Nowcasting: Estimating the Present Before the Data Arrives

Nowcasting is estimating what a figure will say about a period that has already finished. Nowcasting exists because measurement takes time: the quarter has ended and the number describing it has not been published yet. A nowcast reaches backwards into a settled past rather than forwards into an open future, and that direction in time is exactly what separates it from a forecast.

Two earlier ideas carry most of the weight underneath that answer, and neither needs repeating at length. The first is the vintage discipline. One growth figure is published more than once, and the later publications are not corrections of a blunder: growth in the invented Republic of Sankhya reads 6.20 points at first print, 6.50 at first revision and 6.80 at a later vintage, all describing the same finished period. A figure quoted without saying which of the three it is has left out half of itself. The second is the habit of expressing an input as an effect per one unit of movement. A quantity has to be in that shape before it can be combined with any other quantity. Nowcasting lives in the waiting room between those two ideas. A period ends. The figure describing it is not ready. Somebody still has to say something about the period, and there is an honest way to do it that comes with a method, a published set of parts and a scorecard.

What is a nowcast actually estimating?

Start with time. Almost every confusion about nowcasting is a confusion about where in time the estimate is standing.

A period ends on a particular day. Everything that was going to happen in it has happened: the trucks that were going to move have moved, the power that was going to be drawn has been drawn, the goods that were going to be sold have been sold. The quantity that describes all of that activity therefore exists in the world already. The published figure has not been produced yet. Nobody has finished collecting, reconciling and adding up the parts. A nowcast is an estimate of a past quantity, not a statement about anything still to come, and a reader who files it as a prediction has put it in the wrong place in time.

The household version is quick and it sticks. An electricity meter has been running all month. The month has ended. The reading on the meter is a real, settled, already determined quantity, and it is sitting on the wall right now. The bill has not arrived. Somebody still has to read the meter, apply the slabs and print the thing. Walking over, looking at the dial and working out roughly what the bill will say is a nowcast. Nothing has been predicted. A number that already exists, and has not yet reached the household, has been estimated.

The meter on the wall settles what kind of defence a nowcast needs. Nothing about the world has to cooperate for it to be checkable. No event has to occur. The activity is done, the compilation is under way, and the answer is coming out of a process that is already running. The only uncertainty is about how well the visible pieces stand in for the ones that are not.

Why does a finished period need estimating at all?

Because a figure of that size is not observed anywhere. The figure is built, and building takes steps, and the steps have to happen in an order.

Think about what has to occur before a growth number for a whole economy can exist. Firms have to be asked what they did. Quantities have to be counted at ports, at power stations, on rail networks and in tax filings. The counts have to be checked against one another, converted into consistent units, and reconciled where two sources disagree. Only then can the parts be assembled into one aggregateA single figure built by adding and reconciling many separate sources so it stands for a whole economy or a whole sector, rather than a figure that anyone counted directly. and expressed as a rate of change against the matching period a year earlier. Every one of those steps has to wait for the one before it.

The material therefore arrives in a stable order. Surveys of firms come first. A survey asks people what happened rather than counting what happened, and answers travel faster than quantities. Volume counts come next. A quantity has to be gathered first, and only then can it be added. Aggregates come after the volumes. The aggregate is assembled out of them. Revisions come last. A slower source finally arrives, replaces an early stand in, and the figure is restated. The wait belongs to the way a figure is put together rather than to anybody who puts it together. No amount of urgency shortens it.

One thing about that order is worth holding on to. The order is durable, and the timing attached to it is not. Which release logically precedes which is a fact about construction that stays true from one year to the next. When any of them lands in any particular country is a schedule, and a schedule is exactly the sort of thing that goes quietly out of date. The order is what holds.

What is a nowcast built from, and where do the weights come from?

From the material that is already out. Not from a fresh count, not from a hunch, and not from the previous period carried forward.

