Key takeaways
- Use the direct method. An indirect forecast built from projected profit cannot tell you which day the bank balance goes negative, which is the entire purpose of the model.
- Build the collections line from the open invoice listing using behavioural payment lags per customer, then apply a single realisation haircut at aggregate level rather than invoice by invoice.
- Report available headroom, not just cash: undrawn facility less any drawing power constraint, alongside the running position against the tightest covenant.
- Score last week's forecast against actuals and classify each variance as timing, quantum, omission or behavioural. The pattern shows where the model is structurally wrong.
- Keep the model to twenty-five to thirty-five lines with a weekly update under three hours. Complexity, not inaccuracy, is what kills these forecasts.
Most 13-week cash flow forecasts die in the fourth week. The first version is built during a liquidity scare, circulated with some ceremony, and then quietly abandoned once the immediate pressure eases — because updating it takes a day and a half, nobody reconciles last week’s forecast to what actually happened, and the numbers stop being believed.
A forecast that survives has three properties: it is built directly from open items rather than derived from the P&L, it is updated on a fixed weekly rhythm that takes under three hours, and every week its previous version is scored against actuals. Those three things, not modelling sophistication, are what make management use it.
Why thirteen weeks, and why the direct method
Thirteen weeks is one quarter. It is long enough to see a covenant test, a tax payment cycle and a seasonal trough coming, and short enough that receipts and payments can be forecast from identifiable items rather than from statistical assumptions. Beyond roughly a quarter, the item-level basis erodes and the forecast becomes an extension of the budget.
The distinction that matters more is direct versus indirect. An indirect forecast starts from projected profit and adjusts for non-cash items and working capital movements. It is the right tool for annual planning and for the statutory cash flow statement. It is close to useless for liquidity management, because it cannot tell you which Thursday the bank balance goes negative.
A direct forecast starts from the bank account and works forward: every expected receipt and every expected payment, by week, sourced from open invoices, contractual commitments and the statutory calendar. It ties to a bank balance, not to an accounting profit, and it is verifiable against a bank statement every Monday morning.
The structure
Keep the line count disciplined. Between twenty-five and thirty-five lines is usually enough; models with a hundred lines are not updated. Group them so that a reader can see, in one screen, where the cash comes from and what it is committed to.
| Block | Lines | Basis of estimate |
|---|---|---|
| Opening balance | Bank balances by account, less unpresented items | Bank statement, actual |
| Operating receipts | Collections from trade receivables, split by top customers and the tail; advances; cash and card sales | Open invoice listing with expected collection date |
| Other receipts | GST refunds, export incentives, insurance claims, interest, asset disposals | Filing and claim status, expected credit date |
| Operating payments | Vendor payments by category, payroll, rent, utilities, freight, contract labour | Open payables with due date, payroll register, contracts |
| Statutory payments | GST, TDS, advance tax, provident fund, ESI, professional tax, customs duty | Statutory due-date calendar and computed liability |
| Financing | Term loan principal and interest, working capital facility drawdown and repayment, lease rentals, dividends | Loan amortisation schedules, sanction terms |
| Capital and one-off | Capex commitments, retention releases, legal settlements, transaction costs | Purchase orders, agreements, board approvals |
| Closing balance | Closing bank balance; headroom against sanctioned limits; covenant metrics | Computed |
Two lines are non-negotiable at the bottom: available headroom — cash plus undrawn facility minus any drawing power constraint from the borrowing base — and, where relevant, the running position against the tightest financial covenant. Cash on hand alone understates or overstates flexibility depending on facility structure.
Where the data comes from
The receipts line
Collections are the hardest line to get right and the one that determines whether the model is credible. Build it from the open receivables listing at invoice level, and assign each invoice an expected collection week rather than its contractual due week. The two differ, and the difference is exactly what the model exists to capture.
A workable approach: take the last six to twelve months of settled invoices and compute, per customer or per customer segment, the median days from invoice date to receipt. Use that behavioural lag for the tail. For the top customers that make up the majority of the balance, override the statistical lag with information from the collections team — remittance advices received, payment runs confirmed, disputes open. Then apply a haircut to the aggregate, not to individual lines; a single realisation percentage applied to the total forecast collection is easier to calibrate and easier to explain than dozens of invoice-level probabilities.
Disputed and deduction-laden invoices should be shown separately. A receivable that is in dispute is not a timing problem, it is a revenue problem, and hiding it inside a collection lag obscures it.
