Leading, Coincident and Lagging Indicators Compared

An indicator leads, coincides with or lags the cycle according to where in one chain it is measured. A decision is taken, then a transaction happens, then somebody bears the consequence. A reading taken at the decision is early. A reading taken at the transaction is level. A reading taken at the consequence is late. Nothing about accuracy is involved.

The classification is the whole of it, and the wording is worth noticing. The rule does not say that early readings are better, or that late ones are stale, or that one class is more carefully compiled than another. The rule says only that the same event can be counted at three different moments, and whoever counts it early is looking at something different from whoever counts it late. Once that is held, the three classes stop being three lists to memorise and become one question that can be asked of any reading.

The reader arrives already able to name the four phases of a cycle, read a growth rate against an inflation rate, and place an economy against what it could be producing. The early signals that show up before a downturn is visible, and the surveys that carry business and household sentiment, have also been met. The timing rule underneath all of it is worked below in years on a single path, so that the gap between the classes is a number rather than an impression.

What decides whether a reading is early, level or late?

Start outside economics altogether. A wedding is booked in January for a date in April. The hall takes a deposit in January, the caterer buys and cooks and serves in April, and the extra hands the caterer took on for the season are let go in July when the bookings thin out. Three people could each say, quite honestly, that they watch the wedding season for a living. The hall manager sees it in January. The caterer sees it in April. The young man who was hired for the season sees it in July, when it is already over.

Nobody in that story is wrong and nobody is slow. All three stand at different points of the same chain, and the chain has a direction. Something is decided. Then the thing happens. Then somebody lives with what it did. Every economic event that a number is ever read about runs through those three stages, and a statistic is simply a count taken at one of them.

An indicator's class is a fact about where on the chain the counting happens, not a fact about how good the indicator is. So a reading never seen before can be placed without consulting any list. Ask one question of it. At which stage does this get counted? If the answer is at the moment somebody decided, the reading leads. If the answer is at the moment the thing changed hands, it coincides. If the answer is afterwards, in what the transaction did to somebody, it lags.

The same test works on a reading nobody has taught. The number of new electricity connections applied for by factories: applied for, so a decision, so it leads. The tonnage actually despatched from factory gates: the transaction itself, so it coincides. The share of workers who have been out of work for more than a year: a consequence, and a slow one, so it lags heavily. None of the three needed a list. All three needed the chain.

What sits at the decision end of the chain?

The readings that lead are the ones taken while the thing is still only intended. An order bookWork a business has taken on paper and still has to produce. The book records promises made to buyers, not goods that have left the yard. counts work that has been promised and not yet done. Approvals for building plans count buildings that somebody has been permitted to put up and has not yet put up. Hiring plans count people a business means to take on. Survey readings of what businesses and households expect to do sit in the same place, and are covered separately.

Notice what these have in common, and it is not vagueness. An order placed in March for delivery in September is a hard fact about September that exists in March. Nobody had to be clever to see it. The fact got written down at the decision stage, and the decision stage happens first. A leading reading is not a forecast; it is a measurement of something that has already happened at an earlier point on the same chain.

Keep those two apart. A forecast is somebody's opinion about a period that has not begun. An order book is a count of commitments already recorded on paper. The order book leads because of when it was counted, not because anybody in it was predicting anything. A leading reading can therefore be checked, argued with and revised, in a way a forecast cannot.

What does a coincident reading actually tell the reader?

Coincident readings are the ones taken at the transaction itself. Output is the obvious one. Industrial productionA measure of the physical volume of goods actually turned out by factories, mines and utilities over a period, counted in quantities rather than in money. is another. Sales measured in real termsA money figure stripped of price movement, so an increase signals extra volume shifted rather than an unchanged volume at a dearer rate. is a third. All three read the cycle at zero offset. In the case of output the offset is zero by definition rather than by measurement: output is not something that tracks the cycle. Output is the cycle.

Which means a coincident reading has no predictive content whatsoever, and a reader who goes looking for some will be disappointed and will probably blame the statistic. A coincident reading gives where things stand and nothing about where they are going. Having no predictive content is not a weakness. Nothing else on the chain can give where things stand at all. The decision stage gives what was intended. Only the transaction stage gives what actually got made.

Here is the whole thing in a household. The shopping list on the kitchen counter is the leading reading: it says what somebody means to buy. The till receipt from the shop is the coincident one: it says what was actually bought, and it settles arguments the list cannot. The card statement that arrives next month is the lagging one: it says what the buying did to the household's money, and it arrives long after every decision in it was taken. Three sheets of paper, one shopping trip, three different moments.

Why does employment respond so slowly?