Some indicators arrive early precisely because they measure less. A survey of firms can be answered in days because nobody has to count anything. A tally of freight moved on a rail network exists because the network was billing for it anyway. A record of power despatched to a grid exists because the grid meters itself continuously. Each of those is narrow and each of them is fast, and a nowcast is what results when several narrow fast things are combined into a statement about the wide slow thing actually wanted. A nowcast is a weighted reading of what is already published rather than a new measurement of anything.

The weights are the whole method, so they have to be published. A weight says how much of the slower figure a faster indicator has usually tracked, and it is exactly the form that sensitivityAn effect stated per one unit of movement in the thing causing it, such as the change in an outcome for each one per cent move in an input. Stating it per unit is what lets two different inputs be set side by side. takes when the input is an indicator rather than a macro variable. Here is Sankhya's, with everything on the table.

The arithmetic is the argument, and it is worth working through once. Forty five per cent of 7.20 is 3.24. Thirty per cent of 5.40 is 1.62. Twenty five per cent of 5.60 is 1.40. The three together come to 6.26 points, and the weights add to 100 so nothing has been quietly left out or counted twice. Anybody holding the same six numbers reaches the same 6.26, and anybody who disagrees can say precisely which of the six they would change. An estimate built that way is a very different object from one that arrives with a name attached and no parts.

How is a nowcast checked once the figure arrives?

By subtraction, and this is the part that makes the whole exercise honest rather than decorative.

An estimate that can never be compared with anything is an opinion with arithmetic attached. A nowcast is not in that position. The figure it was estimating turns up, on a process that was already running when the estimate was made, and at that moment the estimate becomes checkable to the second decimal place. The answer arrives through a known sequence of steps rather than never arriving at all, and that makes a nowcast falsifiable in a way an estimate of the future is not always.

But checkable against what, exactly? The figure arrives more than once, and that is where the whole thing turns. Sankhya's growth for the finished period reads 6.20 at first print, 6.50 at first revision and 6.80 at a later vintage. All three describe the same period. None of the three is a mistake. So a nowcast of 6.26 has three candidate targets and three completely different scores, and picking between them is not a matter of taste.

Score the estimate against the first print. The first print is the figure it was estimating. The nowcast was built from indicators that existed before publication, using weights fitted to what those indicators had tracked, and what it was trying to land on was the number that would come out of the first compilation. The estimate was never aiming at a figure that would be assembled later out of sources that had not arrived yet. Against 6.20, the estimate of 6.26 sits 0.06 points above, and that is the score that happened.

Look down that last column before reading on. The estimate is identical in all three rows. Nothing about the method changed, no weight moved, no indicator was restated. The number it was set beside is what changed, and that alone turns a near miss of 0.06 points into an apparent miss of 0.54 points, nine times the size. The difference between the two outer scores is 0.60 points, precisely the revision the figure underwent after publication. Scoring is not a neutral act: the choice of vintage carries the entire revision straight into the verdict.

Why is a nowcast not a forecast?

Because of one thing, and it is not accuracy, not sophistication and not the amount of data involved. The difference is direction in time.

A nowcast estimates a number that already exists and has not been published. A forecast estimates a number that has not yet come into existence. Everything else that separates the two comes out of that. The nowcast is aimed at a quantity that is fixed and merely hidden by compilation. The forecast is aimed at a quantity that nothing in the world has yet determined. The period it describes has not run.

Watch what that does to the defence each one needs. When somebody challenges a nowcast, the argument is about the bridge: do those three indicators really stand in for the wider figure, are the weights the right size, and does the coverage they offer support the claim being made. Every one of those questions can be settled by evidence that already exists, and the whole thing gets marked when the figure lands. When somebody challenges a forecast, the argument is about assumptions describing conditions that have not occurred. The material that would settle such an argument has not been generated, and no amount of published material settles it.

The same difference shows in how each one gets checked. A nowcast is graded on a schedule of steps that is already in motion. A forecast has to wait for a period to run and then be measured. The scoring itself therefore carries every hazard of measurement on top of the hazard of being wrong. The same difference is why a gathered consensusA single figure summarising many separate estimates of the same quantity, usually the middle value of the set. A consensus summarises what estimators expected and measures nothing. of forecasts is a different object again, and consensus is taken up separately.