The disbursements lines
Payables split cleanly into three behaviours. Contractual and dated items — loan instalments, lease rentals, statutory dues, payroll — go in at their exact date and amount, and should not be smoothed. Committed but flexible items — approved vendor invoices where the business controls the payment run — go in at the intended payment date, which makes the model a decision tool rather than a prediction. Uncommitted run-rate items — consumables, travel, small recurring spend — can be forecast on a rolling average with a single line each.
The most common omission is the statutory calendar. In an Indian entity a typical month carries GST payment and returns, TDS deposit and quarterly statements, provident fund and ESI remittances, and, in the relevant months, advance tax instalments. These cluster and they are not negotiable. A forecast that treats them as an average monthly outflow will misstate the intra-month trough.
The weekly rhythm
Fix the cadence and defend it. A workable pattern:
- Monday morning. Roll the model forward one week. Week one drops off; a new week thirteen is added. Import the actual bank position and the refreshed AR and AP open-item listings.
- Monday midday. Populate actuals for the week just closed against the forecast made for it, and compute variance by line.
- Monday afternoon. Collections and procurement review the top variances and the coming two weeks. Payment run priorities are set here, not by email later in the week.
- Tuesday. One-page output to the CEO and, where relevant, to the lender: closing balance by week, minimum balance in the period, headroom, the three largest changes since last week, and the actions taken.
The rolling discipline matters more than the accuracy of any single version. A forecast refreshed weekly with modest error beats a quarterly forecast built with great care, because decisions get made weekly.
Variance analysis that teaches the model something
Scoring the previous forecast is the step teams skip, and it is the step that makes the model improve. Classify every material variance into one of four causes and record it. Over eight to twelve weeks the pattern tells you where the model is structurally wrong rather than merely imprecise.
| Variance type | What it means | Fix |
|---|---|---|
| Timing | Right amount, wrong week; nets to zero over two or three weeks | Recalibrate the collection lag or the payment-run assumption for that counterparty |
| Quantum | Right week, wrong amount | Check for deductions, short payments, credit notes, rate or volume assumptions |
| Omission | Item was not in the forecast at all | Fix the source feed — usually a commitment made outside the purchase order or approval process |
| Behavioural | Forecast changed the outcome; a payment was held back because the model showed a trough | Record as a decision, not an error; it is the model working |
A practical accuracy target for a mature model is weekly net cash flow within a defined tolerance of forecast — many teams settle on five to ten per cent of gross receipts — with cumulative 13-week closing balance error materially tighter, because timing errors offset over the horizon. Set the target explicitly and report against it; an unmeasured forecast drifts.
How it changes decisions
The forecast earns its place when it starts altering behaviour rather than describing it.
- Payment sequencing. Payment runs are set against a visible trough, so discount opportunities are taken where headroom allows and non-critical payments deferred deliberately rather than by omission.
- Collections focus. The collections team works the invoices that move the trough week, not simply the oldest balances.
- Facility management. Drawdowns are planned rather than reactive, and interest cost falls when the model shows that a drawdown can be delayed by nine days.
- Capex and hiring timing. Committing spend is a dated decision with a visible consequence on the minimum balance.
- Lender and investor conversations. A borrower who arrives with a scored, rolling 13-week forecast and a variance history negotiates from a different position than one who arrives with a request.
- Covenant management. Breaches become visible weeks in advance, when there are still options.
Failure modes to design against
- Building it from the budget. The budget is a target; the forecast is an expectation. Merging them produces a document that is neither.
- One owner, no backup. If the model lives in one analyst’s file, it stops when they take leave.
- No link to source systems. Manual re-keying of AR and AP listings guarantees the three-hour update becomes a two-day update and then stops.
- Optimism in the collections line. The fastest way to lose management’s trust is a receipts forecast that is systematically high. Bias it slightly conservative and let the surprises be favourable.
- No scenario view. A single line is a point estimate. Two additional cases — a downside where the largest customer pays thirty days late, and one where a facility is not renewed — take twenty minutes and change the conversation.
- Reconciliation gaps. If the model’s opening balance does not tie to the bank statement every week, nothing downstream is credible.
Artham Fintech builds and runs rolling 13-week cash flow models as part of virtual CFO and treasury support engagements, including the source-data plumbing from receivables and payables systems and the weekly variance discipline that keeps them alive. If your forecast has stalled or was never built, we are happy to look at what it would take in your environment.