Start a 13-week cash forecast today and it ends on 19 January. That's two and a half weeks before Chinese New Year.
For a Singapore business with a December year-end, the hardest stretch of the cash year sits just past the textbook horizon. Bonuses and the 13th month go out in December. CPF on them falls due on 14 January. Chinese New Year lands on 6 February, with a public holiday on the 8th, and for a fortnight either side your customers' finance teams are paying very little to anyone. Then, on 15 February, the GST for October to December leaves your account by GIRO. Four big outflows in ten weeks, while the receipts slow down.
A 13-week forecast is the right tool for running the business week to week. It's the wrong tool for seeing that stretch coming in October. So this playbook builds two forecasts that meet in the middle: thirteen weeks in detail, and the thirty-nine weeks after them in outline. Together they're a year of cash you can actually see.
Last week I wrote about building next year's budget in two weeks. The budget says what the business plans to earn. This is the other half of the question: whether you'll have the cash to get there. As before, a model like Claude does the legwork, and the calls stay with you.
Before You Start: The Inputs
The room is the same as for the budget: a commercial Claude account, one Project for the forecast, and the spreadsheet as the source of truth. The inputs are different. Cash forecasts are built from what actually moved, not from what was booked.
- Twelve months of bank statements from every account the business uses, exported as CSV files. Not a summary: every line.
- Your invoice and bill history for the same twelve months, with the date each invoice was raised, its due date and the date it was paid. Most accounting systems will export this in one report.
- Today's receivables and payables ageing, and your suppliers' payment terms.
- The fixed calendar: payroll dates, loan schedules, your GST quarters, rent, insurance and licence renewals.
- Next year's budget, by month. If you built it with the two-week method, it's ready.
- Two or three years of monthly history, if you have it. More on why in step 5.
Claude reads exported files. It never gets a bank login, and it never gets the right to move money. More on that below.
Part One: The Next 13 Weeks, Week by Week
The first thirteen weeks are built directly: actual receipts and actual payments, invoice by invoice and bill by bill. This is the part you'll run the business on.
1. Sort out your inflows and outflows. Ask Claude to tag every line in the bank statements: receipts by customer, payroll, CPF, GST, rent, each supplier, loan repayments, owner's drawings, and transfers between your own accounts, which cancel out and come off. A year of bank lines is a day's work by hand. It's a short first pass for a model.
Your call: settle the lines it can't place. Claude will flag them: a payment with a vague reference, a receipt that matches two invoices, a transfer to a director. Only you know what they were.
2. Learn how your customers actually pay. This is the step most forecasts skip, and it's the one that matters most. Have Claude match every invoice to the receipt that paid it, then work out each customer's real days to pay against their terms. A customer on 30 days who pays in 52 is a 52-day customer. Ask it to show the pattern by month as well, because some customers pay noticeably slower in December and around Chinese New Year. Then have it place every open invoice in the week it's likely to arrive, based on that customer's behaviour rather than the due date.
Your call: decide which receipts you believe. Claude has the history. You have the relationship: you know the client whose finance manager just left, the one that's waiting on its own customer, and the one that always pays the week before its year-end. Move those receipts by hand and note why. A forecast built on payment terms is a wish list.
3. Map the payment calendar. Ask Claude to build the payment calendar week by week from the bank history and your fixed calendar: payroll, CPF by the 14th of the following month, GST (more on its timing in step 5), rent, loan instalments, supplier payments at their real timing, the 13th month and bonuses, and the Chinese New Year costs that never appear in a budget line called "Chinese New Year". Have it flag anything the bank history shows that the calendar doesn't, such as an annual subscription that renews in March.
Your call: decide who gets paid, and when. Which suppliers get paid on the due date whatever happens, because the relationship is worth more than a fortnight's float. Which ones could wait a week if they had to. Payroll, CPF and GST are not on the list of things that wait.
4. Set the sleep-easy floor. The floor is your sleep-easy sum: the balance that, when you see it in the bank, tells you everything is fine. Below it, you're allowed to panic. Ask Claude what a sensible floor looks like for your business, for example one month's payroll and CPF, plus the next GST payment, plus your largest regular supplier payment. Then ask how often, over the last year, you'd have dropped below each version of it.
Test the floor against weekly balances, not month-end ones. A month-end figure can hide a trough in the middle of the month. If most of your receipts land in weeks four and five while payroll, rent and suppliers go out in weeks one to three, you can be below the floor, or overdrawn, by week three of a month that ends comfortably in the black. That's why the first thirteen weeks are weekly, and why, in the monthly part of the forecast, you should have Claude phase each month by week using your real receipt pattern, so the mid-month dip stays visible.
