How to build a Cash Forecast: Six Steps and the Weekly Rebuild

Build it in six steps: fix the horizon and the bucket, list the lines that move cash, take the opening position from a reconciled record, drive every line from its own driver rather than from an accounting figure, roll it forward one week at a time, and measure the error separately at every horizon. At Nirjhar Industries Limited, invented, the sixth step decides whether anybody uses the other five.

A cash forecast is built, not derived, and the building has an order to it. Whether a forecast gets believed, what its error curve means and where the line falls between a week a treasury decides on and a week it only takes warning from are covered under forecast credibility, and that boundary is taken here as settled. The treasury doing the building is that of Nirjhar Industries Limited, an invented steel and alloys maker.

What are the six steps, and what does each one decide?

Every attempt at a cash forecast that stalls, stalls in the same place, and the reason is almost never the arithmetic. The reason is that somebody started at step four. The builder opened the receivable ledger and began working out what would be collected, before anybody had decided how far ahead the document looks, how fine its columns are, or which lines it is made of. Two days later they had beautiful detail on collections and no idea where to put it. The six steps run in a fixed order because each one consumes the output of the one before it, and the commonest way to waste a week is to build detail for a line that the line list would have marked for deletion.

The everyday version has the same shape at every scale. A household about to work out whether it can get to the end of the month does not begin by adding up the grocery bills. The household begins by deciding it is looking four weeks ahead in weekly slices, then lists what actually arrives and leaves, then checks what is in the account today, and only then works out what each item will be. Taken in the wrong order, the evening goes on pricing groceries for a month nobody has defined.

Why thirteen weeks, and why weekly buckets?

Step one settles two things at once and they are separate decisions. The first is the horizonHow far ahead a particular column is looking, which is what error has to be measured against., meaning how far out the last column reaches. The second is the bucketThe unit of time each column covers, which has to be short enough that the thing being looked for cannot hide inside one., meaning what one column covers. Nirjhar's treasury runs thirteen weeks in weekly buckets, and both halves of that are choices rather than rules.

Take the bucket first. The bucket is the one people get wrong, and the reasoning about it is exact. The bucket has to be shorter than the problem it is meant to reveal, or the problem hides inside a single column and the document reports nothing. A monthly column can show a comfortable month that contained eleven days with nothing in the account. Nobody is short of money for a quarter. A business is short of it on a Tuesday, and the column has to be short enough to have a Tuesday in it. In the other direction, daily buckets for thirteen weeks produce 91 columns nobody will read and a level of precision the drivers cannot support anyway.

Now the horizon. Thirteen weeks is one quarter and a day, so it reaches a quarter end from wherever it starts. A quarter end matters because a group's own reporting, its lenders and its counterparties all look at the same date. Thirteen weeks is also short enough that the far columns are still worth drawing at all. Push the horizon to twenty six weeks and the last columns carry an error nobody can act on, so the effort of building them buys a warning that a single line of prose could have given. Thirteen is Nirjhar's own choice, made by Girish Talwalkar, its group treasurer, and it is not a standard anybody sets.

Which lines go in, and how is the list known to be finished?

Step two is the shortest step and the one that decides whether the rest of the work is possible. Step two lists the things that move cash into or out of a bank account, and the list is closed: seven lines at Nirjhar, and every rupee that hits any of its 26 accounts belongs to exactly one of them. A line called other is not an eighth line. A line nobody can name is a line nobody can drive, check or improve, so a line called other is a report that step two is not finished.

The completeness test is worth stating as a test rather than a wish. A month of actual bank statement lines is taken and every one of them allocated to a named line. Nirjhar makes about 1,840 payments a month across roughly 20 working days, or about 92 payments on a working day, and 92 is a number one person can genuinely walk through. If everything lands, the list is closed. If something does not land, the missing line has been found. The honest response is to add it by name rather than to open a bucket called other and let unclassified items accumulate in it, and they always do.

Where does the opening position come from, and what if it is incomplete?

Step three is a single cell and it is the most load-bearing cell in the document. The opening positionThe cash the group actually holds at the start of the forecast, taken from a reconciled and complete record. is what the group actually holds right now, and at Nirjhar that is Rs 540 crore sitting across 26 bank accounts with 4 banks in 3 currencies. Every other number in the forecast is a movement away from that figure. An error in the opening position therefore does not stay where it is. The error propagates: present in week 1, and still present, unchanged and unmeasured, in week 13.

Two conditions have to hold, and they are different conditions. The first is that the record is reconciled, meaning the balance being read has been matched against the bank's own statement rather than taken from the group's ledger and hoped for. The second is that it is complete, meaning every account is in it. A position missing one account is not a slightly smaller position, it is an unknown one, and forecasting forward from an unknown starting point produces thirteen unknown weeks rather than one uncertain cell. The coverage question at Nirjhar is therefore answered at 26 of 26 and never at 25 of 26. Nobody can know which way the missing account leans, so the near miss is worth nothing.