Employment is the class that gets read wrong most often, so the mechanism is worth working through rather than asserting. Picture a tile works whose orders come in light one week. Does it dismiss anybody that week? Of course not. The works has no idea yet whether the week means anything. There was a festival. One buyer's payment is stuck. A site got delayed by rain. Acting on one light week costs money and might be acting on nothing at all.

So the works waits, and while it waits it does the cheap things in order. First it cuts overtime. Cutting overtime costs nothing to reverse. Then it stops the night shift. Then it lets a temporary contract lapse instead of renewing it. Then it stops replacing people who leave of their own accord, so headcount falls by attritionA workforce shrinking because people who leave are not replaced, rather than because anybody was dismissed. without a single dismissal being announced. Only when the works has concluded that the fall in orders is permanent does it dismiss anybody, and only at that point does anything appear in an unemployment reading at all.

Now run the recovery. The recovery is slower still. Orders return. Does the works hire the following week? No. The works first runs the people it already has for longer hours, and longer hours are free. Then it brings back temporary and contract help, cheap to stop again. Only when it believes the recovery will hold does it take somebody on permanently. Taking somebody on is a commitment. Not dismissing somebody is only a delay.

The employment lag is not sluggishness or bad measurement; it is a business refusing to pay the price of acting on information it does not yet trust, and being wrong about a hire costs more than being wrong about a wait. That asymmetry is why the hesitation on the way up runs longer than the hesitation on the way down. The same asymmetry explains why the unemployment rate, the share of the labour forceThe count of people in work added to the people hunting for it. An unemployment rate divides by that count, never by the whole population. without work, moves last of everything on the chain.

What do the three Sankhya series do in the same six years?

The Republic of Sankhya, an invented economy, runs on a path already fixed. Output runs from Rs 17,47,200/- crore in year 3 to Rs 19,99,066/- crore in year 8, growing 6.00 and then 6.50 per cent, falling 1.00 and then 2.50 per cent, then growing 5.00 per cent again. Unemployment runs 5.45, 4.80, 4.20, 5.10, 6.90 and 7.40 per cent over the same years. Both of those are published figures on this path and are used here as they stand.

There is no published Sankhya reading at the decision end, so the third series has to be built. Call it the order intentions reading, an invented series. The reading is constructed as 50.00 plus 1.20 times the growth that the decisions taken in a year would deliver the following year if every one of them were carried through. The construction is an assumption, not a measurement, and the assumption is precisely what makes the reading lead by one year: the reading is next year's transactions, counted at this year's decisions.

Read down the three number columns and the marching becomes obvious. The order intentions reading is at its highest in year 4, at 57.80, and at its lowest in year 6, at 47.00. Output is at its highest in year 5, at Rs 19,72,414/- crore, and at its lowest in year 7, at Rs 19,03,873/- crore. Unemployment is at its lowest in year 5, at 4.20 per cent, and at its highest in year 8, at 7.40 per cent. Three series, three turning points, and each one arrives a year after the one above it. The chain from decision to transaction to consequence has shown up as a calendar.

One caution before the picture. Year 7 is the trough of output on this path, and it is also where the output gapHow far actual production sits away from an estimate of what could have been turned out without straining. Worked in full where potential output is taught. is at its widest. The output gap is covered separately. Year 7 is genuinely the bottom of this path, not an artefact of the series chosen.

How large is the lag, in years?

Now do the subtraction. An adjective is not an answer. The order intentions reading bottoms in year 6 and output bottoms in year 7, so the intentions reading turns 6 minus 7, one year early. Output has an offset of zero against itself. Zero is true by definition here, and worth saying out loud so that nobody treats it as a finding. Unemployment peaks in year 8 and output bottomed in year 7, so 8 minus 7, one year late.

Minus one, zero, plus one. Say the lag as a number of years rather than as an adjective. Knowing that unemployment lags gives nothing to act on. Knowing that unemployment peaked one year after output bottomed gives exactly how late a reader of it would have been.

One honest complication, and it strengthens the point rather than weakening it. At the top of this cycle the lag does not show. Output is at its highest in year 5, and unemployment is at its lowest in year 5 too, so measured on a grid of whole years the lag at the top is zero. A zero lag at the top is not a contradiction and the series has not misbehaved. An annual grid does that to any delay shorter than a year: the delay is real, it is simply smaller than the smallest thing the grid can show. The delay at the bottom is a full year and therefore visible. And the two delays being different sizes is exactly what the tile works predicted. Holding on to people is cheap to reverse and hiring is not, so the hesitation on the way up runs longer than the hesitation on the way down.

If unemployment is at its worst, has the recovery not begun?

