gaps of zero. Forecasting those requires predicting when a sale happens as well as how many, and that is where the advantage of complex methods narrows sharply. A shop with 200 products, most of which sell a handful of units a week, is an intermittent demand problem wearing a retail costume.

Does a better model help a small shop?

Usually less than better inputs do, and the reason is structural rather than a matter of effort. Your series are short, sparse and full of events the data cannot explain.

Consider what a model would need to learn to be useful. It needs enough history to separate seasonality from noise, which for weekly patterns means years rather than months. It needs your promotions labelled, because an unlabelled promotion teaches it that demand randomly tripled. It needs to know that you were closed for a fortnight in August. Very few small businesses have that data, and assembling it is more work than the forecast is worth.

The research community has been making a version of this argument for a while. A paper on fast and frugal retail forecasting questions the assumption that large families of complex models are needed at all, and shows a deliberately reduced set performing well while costing far less to run. For a business where the forecast exists to answer one question, whether there is enough money in three weeks, that trade is obviously correct.

The pattern in that table is that automation earns its place where the work is repetitive arithmetic on data you already have. It stops earning its place at the exact moment judgement enters, which is the same boundary that shows up in every other back office task, and it is the same reason our piece on reordering stock when you have thin sales history lands on a similar answer for inventory.

Which numbers should the AI touch?

The ones that repeat. A tool reading your bank feed can extend recurring lines forward with high reliability and almost no risk of an interesting mistake.

That covers more ground than it sounds. Rent, salaries, loan repayments, insurance, software subscriptions, regular supplier orders and tax payments on known dates are a large share of a small firm's outflow, and projecting them is genuinely tedious by hand. Getting that half automated frees the attention for the half that needs you.

Where it should not be trusted is anything requiring a view about a specific counterparty. A model that learns your average customer pays in 34 days will quietly apply 34 days to the customer who has never once paid before day 60. Averages hide exactly the concentration risk that sinks small businesses, and a shop with one customer worth 30 percent of revenue does not have an average, it has a dependency. Treat that customer as a separate line with its own assumption, and if your business has meaningful trade credit exposure the checks in our guide to running a trade credit check on business customers belong upstream of the forecast.

What does the law say about when you get paid?

More than most owners realise, and the default matters because it is what applies when nobody wrote a term down. In the UK, if no payment date was agreed, the government's guidance on late commercial payments and debt recovery states that payment is late 30 days after the customer receives the invoice or after you deliver, whichever happens later.

Two things follow for a forecast. The clock often starts on delivery rather than on invoice, so a slow invoicing habit does not merely delay payment, it delays the point at which the payment is even late. And the guidance notes that a longer period than 60 days can be agreed for business transactions only where it is fair to both sides, which means unusually long terms pushed by a large buyer are not automatically binding.

The forecasting consequence is that your payment terms are an assumption you can change, not weather you have to endure. Shortening the gap between delivery and invoice is the cheapest cash flow improvement available to most small businesses, and it needs no model at all.

A routine that survives contact with a real month

Five steps, once a month, and the fourth one is the step almost everybody skips.

Stressing the forecast is what turns it from a document into a decision tool. Delay your single largest expected receipt by thirty days and look at what happens. If the answer is that nothing much happens, you have more resilience than you thought and can stop worrying. If the answer is that you breach your overdraft in week three, you have learned it in advance, which is the entire purpose.

The fifth step, comparing last month's forecast against what actually happened, is the only way the thing ever gets better. It also tells you something no vendor will: whether the tool is adding accuracy or just adding confidence. Two months of that comparison is usually enough to know.

Why thirteen weeks rather than a year?

Because it is the longest horizon over which the dates are still mostly knowable, and the shortest over which you can still do something about a problem. Anything beyond it stops being a cash forecast and becomes a budget.

