The Cash Forecast: Building One That Is Actually Used

A cash forecast that gets used states a horizon beside every number. The error grows with the horizon, so the same document is accurate enough to decide on in its first four weeks and only accurate enough to warn with after that. Publishing that measured error at each horizon, rather than hiding it, is what turns a forecast people ignore into one they act on.

Somewhere in almost every business there is a spreadsheet nobody argues with and nobody acts on. The spreadsheet arrives on a Monday, it has a number at the bottom, and everyone in the room has quietly decided how much of it to believe without saying so out loud. The silent discount is the whole problem. The forecast has not failed a test. Nobody set it one. The forecast has simply never told anybody how wrong it usually is, so every reader supplies a private guess, and private guesses are always harsher than a measurement.

Nirjhar Industries Limited, an invented steel and alloys maker, publishes one such document from its treasury every week. Girish Talwalkar is the group treasurer. The group holds Rs 540 crore of cash across 26 bank accounts with 4 banks in 3 currencies, and it makes about 1,840 payments a month. Nirjhar is also the largest single-name exposure on the books of Vindhya Commercial Bank Limited, also invented, where it is counterparty C1, so the same group appears in this material twice, once as a borrower being measured and once as a treasury doing the measuring. Every rupee, rate, count and limit in this guide belongs to that invented group and to nothing else.

Two things are assumed here rather than rebuilt. The first is that a treasury's first job is knowing the group can pay, the point at which this subject area starts. The second is that a complete opening positionThe cash the group actually holds at the moment the forecast starts, which has to be known before anything can be projected forward from it. exists before anything can be projected from it. At Nirjhar the group's cash position is known by 10 am rather than by the third working day, and that single operational fact is what makes the near end of this forecast worth anything at all. How that position is assembled is settled separately in this sequence and is not reopened here.

Why does almost nobody act on a cash forecast?

A treasurer asked why the finance director ignores the forecast will say that the business does not understand cash. The finance director, asked the same question, will say that the forecast has been wrong before. Both are describing the same missing thing, and it is not accuracy. A bare number invites exactly one response, and the response is disbelief. A forecast published as a single number with nothing attached to it therefore gets ignored.

The shape of this is clearest away from finance. Somebody says they will arrive in twenty minutes. The speaker is known, and so is roughly how their twenty minutes usually behaves, so a listener knows instantly what to do with it. Somebody else says they will arrive on the fourteenth of next month at 4.20 in the afternoon. The second statement is more precise and far less useful, and nobody would plan a day around it the way they would plan around the first. Nothing about the speaker changed. The distance changed, and the distance is doing all the work.

Distance is the entire subject here. A cash number for next Tuesday and a cash number for the Tuesday thirteen weeks after it are not two readings of the same quality. The two are different kinds of statement wearing the same font. A document that prints them in one column, with the same number of decimal places and no other mark, has told the reader something untrue by omission. The reader knows it, and the document sits unread.

What is a rolling thirteen week forecast, and why thirteen weeks?

Nirjhar's treasury runs a rolling forecastA forecast whose horizon moves forward every time it is rebuilt, so it never runs out of weeks and never stops halfway through a quarter. of thirteen weeks, rebuilt every week. Two words are carrying weight there and both are worth slowing down for.

Thirteen weeks is a quarter, near enough, and a quarter is the natural unit for the things a treasury is trying to see coming: a tax payment, a loan instalment, a large customer settlement, a seasonal build in stock. Shorter than that and the view ends before the events that matter arrive. Much longer than that and the weekly detail stops being detail and starts being invention. Nobody knows which week of month five a particular receipt will land in. Thirteen weeks is where weekly granularity is still meaningful.

Rolling is the more important word. Every week the treasury does not extend the old forecast by one line; it rebuilds the whole thing, and the view slides forward so it is always thirteen weeks long. The rebuild has an effect most people miss on first reading. Because the view rolls, every single calendar week gets forecast thirteen separate times before it arrives, and those thirteen sightings are the only reason the accuracy of this document can be measured at all.

Follow one week through. A week that is thirteen weeks away today is seen first at the far end of the view. Next Monday it is twelve weeks away, then eleven, and so on, until the Monday when it is next week and the treasury is looking at it for the thirteenth time. Twelve rebuilds separate the first sighting from the last. Every one of those thirteen sightings can afterwards be compared with what really happened, and that comparison is where every number in the rest of this guide comes from.

A household version shows the same mechanism. Somebody writes down, every Sunday, what they expect their bank balance to be on the last Sunday of the quarter. The habit produces thirteen guesses about one day. Laid beside the balance that actually turned up, they show something no single guess could ever tell: not whether the guesser is any good, but how the guessing improves as the day gets closer. Nirjhar's treasury keeps exactly that record, and the record is the difference between a forecast that has an opinion about itself and one that does not.