Where the complete position comes from operationally, and what it takes to have it by ten in the morning rather than on the third working day, is a separate subject and is settled under the daily cash position. Step three only asserts what the construction needs: a reconciled figure covering every account, dated today.

How is each line actually driven, and why not from the profit and loss account?

Step four is where the work is, and it has one rule. Every line is built from its own driverThe underlying record or pattern a forecast line is built from, such as a receivable ledger or a loan schedule., meaning the record that actually determines when the cash moves. Collections come from the receivable ledger and the pattern of how that group's customers have historically paid against it. Payroll comes from the payroll run. Statutory payments come from the filing calendar. Interest and repayment come from the loan schedule. Building each line from its own driver is what is meant by the direct methodBuilding a cash forecast from expected receipts and payments rather than from an accounting profit figure., and the alternative is not a different technique so much as a different document.

An accounting line is the wrong input for three separate reasons, and each one alone is enough to disqualify it. A sale recognised is not a sale collected, so profit is not cash. Twelve months of twelve equal quarters do not divide into 52 weeks without leaving a remainder somewhere, so a month is not four weeks. And an accrual carries an amount without a date. The date is the one thing the forecast exists to supply. An accrual states that Rs 36 crore of interest belongs to this month. The accrual does not state which day the money leaves the account, and a treasury that does not know the day does not know anything useful.

The interest line makes the point with locked numbers. Nirjhar carries total debt of Rs 4,320 crore at an average contracted cost of 10.0 per cent, an interest cost of Rs 432 crore for the year, or Rs 36 crore a month. The annual figure divided by 52 is Rs 8.31 crore a week. Put into every column, that forecasts Rs 108.00 crore across thirteen weeks, and three monthly payments of Rs 36 crore total exactly the same. The two methods agree to the rupee on the block and disagree in every single week inside it.

The cost of that is exact. In each of the three payment weeks the even spread is short by Rs 36 crore less Rs 8.31 crore, being Rs 27.69 crore. In each of the other ten weeks it is over by Rs 8.31 crore. The absolute misses add to 3 times 27.69, being Rs 83.08 crore, plus 10 times 8.31, being another Rs 83.08 crore, so Rs 166.15 crore of error on a line whose total is right to the rupee. The error is 1.54 times the size of the line itself. The reconciliation is perfect and the document is useless, and the trap is exactly that pairing.

One warning belongs specifically to this group, and it is the most expensive confusion available in a treasury. Rs 432 crore appears twice inside Nirjhar and means two different things. Rs 432 crore is the interest cost for the year, at 10.0 per cent on Rs 4,320 crore of debt. The same Rs 432 crore is bucket MP1 of the group's debt maturity profile, the principal falling due inside year 1. A forecast that writes a bare Rs 432 crore beside a repayment schedule has merged a principal repayment with an interest charge, and the fix is to name the object in the cell every single time. Both belong in the interest and repayment line, on their own dates, as two things.

Scale checks come last, and they get mistaken for forecast lines. Nirjhar's revenue is Rs 10,800 crore a year, or Rs 900 crore a month and Rs 207.69 crore a week. The weekly figure is useful for sanity: if the collections line is forecasting Rs 600 crore in a single week, something is wrong. But it is a check and not an input. Collections come from the receivable ledger and the collection pattern, never from revenue divided by 52, and the difference between checking a number and generating one is the whole of step four.

What does rolling it forward weekly actually involve?

Step five is a cadence rather than an event. A rolling forecastA forecast rebuilt on a fixed cadence, dropping the week that has happened and adding a new one at the far end. is redone on a fixed rhythm, and at Nirjhar that rhythm is weekly. The loop has four moves, and the second one is the move that gets dropped when people are busy. Dropping it is precisely what converts a rebuildRedoing every remaining week from its drivers, as against copying the previous version forward and editing the front. into a copy.

One feature of rolling is rarely said out loud, and once it is seen the whole document reads differently. Because the forecast rolls weekly and reaches thirteen weeks out, one calendar week does not get forecast once. The same week gets forecast thirteen separate times, first at a horizon of thirteen weeks and last at a horizon of one, twelve rebuilds later. The document published on any given day is a snapshot of thirteen different weeks at thirteen different stages of being understood, and reading it as though every column were equally solid is the single commonest misreading of a rolling forecast.

How many times does one calendar week get forecast?

Follow one calendar week through, from the first time anybody puts a number against it to the week itself. Nirjhar's treasury has measured its own error by horizon over the last thirteen weeks and it holds four measurements: 1.4 per cent at week 1, 4.8 per cent at week 4, 9.6 per cent at week 8 and 16.2 per cent at week 13. The four readings are the group's own measurements. Every reading between them is drawn straight between the measured points and is not itself a measurement.