The thirteen week window is borrowed from restructuring work, where it is the standard instrument for a business under pressure, and the logic transfers cleanly to a small shop in good health. Inside thirteen weeks you know your rent dates, your payroll dates, your tax dates and most of your supplier orders. You know which invoices are outstanding and roughly who pays late. The unknowns are bounded.

Stretch to twelve months and almost every line becomes an assumption. Sales are projected rather than invoiced, costs are estimated, and the compounding of small errors makes week 40 meaningless. Worse, a long horizon invites a comforting shape, because a year of projected growth always absorbs a bad fortnight. The bad fortnight is the thing that closes businesses.

There is a practical benefit too. Thirteen weeks fits on one screen at weekly granularity, which means you will actually look at it. A forecast nobody opens has an accuracy of zero regardless of the model behind it.

What do you do when the forecast shows a gap?

You work the levers in order of how fast they move money, and that order is almost the reverse of how people instinctively approach it. The instinct is to cut spending, which is the slowest lever in the list.

Collection comes first, because the money already belongs to you. Chasing the three largest overdue invoices moves more cash in a week than a month of expense trimming, and a forecast that ages your receivables tells you exactly which three. This is the single highest return use of an automated ledger, and it works because the arithmetic is boring rather than clever.

Timing comes second. Moving a discretionary payment by two weeks costs nothing if the supplier relationship tolerates it and you ask rather than simply paying late. Paying late without asking is a different act with a different price, since it damages the one thing that gets you terms in the first place.

Inflow acceleration is third. Deposits on large orders, part payment on delivery, or a small discount for early settlement all pull receipts forward. Each has a real cost, so they belong below collection and timing rather than above them.

Financing is fourth, and the Federal Reserve numbers are the reason it sits this low. Applying takes time you may not have, and full approval is not the common outcome. A facility arranged before you need it is a different proposition from one sought during a squeeze, which is the argument for opening the conversation in a quiet month.

Cost reduction is last, not because it does not matter but because it is slow. Most costs in a small business are contracted, and the ones you can cut this week are the ones that were already small. Cutting is a structural fix for a structural problem, not a response to a three week gap.

What ties the order together is that the first three levers all depend on information you already hold and a willingness to make uncomfortable phone calls. No model is involved. The forecast's job was only to tell you which week to be worried about and how large the hole is, and once it has done that the work is human.

What is actually moving small business cash right now?

Costs, more than demand, according to the most recent national data. That matters because a forecast built only on sales projections is watching the wrong side of the ledger.

The Federal Reserve's 2026 report on employer firms, drawing on 6,525 responses collected between 3 September and 14 November 2025, found rising costs of goods, services and wages to be the most common financial challenge reported. More than 40 percent of firms identified tariff related cost increases as a problem, with retail at 69 percent among the hardest hit sectors. On financing, of the 60 percent of firms that applied, only 42 percent received the full amount sought.

Read that last figure next to your forecast. A plan that resolves a projected shortfall by assuming credit will arrive is making an assumption that failed for well over half of the firms that tested it. The honest version of a cash forecast includes what you will do if the funding does not land, and that contingency is a sentence you write, not an output a model produces.

The part worth keeping

Cash flow forecasting has been improved by these tools in a narrow and real way. The bookkeeping half, extending known lines and ageing unpaid invoices, has gone from an afternoon to a click, and that is worth having.

The part that decides whether you run out of money has not been automated and shows no sign of being. It consists of the dates other people choose, the decisions you have not made yet, and the events that have no precedent in your history. A competition on the cleanest retail data available could not get most expert teams past a simple benchmark, which should end the idea that a smarter model is what stands between a small shop and financial visibility.

What stands between them is usually a list of committed payments that nobody has written down, and a habit of checking it against reality once a month. Owning your own sales data in a system you control makes that list easy to assemble, which is one of the quieter arguments for running your shop on infrastructure you hold rather than rent, and our store builder for online sellers is built on that principle. The forecast itself can stay in a spreadsheet. It is the discipline that has to be somewhere reliable.