What is missing when somebody quotes one cash number?

The central sentence is this. A forecast is not one number, it is a number and a horizonHow far ahead a particular number is looking. It is part of the number rather than a caption on it, because the size of the likely miss depends on it entirely., and quoting one without the other is quoting half of it.

Treat the horizon as part of the value, the way a currency is part of a price. Nobody would write down an amount and leave off whether it was rupees or dollars and then defend the omission by saying the digits were correct. The digits are not the statement. In exactly the same way, Rs 480 crore at week 1 and Rs 480 crore at week 13 are not the same reading. One carries an expected miss of Rs 7.56 crore and the other carries one of Rs 87.48 crore, on the same document, produced by the same people on the same morning.

A good forecast document therefore never has a single headline figure. The document has a row of weeks, each one labelled with how far ahead it is looking, and the reader is never allowed to lift one figure out of the row and carry it into a meeting on its own. The moment a number leaves the document without its horizon attached, it acquires a false precision the forecaster never claimed and cannot withdraw.

How wrong is this forecast, and how fast does that get worse?

Nirjhar's treasury measured itself over the last thirteen weeks, comparing every forecast reading against the actualWhat really happened in a week once it arrived. Every error measurement in this guide is computed against it rather than against a later version of the forecast. that eventually arrived, and published the result. The measure it used is mean absolute errorThe average size of the miss, ignoring whether the miss was over or under, so that a run of overs and unders cannot cancel out and flatter the record.. A measure that let overs and unders cancel would report a treasury that is wildly wrong in both directions as one that is nearly perfect.

The four readings are one group's measurement of its own document, and they are not a norm anybody else has to match. Read the two ends together and the point arrives on its own. The far end of this forecast is 16.2 over 1.4 = 11.57 times as wrong as the near end, on one document, produced by one team, on one morning.

Notice what that sentence does not say. The sentence does not say the treasury is careless at week 13 and careful at week 1. The same people did both, with the same care and the same inputs. Nothing about the quality of the work changes across the row. The only thing that changes is how far ahead each reading is looking. The honest description is that the error belongs to the horizon and not to the forecaster.

The case fixes four readings and no more. Everything shown for the other nine weeks is a straight line drawn between the fixed points, at 1.1333 points a week from week 1 to week 4, 1.2 a week from week 4 to week 8 and 1.32 a week from week 8 to week 13. The joins give 1.40, 2.53, 3.67, 4.80, 6.00, 7.20, 8.40, 9.60, 10.92, 12.24, 13.56, 14.88 and 16.20 per cent across the thirteen weeks. Four of those thirteen numbers are measurements and nine of them are joins, and every drawing in this guide says which is which rather than letting a smooth line imply a precision the record does not carry.

An answer committed to before the control below is moved is worth more than one formed afterwards. Predicting first and then testing is how a reader finds out whether a shape is genuinely obvious or merely obvious after the fact.

What does that error look like in rupees?

Percentages are easy to nod at and hard to feel. A room that shrugs at 16.2 per cent will sit up at the rupee figure, and the rupee figure is the same statement translated. A percentage without a base is another number missing half of itself, so the percentage is applied to the group's Rs 540 crore of cash and the base is stated alongside it.

Being wrong by Rs 7.56 crore on a Rs 540 crore position does not undo a decision, and being wrong by Rs 87.48 crore undoes almost any decision anybody would take on a number that size. The distinction is entirely practical, and it took a rupee conversion rather than an argument to make it obvious.

One thing this shape does not do, and the absence is worth stating plainly rather than leaving a reader to assume it. The group's own treasury policy sets TL5, a minimum liquidity of Rs 120 crore to be held at all times, so from Rs 540 crore of cash there is Rs 420 crore of headroom above that floor. The largest expected miss anywhere in the thirteen weeks is Rs 87.48 crore, or 87.48 over 420 = 20.8 per cent of that headroom. The measurement error never reaches the group's own liquidity floor at any horizon in this view, and a shape that implied a crossing would be teaching something the record does not contain. A group is put through a liquidity floor by the flow, meaning money genuinely leaving faster than it arrives, and not by the size of the error bar on the balance.

What does an average error hide that a hit rate shows?

An average is a flattering measure and it is flattering in a specific way: it gives the typical size of the miss and says nothing whatever about how the misses were distributed. Two forecasts can average the same error and be entirely different documents to live with. One is quietly a little wrong every week. The other is nearly perfect most weeks and catastrophically wrong occasionally. A treasury can be run on the first and nobody would touch the second.