Keep the everyday pair in mind. Reading the meter on the wall to work out what the bill will say is a nowcast. Working out what next month's bill will be, when next month has not started and nobody knows how hot it will get or who will visit, is a forecast. Both are estimates and both can be done carefully, but only one of them is about a quantity that has already been settled.

What does a nowcast not deliver?

Three things. The last of them is the one a clean two decimal answer quietly talks a reader out of remembering.

A nowcast does not deliver the revised figure. The estimate aims at the first print, so the first print is all it can claim. The reading at a later vintage depends on sources that had not arrived when the estimate was made, and any statement about that reading is a statement about material nobody has seen. A nowcast delivers nothing about a period that has not finished. Nothing has happened yet to be compiled, and no hidden quantity is waiting there to be estimated. And it does not deliver more precision than its inputs carry. The third limit is the one worth sitting with.

The three Sankhya indicators touch a limited part of the economy. Freight tonnage covers activity worth 8.00 per cent of output. Power despatched covers 7.00 per cent. The manufacturing survey covers 5.00 per cent. Added together, the inputs between them touch 20.00 per cent of the economy. The other 80.00 per cent is ground none of the three went anywhere near. A nowcast built from indicators covering a fifth of the economy is an estimate of the whole built from a fifth, and printing it to two decimals does not change that.

Narrow coverage is not a flaw to be fixed by finding better indicators. Narrow coverage is the shape of the exercise. Fast material is fast because it is narrow: the grid meters itself continuously precisely because metering a grid is a smaller job than compiling an economy. Every extra scrap of speed is bought by covering less. Say what the coverage is and publish it beside the weights. A reader can then weigh the estimate against how much of the economy it actually touched.

Two more limits belong in the same breath, and both are about what the estimate has not been asked to do. A nowcast says nothing about whether a figure is seasonally adjustedTreated to remove the regular within year pattern that repeats every year, such as a festival month or a harvest, so that one period can be set against the one before it rather than only against the same period a year earlier.. Adjustment is a property of the figure being estimated rather than of the estimate. And the target is a growth rate measured year on yearComparing a period with the matching period twelve months earlier rather than with the period immediately before it. The comparison removes the repeating within year pattern but makes the reading depend on what happened a year ago.. Everything about how the base period behaved is already baked into that target.

What makes one nowcast worth keeping and another worth ignoring?

A record of what it has done before. Not the method, not the number of inputs, and not who published it.

The practitioner's version of all this is short. An estimate whose past errors have never been measured is an assertion with decimals on it. An estimate that comes with a list of its previous misses, each scored against the figure it was aiming at, can actually be weighed. How far it has typically landed from the target, and whether it leans one way, are both visible. The answer arrives every period on a process already running, and that makes the measurement available for a nowcast in a way it rarely is for other kinds of estimate. Declining to keep the record is a choice rather than a limitation.

Sankhya's record over five periods is below, and both numbers are printed for each period so any of the five errors can be rechecked line by line.

Two readings come out of that table and they say different things. The average size of a miss is 0.22 points, a measure of how far this estimate typically lands from its target. The average error once direction is kept is minus 0.06 points, a sign that the estimate has run very slightly low across the five rather than leaning hard in either direction. A record with a large average size and a near zero signed average is a noisy estimate that is not biased; a record with both large in the same direction is an estimate that is systematically off. The two readings cannot be told apart from a single period. The record exists for exactly that reason.

One boundary belongs here before anybody carries this too far. A nowcast of Sankhya's growth is an estimate of an economy wide figure, and it is not an estimate of any single company's numbers. A wider move works through the revenue, costs and operating profitRevenue less the cost of running the business, counted before interest and tax. For the invented company carried through these notes it stands at Rs 100 crore against revenue of Rs 1,000 crore. of the invented company Nirvi Engineering by separate arithmetic with separate inputs, and that arithmetic is covered separately.