Inside Sankhya year 8 the two numbers read side by side are where the timing rule earns its keep. Output in year 8 is Rs 19,99,066/- crore. Output in year 8 is 5.00 per cent above year 7, a rise of Rs 95,194/- crore. Output in year 8 also sits Rs 26,652/- crore above the year 5 peak of Rs 19,72,414/- crore, a rise of 1.35 per cent. So output has not merely turned upward. Output has passed the highest level it reached before the contraction. On any reading of the transaction stage, this is a recovery that has already happened.

Now the other number. Unemployment in year 8 is 7.40 per cent. Unemployment is up 0.50 points from 6.90 per cent in year 7. The 7.40 per cent is the highest unemployment reading anywhere on years 3 to 8, higher than either contraction year, and it is still rising. On any reading of the consequence stage, things have never looked worse.

Both readings are correct, and they disagree only because one of them is reporting year 8 while the other is reporting what year 7 did to people. The unemployment reading in year 8 is the accumulated consequence of the dismissals that year 7's fall in orders finally forced, arriving on the calendar exactly when the mechanism said it would. The reading is not stale, not broken and has not failed. A lagging reading has exactly one job: reporting a consequence after the event that caused it. Year 8 unemployment is doing that job.

What is a lagging reading actually for?

If it is always late, why publish it at all? Because being late is the only way to be sure. A turn suggested by a reading at the decision end has not yet been tested against anything. Somebody intended something. The intention may or may not have survived. A lagging reading is where it emerges afterwards whether the early signal was real. Nowhere else can it be established. Only the lagging reading is taken after the consequence has landed.

So a turn that a leading reading suggests and a lagging reading has not yet confirmed is an unconfirmed turn, and calling it anything stronger is a choice about risk rather than a statement about evidence. The three classes are not competing to describe the same thing better or worse. The three classes answer three different questions. Where are things going. Where do things stand. Did it actually happen.

Wanting one reading to answer all three questions is the error underneath most misuse of indicators. The error is one of expectation rather than of arithmetic, so no amount of better data fixes it. A reading cannot be early and confirmed at the same time, for the same reason a weather forecast cannot be both a forecast and a record of yesterday's rain.

Why do early readings so often turn out to be wrong?

Because they measure intentions, and an intention can be reversed at no cost whatsoever. An order can be cancelled before a single tile has been made. An approved building plan can sit in a drawer for three years. A hiring plan can be dropped in one meeting and leave no trace anywhere. Every one of those reversals shows up in the leading reading and never shows up in output. There was never any output to show.

Now put that next to the reason the reading is early in the first place. The reading is early because it is taken before anything has been made. And it can be undone because nothing has been made. Reversibility is not a defect sitting beside the earliness. Reversibility and earliness are the same property looked at twice. A decision is early news precisely to the extent that it is not yet a commitment.

The pairing is worth holding as one fact rather than two, in the same way that a business with heavy fixed costs is both a fast riser and a fast faller for one reason rather than for two separate reasons. Attempting to keep the earliness and remove the false signals leaves nothing to remove: a reading that could not be reversed would be a record of something already built, and a record of something already built is not early.

Drag the slider down to 70.00 per cent and look at what has happened. The order intentions reading has not changed by a single point: it still falls from 57.80 in year 4 to 47.00 in year 6, and anyone watching it would have said a contraction was coming. Output never falls. Year 6 grows 1.10 per cent, year 7 grows 0.05 per cent, year 8 grows 5.30 per cent, and there is no trough anywhere on the path to point at. Nothing about the early reading was wrong at the moment it was taken. The decisions really were made, then unmade before they could turn into anything, and the reading has no way to know which of those two futures it is looking at. That is a false signal, and it is not a measurement error.

What does an analyst or a lender actually do with all three?

The working habit is duller than it sounds and it comes straight from the three questions. An analyst uses the decision end to know where to look, the transaction to know where things stand, and the consequence to find out whether the earlier reading was right. Nothing is used for another one's job, and no single reading is asked to carry a conclusion on its own.

Two practical habits sit on top of that. The first is that early readings are treated as provisional. A reading taken at the decision end will be revisedA published figure being restated later as fuller information arrives. Early estimates get revised more often and by more than settled ones. more often and by more than a settled one. The second is that a series with a seasonal shape gets read seasonally adjustedA series with its regular calendar pattern removed, so that a festival month or a harvest month does not look like a turn in the cycle.. A festival month is then not mistaken for a turn. Neither habit changes the classification. Both stop a reader manufacturing turning points that the chain never produced.

And a household does exactly the same thing without naming it. For someone working out whether the job market where they live is turning, the vacancies being advertised are at the decision end and will move first. Whether that household's own overtime has come back is the transaction, and it settles what the advertisements only hinted at. Whether the neighbours who lost work last year are back in work is the consequence, and it will be the last thing to change. Being last is precisely what makes it the one that finally confirms the turn was real.