So Nirjhar's treasury publishes a second measure beside the first. The treasury picked a tolerance of 2.0 per cent, its own choice and not a norm of anybody's, and counted how often each horizon landed inside it. The count is the hit rateHow often a forecast landed inside a stated tolerance. It answers a question the average cannot, which is whether the misses are steady or lumpy., and it is the number that makes the average honest.

The week 1 reading landed within 2.0 per cent of the actual on 11 of the 13 weeks measured, and the week 13 reading did so on 2 of the 13. Read those two counts back to back. The first says a treasurer can plan the week ahead and expect the plan to survive contact with the week. The second says the far end of the document is a direction of travel and nothing more, and no amount of care in producing it will change that.

The counts speak far more bluntly than the averages do. An average of 16.2 per cent still sounds like a number that could be worked around by adding a margin. A record of 2 weeks out of 13 inside tolerance sounds like exactly what it is, a reading that will usually be outside anything a plan could be built on. Same document, same thirteen weeks, two measures, and the second is the one that stops an argument.

Which weeks does this treasury decide on, and which does it only take warning from?

Here is where the measurement turns into a working rule. Nirjhar's treasury uses weeks 1 to 4 for a decision and everything from week 5 to week 13 for a warning. Weeks 1 to 4 and weeks 5 to 13 are two different uses of one document, and the line between them was chosen by people and written down rather than emerging on its own.

A decision weekA horizon accurate enough that the treasury will move money on it, because being wrong by the expected amount would not undo the decision. is a week where the treasury will act. A warning weekA horizon accurate enough to start a conversation and not accurate enough to act on, so what it triggers is enquiry rather than a transaction. is a week where the treasury will talk. Both are real uses. Only one of them costs money on the day it is exercised, and that asymmetry is the reason the boundary has to be written somewhere a worried person cannot move it in a meeting.

Why week 4 and not week 6? Because Rs 25.92 crore of expected miss on a Rs 540 crore position leaves a sweep, a placement or a repayment standing, and the group judged that it does not. The boundary is a judgement about this group's own tolerance for being wrong, made by this group's own treasury, and it is not a rule anybody else has to adopt. Having the line somewhere, and having it in writing, is not optional. A treasury that has never written the boundary down will be argued into crossing it by whoever is most worried in the room, and being most worried is not an analytical qualification.

What does a decision week actually let somebody do?

Being concrete matters here. The phrase "act on it" sounds like it means something and often does not. Inside weeks 1 to 4 this treasury does five things, and each one is a transaction with a date and a counterparty.

The treasury sweeps, meaning it moves cash out of the operating accounts into the header account so the position is in one place. It places, meaning it puts a surplus somewhere for a stated period inside the group's own tenor cap of 12 months under limit TL2, choosing from the instruments its own policy permits under TP1 to TP6, each of which is named there and covered under the treasury policy. It draws on a facility where the week is short and repays where the week is long. And it moves the date of a payment, the cheapest of the five and the one most often forgotten.

Every one of those five is reversible or cheap, and every one of them is worth doing only if the week it is aimed at is roughly the week that turns up. The connection between the measurement and the action sits exactly there. A sweep aimed at a week that misses by Rs 7.56 crore still lands. A sweep aimed at a week that misses by Rs 87.48 crore may have been aimed at nothing.

What is a warning week worth, if nobody may act on it?

A reader who has followed the argument this far usually asks the sharp question: if week 11 cannot be acted on, why publish it? Why not print four weeks and stop?

Because time is the one thing a treasury cannot buy later. A shortfall that shows at week 11 gives ten clear weeks to find out whether it is real; the same shortfall discovered at week 1 gives none. The value of the far end is not its accuracy. The value is earliness, and early is what lets somebody ask a question while the answer can still change something.

Think about what actually happens in those ten weeks. Somebody asks the sales side whether the large receipt in that week is firm. Somebody checks whether a supplier payment can be moved a fortnight. Somebody looks at whether an instalment falls in the same week as a tax payment and whether either can shift. None of that costs anything. All of it makes the week look different by the time it becomes a decision week, and by then the treasury is deciding on a 4.8 per cent number rather than a 16.2 per cent one.

The far weeks are not a worse version of the near weeks. They are a different instrument entirely: an early enquiry trigger, and the enquiry is free.

How is a forecast's accuracy measured honestly, and what does publishing it cost?

Everything above rests on one habit, and it is the habit most treasuries skip. Nirjhar's treasury keeps every forecast it ever published, compares each reading against what actually arrived, and prints the resulting error beside the next forecast at the same horizon. Nothing exotic is involved. The rolling rebuild has already produced thirteen sightings of every week, so the record is a by-product of the process rather than an extra project.

The cost is comfort. A treasurer who publishes 16.2 per cent at week 13 is handing the board a stick, and the first meeting after that number appears is not a pleasant one. The discomfort is exactly why so few do it, and why the ones who do are believed.