Read backwards, those readings are the life story of one week. Its thirteen successive sightings run 16.20, 14.88, 13.56, 12.24, 10.92, 9.60, 8.40, 7.20, 6.00, 4.80, 3.67, 2.53 and 1.40 per cent. The fall from first sight to the week itself is 16.20 less 1.40, being 14.80 points, or 91.4 per cent of where it started. Nothing about the week changed across those thirteen sightings except how close it was, and that alone took nine tenths of the uncertainty out of it.

The crossing is not what people expect. The error improves gradually, one reading at a time, but nothing about what the treasury may do with the week changes at all until the tenth sighting. Then it changes completely, and it changes because of a boundary somebody chose in advance rather than because of anything that happened in the business. Nine sightings of informative and not actionable, then four of actionable. Twelve rebuilds happen between the first sighting and the week itself, and only the tenth one alters what anybody is allowed to do.

The readings convert into rupees in one step, and the step is worth stating outright. The error is a percentage, so applying it to the group's Rs 540 crore of cash gives an expected miss of Rs 87.48 crore at horizon 13, Rs 51.84 crore at horizon 8, Rs 25.92 crore at horizon 4 and Rs 7.56 crore at horizon 1. Set those against the group's own minimum liquidity requirement, treasury limit TL5 at Rs 120 crore. The headroom is Rs 540 crore less Rs 120 crore, being Rs 420 crore. The largest miss is 87.48 over 420, being 20.8 per cent of the headroom. The curve does not cross the floor at any horizon in the thirteen weeks. A group is put through its liquidity floor by the flow, not by the measurement error on the balance.

How is the error measured, and what is published?

Step six is one row on the forecast and it decides whether the other thirty rows get used. The measurement is like for like measurementComparing each forecast against the actual for the same horizon, so week 1 is judged against week 1.: every week 1 forecast is compared with the actual for the week that followed it, every week 13 forecast is compared with the actual for the week thirteen weeks later, and the two are never mixed. Publishing one blended accuracy figure for the whole document mixes a 1.4 per cent reading with a 16.2 per cent one and produces a number that describes neither column, so a reader cannot tell which parts of the document to act on.

The published number is therefore not a headline accuracy but a row: the error measured at each horizon, sitting under the column it belongs to. The number a reader needs sits directly under the number they are reading, so the interpretation gets done correctly. Without that row every reader has to invent their own confidence, and they will not all invent the same one.

What goes wrong, and how does it show?

Who actually picks this up, and what do they do with it?

Three readers use the same thirteen columns for three different things, and the differences are worth naming because they explain why the error row matters so much.

Girish Talwalkar, the group treasurer, reads columns one to four as instructions and columns five to thirteen as weather. A week 1 number tells him whether to leave a placement where it is or bring it back. A week 11 number tells him to start a conversation with a bank, not to move money. The Rs 24 million United States dollar payable the group has committed to, booked at its own contracted rate of Rs 84.00 to the dollar and worth Rs 201.6 crore, falls due in 90 days, and 90 divided by 7 is 12.86, so it lands in week 13 of the view running today. The payable is 37.3 per cent of the group's whole Rs 540 crore of cash and 0.97 of one week's revenue at the Rs 207.69 crore weekly run rate. The exposure is already known and certain while the cash to settle it sits at the horizon where the forecast is least reliable, and those are two clocks rather than a contradiction. The forecast error does not apply to the payable itself, a contracted amount. The error applies instead to the cash position the payable will be settled out of.

Devendra Achar, head of treasury at Vindhya Commercial Bank Limited, invented, sits on the other side of the same facilities. He never sees Nirjhar's forecast, so he is not reading it as a document he can rely on. He is reading whether the borrower behaves like a group that has one: whether drawings on its committed lines arrive on notice or as a surprise on the day, and whether requests to move a repayment date come three weeks out or on the morning. A borrower that can state what it needs four weeks ahead is a borrower with a forecast, and that observation costs the lender nothing to make.

The mechanism does not care about scale, so the household version works the same way. A person on one salary who writes down what arrives and what leaves for the next four weeks, in weekly rows, and then rewrites it every Sunday, has built exactly the same instrument. The same thing turns up: the coming week is almost exactly right, the fourth week is a guess, and after a couple of months they know which of those two they can act on. The knowledge, and not the arithmetic, is the product.

How is it known that the whole thing is working?

Three tests, and none of them is about the size of the error.

The first is that the near horizon error falls over the first few months and then stays down. A falling near horizon error is the only evidence that the rebuild is a rebuild. The second is that somebody actually does something differently because of a week 1 number: leaves a placement in, brings one back, moves a payment run by two days. A forecast nobody has ever acted on is a report, and reports do not need error rows. The third test is the one people forget: a far week should start a conversation rather than a transaction, and if week 11 is generating instructions, the boundary between decision and warning has been forgotten and somebody is acting on a reading with three times the error of the one they think they have.

Notice what is not on that list. The size of the error is not a test. A group with a 1.4 per cent week 1 error that publishes it and acts on it is in far better shape than a group with a claimed 0.5 per cent error nobody measured and nobody uses.