Your call: choose the number. How much risk you're willing to carry is an owner's decision. Once it's set, the forecast has a line to measure every week against.
Part Two: The 39 Weeks After, Month by Month
Past week thirteen, invoice-by-invoice forecasting stops being honest. You haven't raised those invoices yet. So the rest of the year is built from the budget, by month. It's less precise and it doesn't need to be. Its job is to show you the shape of the year: where the low points are, roughly how deep, and why.
5. Build the long-range forecast. Start by giving Claude more history than you think it needs: two or three years of monthly receipts and payments, not just last year. One year shows you what happened. Three years show you what keeps happening. Ask Claude to chart the seasonality: which months collections dip, which customers bunch their payments around their own year-end, how long receipts take to recover after the holidays. Chinese New Year moves between late January and mid-February, so ask it to compare the weeks around the holiday rather than the same calendar month each year. The patterns it finds can surprise you, even in a business you've run for a decade.
Then give it the monthly budget and ask it to turn profit into cash. Revenue becomes receipts at your real collection speed from step 2. Costs become payments at your real payment speed.
Then comes the part almost every owner misses.
The cash your P&L never shows you. The profit and loss records what you earned and what it cost. It leaves out a long list of things that move real money. Ask Claude to find every one of them in your bank history and put each in the forecast on the date the money moves:
- GST. What you collect on sales isn't yours, and it sits on the balance sheet, not in the P&L. It leaves your account in a lump, weeks after the quarter ends. The schematic below shows how.
- Loan repayments. Only the interest shows up as a cost. The principal comes out of the bank every month regardless.
- Equipment and fit-out. The P&L spreads the cost over several years as depreciation. The bank pays for it in one go.
- Corporate tax. The P&L shows a provision. The bank sees instalments, on a timetable you partly control (more below).
- Deposits. Rental deposits you pay, deposits clients pay you up front, and the refunds of both.
- Dividends and director's loans. Money in and out between you and the company.
To put GST in your forecast, take each quarter's net amount (the GST on your sales minus the GST on your purchases) and place it as a single payment on the 15th of the second month after the quarter ends. Ask Claude to estimate the running total from your sales and purchases as the quarter goes, so the payment is never a surprise. For a December year-end, notice where the fourth quarter's payment lands: mid-February, a week after Chinese New Year.
Corporate tax works the same way: the timing is partly yours to choose. If your company pays by GIRO, IRAS gives more interest-free instalments the sooner you file your estimated chargeable income after the year-end: ten if you file within one month, eight within two, six within three, and none after that. For a December year-end, filing in January rather than March spreads the same tax bill across ten months instead of six. Claude can show you what that does to your low point.
Your call: move the timing, not just the totals. This is where the forecast earns its keep. Run last year's plan through it first, unchanged: the bonus in December, the equipment bought in February, the big contract renewing in March. That's the status quo, and the forecast will show you exactly where it breaks. Then you can do something about it while there's still time. Pay the bonus in January, after the biggest client's payment clears. Buy the equipment in April, once the post-holiday receipts are in. Ask the client to renew a month early, with a small discount for paying up front. Claude will lay out the options and what each does to the low point. Choosing between them is running the business.
6. Connect both forecasts. The two forecasts meet at week thirteen, and the join is where errors hide. Ask Claude to run the monthly method over the first three months as well, and compare it with the weekly forecast. If they agree, the long view is calibrated. If they disagree by much, one of them is wrong, and Claude should list the likely reasons: a payment pattern the budget doesn't know about, a cost the bank shows that the budget doesn't, a timing difference.
Your call: decide which one is wrong. Usually it's the monthly one, because the weekly one is built from what actually happened. But not always. If the weekly forecast quietly assumes your biggest customer pays on time for the first time in a year, the budget is the one telling the truth.
7. Plan for contingencies. In the army, every plan made during Battle Procedure goes through a round of contingency planning before it's issued. The reason is simple: assumptions start to unravel the moment a plan is fixed, and a commander needs to know what could go wrong, and what they'll do about it, before it does. A cash forecast deserves the same discipline.
Ask Claude to draft the three to five most realistic risks for your business, using your own history: the customer who has paid late before, the contract that's up for renewal, the receipts that slipped around last Chinese New Year, the supplier who has raised prices twice. For each, have it propose what you'd do if it happened.
Then ask Claude to build the risks into your spreadsheet as switches: one cell per risk, on or off, with its effect on the timing and size of the cash flows, and a chart that redraws as you flip them. Try the combinations. Your biggest customer paying a month late, on its own. That, plus the contract not renewing. All of it at once.
Within an hour, you'll have a feel for your best, base and worst cases, built on your own instinct for which risks are real rather than on a formula's.