Without a published error, every reader supplies a private guess about how wrong the number is, and the private guess is almost always harsher than the measurement. The trade is exactly that. Publishing 1.4 per cent at week 1 is what buys the right to have week 1 acted on. Hiding 16.2 per cent at week 13 does not make week 13 more believable; it makes the whole document less believable, including the part that was genuinely good.

What happens when a committed payable lands in the worst week of the view?

Nirjhar has a payable of 24 million United States dollars (USD 24 million) falling due in 90 days. Its overseas trading entity E3 buys in dollars while the group reports in rupees, and the amount is booked at the group's own contracted rate of Rs 84.00 to the dollar rather than at any market rate, making it Rs 201.6 crore. The payable is 201.6 over 540 = 37.3 per cent of the group's entire cash.

Now put it on the calendar of the forecast. Ninety days is 90 over 7 = 12.86 weeks, so the payable lands in week 13 of the view running today. The single largest committed outflow the group can see falls in precisely the week where its forecast is least reliable, and that is a clash of two clocks rather than a contradiction in the numbers.

Exactness about what is uncertain matters here. The payable is the easiest point in the subject to get backwards. The 16.2 per cent error is on the cash position and it is not on the payable. The payable is a contracted amount: Rs 201.6 crore is what it is, and no forecast error applies to it, ever. The 16.2 per cent belongs to the forecast of how much cash the group will be holding in the week the payable settles. The amount is certain and the funding of it is not, and those are two different anxieties with two different fixes.

Getting that backwards has a real cost. A treasury that reads the 16.2 per cent as uncertainty in the payable will start managing an exposure that is already fixed, and hedging a number that cannot move is an expensive way of feeling organised. The exposure decision on that payable is taken under the group's own treasury policy limit TL3, setting a minimum hedge ratio of 60.0 per cent on a committed foreign currency exposure. The 60.0 per cent is this group's own figure rather than anybody's requirement. The hedging decision belongs to the day the exposure was committed, and it is not the forecast's question. A forward is named in that policy under TP1 to TP6 and is covered separately, under instruments.

What goes wrong when somebody treats week 13 as though it were week 1?

What can a cash forecast not tell?

Three things, and a treasury that is clear about them argues less.

A forecast cannot make a payment arrive. A forecast is a description of what the treasury expects, and describing a receipt does not cause it. The week 11 shortfall on the last build was a statement about the world, not a lever on it, and the ten weeks of enquiry it triggers are where any actual change comes from.

Nor can it make a customer pay. The customer version is the same point with a name attached, and it is worth separating because it is where forecasts get blamed for things that are not forecasting failures. If a large customer takes an extra fortnight, the forecast that said otherwise was not wrong about arithmetic; it was wrong about somebody else's behaviour, and behaviour is not a thing arithmetic reaches.

And its error never attaches to a contracted amount. The payable point returns here as a general rule: the error lives on the position and not on any figure fixed by a contract.

One collision is worth naming here because two entirely different objects in this group carry the same number. Nirjhar's profit after tax for the year is Rs 324 crore, and the net position across the 18 bank accounts in its physical concentration, after the daily sweep, is also Rs 324 crore. The first is a result for a whole year on the profit ladder; the second is a balance sitting in bank accounts on a day. The two share a figure and nothing else, and a forecast that quietly treats one as the other has produced a number no reader can use.

Who actually reads this document, and what do they do with it?

Four readers, and each one takes something different out of the same document.

The group treasurer reads the near weeks and acts. Girish Talwalkar is deciding whether to sweep, place, draw or repay this week, and he needs the week 1 and week 2 numbers to be close enough that a placement made on Tuesday is not unwound on Thursday. For him the error column is a permission slip.

The finance director reads the far weeks and asks questions. The value of week 9 to a finance director is not that it is right; it is that it is early enough for a question to change something, and the error column tells him which questions are worth asking and which are noise.

A lender reads it as evidence about the borrower rather than as a prediction. Vindhya Commercial Bank Limited holds Nirjhar as counterparty C1, its largest single-name exposure. A lender rarely takes the number from a borrower's cash forecast and almost always takes the fact that the borrower measures itself. A treasury that publishes its own error is a treasury that knows where its cash is. A borrower who cannot say how wrong its forecast usually is has told the lender something without meaning to.

And a household reads the same shape without calling it anything. Anybody running a home on one salary already knows the difference between what next week costs and what March costs. Next week can be planned around to the rupee, and March can only be prepared for in general. One person holds both ends of that boundary at home, so nobody writes it down and nobody needs to. A group with 26 accounts, 4 banks and 1,840 payments a month has no such luxury: the boundary has to leave somebody's head and become a line in a document, or it will be argued about every single week.

Where does this subject stop, and what holds the rest?