Herding: Following the Crowd, and Why It Can Be Rational

Herding means choosing what other people chose because they chose it. Copying is not automatically an error: where others hold information a decider lacks, imitating them is a defensible reading of the evidence. Copying turns into an error once everybody is imitating and nobody is examining. A crowd assembled that way carries no information at all.

The hard part of this subject is not the definition. HerdingDoing what others are doing because they are doing it. rests on a single distinction, and that distinction decides everything downstream: did the people being copied look at the thing itself, or did they look at somebody else who looked at it? A queue of people who each examined the thing is powerful evidence, and the longer that queue runs the stronger the evidence gets. A queue of people who each examined the queue is one examination repeated over and over. The two queues look exactly the same from the street, and nothing about their appearance will ever tell a passer-by which one has been found. So the useful move is a question about the people in a queue rather than a rule about queues.

What is herding, precisely, and how is it told apart from agreement?

Herding is doing what others are doing because they are doing it. The word doing the work in that sentence is because. Herding is a claim about the reason for a choice, not about the choice itself. The question can never be settled, then, by looking at what people ended up holding. Two neighbours who each read the same statement of eight printed sides, each work through the same numbers and each conclude the same thing have agreed. Two neighbours where the second glanced at the first, thought that will do, and matched it have herded. The paperwork afterwards is identical in both cases.

The shape is older than money and appears well outside finance. A traveller new to a town and hungry passes two eating places side by side. One has a queue at the door and the other is empty. Almost everybody walks towards the queue, and almost everybody is right to. The queue is evidence. Sixty people have been to this street already and have voted with their evening, and unless the traveller knows something about the empty one that they do not, matching them uses information no single person could ever have gathered alone. Copying is not a character flaw and it is not laziness: for most of human history it has been the cheapest reliable way to learn anything.

Now change one detail and watch the evidence drain out of the picture without the picture changing at all. Suppose the first three people joined that queue because the place looked clean, and everybody after them joined because there was a queue. The queue is now sixty long and it still contains three opinions about the food. The queue looks identical. A longer queue even feels more convincing than a short one. That feeling is exactly backwards: the twentieth person in a queue is weaker evidence than the third was, not stronger. By the twentieth, the queue itself has become the reason people are joining.

Two words carry the whole mechanism, so they are worth fixing now. A private signalThe reading somebody has of their own, before seeing what anybody else did. is whatever a person worked out before looking around: the statement read, the shop visited, the arithmetic done on the back of a bill. An independent observerSomebody who looked at the thing themselves rather than at other people. is somebody whose choice was driven by their own signal. Herding is what happens when the second input, what others visibly did, outweighs the first. Setting a private signal aside can be entirely reasonable. Where a person's own reading is weak and forty people who know the trade have gone the other way, setting that reading aside is the sensible move rather than the weak one.

When is copying other people a sound use of evidence?

Most of the writing on this subject treats copying as the mistake and independent thinking as the virtue, and that framing is wrong often enough to be worth correcting properly. People copy constantly and are right to. A patient takes the medicine most doctors prescribe rather than reading the trials. A pedestrian crosses the road when the people alongside cross. A household sends a child to the school other households in the lane chose. In every one of those cases other people have done work that cannot be repeated, and their choices carry that work forward at no cost. The question is never whether to use what other people know, but whether the thing being looked at contains anything they know.

Three conditions have to hold before a crowd is worth copying, and all three are checkable in advance rather than in hindsight. The first is that its members examined the thing rather than each other. The second is that their interests point roughly where the decider's do. Their choice then signals something about the thing rather than about what they were paid to say. The third is that they chose separately rather than in a visible sequence. A sequence lets each person see the ones before and quietly stop contributing anything of their own. If any one of the three fails, the crowd's size becomes decoration.

The eating place runs through the three. Did those sixty people examine the thing? The first arrivals did. There was nothing else to look at. Do their interests run with the newcomer's? Yes, nobody in the queue gains from one more person joining it. Did they choose separately? Here the test fails, and it fails in the ordinary way: they arrived one after another, in full view of each other, and the later ones could see the queue before deciding. So the queue passes two conditions of three. Two of three makes it decent evidence rather than excellent evidence, and the sensible move is to walk towards it while remaining willing to be disappointed.

Now run one line of the Palash decision log through the same three. The Palash log, an invented record kept by Palash Advisory Services Private Limited where Devika Rao advises, holds 240 decisions taken by 60 investors over eight quarters. On 19 February a television segment named Suvarna Chemicals Limited, and Meera Sundaram put another Rs 1,00,000/- into it the same evening, taking that position from Rs 3,00,000/- to Rs 4,00,000/- and the whole holding from Rs 12,00,000/- to Rs 13,00,000/-. She did not act because a crowd was visible. She acted because a broadcast was, and a broadcast is a claim about how many people are about to look. The condition that failed was the first one: nobody involved, Meera included, could have named what the people reacting to that segment had examined.

What is an information cascade, and how does one form step by step?