Your call: pick the risks and the responses. Decide now what you'll do in each case, and at what point on the chart: chase earlier, delay a hire, defer a purchase, or speak to the bank. Speak to the bank early. Banks lend more readily to a business that comes in with a forecast showing a dip in February than to one that rings them in the second week of it.
The Weekly Refresh
A cash forecast is only worth something if it's current. Once it's built, keeping it current takes less than an hour a week, and most of that is legwork.
- Roll it forward. Load last week's bank statement. Claude compares what came in and went out with what the forecast said, drops the week that's passed and adds a new week thirteen.
- Explain the misses. For every difference, Claude proposes whether it's timing (the money arrives next week) or permanent (it isn't coming). You confirm which. Timing misses move. Permanent ones come off.
- Watch the drift. Have Claude track each customer's days to pay over time. A customer who has paid a few days slower three months running is telling you something before the overdue report does.
- Refresh the long view monthly, after the month-end close, not weekly. The budget doesn't change every Monday, and neither should the thirty-nine weeks built on it.
- Build it to learn. Claude doesn't remember last month's lessons on its own. Each conversation starts from what it's given. A forecast gets better over time only if the system around it is designed to learn, and that takes three files:
- A variance log. Every miss, its cause, and whether it was timing or permanent, added by Claude during each refresh.
- A rules file. The floor, the payment policy, each customer's real days to pay, every exception and the date it was agreed. In a Claude Project, this lives in the Project's instructions. In Claude Code, it's the
CLAUDE.mdfile in the forecast's folder. Either way, it's read at the start of every session. - A monthly review. Once a month, ask Claude to read the variance log and propose changes to the rules: this customer now pays in 58 days, not 45; this subscription needs adding to the calendar; this risk has happened twice and belongs in the base case. You approve each change before it goes in. Over time, the misses get smaller. That's the system learning, and you can measure it: track how far off week one's forecast was, month by month, and watch it narrow.
I covered how to keep a long piece of work on track inside a single conversation in the budget playbook. The three files above are what carry it from one week to the next.
What I Don't Let Claude Decide
The rule is the same as for the budget: Claude can inform any decision, but it doesn't make one. Six calls stay with you.
1. Who gets paid, and when. Claude can rank every supplier by terms, late-payment penalties and how hard they'd be to replace. Deciding who gets paid late is a decision about relationships, and if someone has to wait, you're the one who calls them before the due date, not after it.
2. Whether a customer will pay. Claude has the payment history and it's good at spotting drift. It doesn't know that a client has just lost its biggest account, or that a new finance director is reviewing every supplier. You do. Your view overrides the pattern, and the note explaining why goes into the rules file.
3. Whether to borrow, and when. Claude can work out how much headroom you have on an overdraft or credit line, what the interest costs at different levels of drawing, and what a lender is likely to ask for. Whether you draw on it, and when you talk to the bank, stays with you.
4. Where the floor sits. Claude can show you the cost of holding more cash and the risk of holding less. How much risk the business carries is the owner's call.
5. What gets cut if the low point is too low. Claude can list the levers and what each one is worth: deferring a hire, delaying a purchase, offering a client a discount to pay early. Which lever you pull is a decision about the business you want to have at the end of the year.
6. Moving money. Claude reads exported statements. It never logs in to the bank, never holds a payment token and never initiates a transfer. Every payment is approved by a person, under the same controls you'd apply to any member of the team. That's not a limitation of the tool. It's a rule of the business.
Where This Breaks
- Unreconciled books. If the bank and the ledger don't agree, step 2 will match invoices to the wrong receipts very confidently. Get the month-end close reliable first.
- Lumpy revenue. If one milestone payment is a third of the quarter, don't let any pattern forecast it. Give it its own line, your own date and a contingency where it slips.
- Several currencies. Forecast each currency on its own and convert at a rate you choose and write down. Don't let the exchange rate change quietly from one week to the next.
Cash Doesn't Surprise You. It Warns You.
Cash is the lifeblood of a company, and the forecast is its blood pressure monitor.
Nobody waits for a heart attack to check their blood pressure. You take regular readings, you know your healthy range, and you watch the trend, because the trend shows trouble long before you feel it. A bank balance checked once a day is a single reading. It tells you how you are this morning. A forecast checked every week is the chart, and the chart shows you February's spike while it's still a number on a screen, with time to change your diet.
Almost every cash crunch gives that warning weeks before it arrives: a receipt that's later than last time, a quarter where GST and bonuses overlap, a holiday that falls where it always does. The forecast's job is to put the warning in front of you while there's still time to act on it.
Thirteen weeks shows you the next payroll. Fifty-two shows you next February. Give the legwork away, and keep the calls.
Tested on Claude, October 2026.