An information cascadeA sequence where each person rationally ignores their own reading and follows the visible choices. is the sharp version of herding, and it was set out twice in the same year by people working separately. Banerjee, in A Simple Model of Herd Behavior in the Quarterly Journal of Economics in 1992, and Bikhchandani, Hirshleifer and Welch, in the Journal of Political Economy in 1992, both showed the same uncomfortable thing: a sequence of people can each behave perfectly sensibly and the group can still end up in a position that almost none of the private evidence supported. Nobody in the sequence is careless. Nobody is panicking. Each one is doing the rational arithmetic, and the arithmetic tells them to stop using what they know.

Watch it happen with four people choosing between two doors. The mechanism has nothing to do with money, so a door serves as well as a share. Each person gets a weak private hint about which door is right, and each can see which door the people before them walked through, but not what hint those people got. The first person has nothing to go on but their hint, so they follow it. The second has their own hint plus one visible choice, and if the two disagree they are weighing one hint against one choice, evenly matched, so they can go either way. The third person is where the structure changes. Two visible choices now point one way against one private hint pointing the other, and the honest arithmetic says to set the hint aside.

Now say what happened to the fourth person. The fourth sees three matching choices. But two of those three were made by people who had already stopped using their own hints, so the three visible choices contain one hint, or at most two. The fourth person cannot tell that. From where they stand, three is three. The fourth person sets that hint aside as well, and so the sequence has quietly converted itself from a device that pools information into a device that repeats it. The private hints of everybody from the third onward are simply lost. Those hints were never wrong; they were never used.

How fast does a cascade lock, and why is the number so small?

One number carries the weight of this section. The sequence above takes a scale: the share of people who follow the visible choice instead of their own reading, plotted against how many people are visibly ahead of them. The scale starts at 30.0 per cent, simply how often a private reading happens to point at that choice anyway. Each additional visible chooser closes six tenths of the remaining distance to a ceiling of 90.0 per cent. The shape is assumed rather than measured, and it is picked to make the important feature legible rather than to be the true curve.

Work the readings out. With nobody ahead, 30.0 per cent. With one person ahead, 54.0. With two, 68.4. With three, 77.0. With five, 85.3. With ten, 89.6. The share never actually reaches 90.0, though from the fourteenth person onward it rounds to it. The striking thing is not the ceiling, it is how early the climbing stops: the first three visible choosers deliver 78.4 per cent of the entire movement, and the seventeen people after them share out the remaining fifth between them.

Say the same thing a second way. The curve makes the fall look gentle and it is not. Ask what each successive person adds. The first visible chooser adds 24.0 points. The second adds 14.4. The third adds 8.6. The fourth adds 5.2, the fifth 3.1, the sixth 1.9, the seventh 1.1, the eighth 0.7. By the eighth person a whole human being with a whole opinion is worth less than a point, and the queue is still growing at exactly the same speed. The look of agreement and the weight of the evidence have come apart, and only one of the two is visible from outside.

Herd Behaviour vs Information Cascade: which word applies when?

Herd behaviour and information cascade get used as though they were spare parts for each other, and they are not. Herd behaviour is the pattern that is visible: people converging on one choice. An information cascade is one particular machine that produces that pattern, and it needs two ingredients that herding in general does not require. A cascade needs a sequence, so some people must decide after others. And it needs the earlier choices to be visible while the earlier reasons are not. With either ingredient removed the machine will not run, though the pattern can still appear by other routes.

So the containment runs one way only. Every cascade is herd behaviour. Most herd behaviour is not a cascade. Twenty people who each read the same report and reach the same view have converged without watching anybody, and the twenty views really are twenty views. A hundred savers who all move at the end of the tax year are following a calendar, not each other. A room of professionals who all avoid the unusual choice because being wrong alone is punished more heavily than being wrong together are herding for reasons of reputation rather than information. Naming the mechanism matters because each one is broken by a different thing, and a remedy aimed at the wrong mechanism does nothing at all.

The distinction earns its keep the moment somebody tries to fix something. If a room converges because everybody read one report, the remedy is a second source, and asking people to decide privately will change nothing. If a room converges because each person spoke after hearing the last, the remedy is to collect views in writing before anybody speaks, and a second report will change nothing. If a room converges because nobody wants to be the one who was wrong alone, neither remedy touches it, and only changing what gets punished will. Same visible consensusApparent agreement, which may rest on many observations or on very few., three different machines underneath, three different levers.

One more word, and it will come back. A cascade is said to lockThe point at which further choices stop adding information to a cascade. once further choices stop adding anything to what is known. The scale above puts that moment absurdly early. Locking is not a wall, and nothing prevents a later person from breaking ranks. Locking is simply the point past which breaking ranks is no longer arithmetic. The arithmetic has been telling everybody to conform since the third person joined.

What is Performance Chasing, and why is it the commonest form?

Performance chasingMoving towards whatever has done well recently, because it has done well recently. is moving towards whatever has done well lately because it has done well lately. Performance chasing is the commonest form of herding there is, and the reason is worth stating plainly: it does not feel like following anybody. There is no queue in sight and no crowd to join. There is a list, sorted, and the sorting looks like a fact about the things in it rather than a record of what other people already did. Sirri and Tufano, in Costly Search and Mutual Fund Flows in the Journal of Finance in 1998, examined how money moves towards what has recently done well and showed that the flow responds far more strongly to good recent showings than it does to poor ones, with the cost of searching helping to explain the asymmetry.

A sorted list is worth looking at for what it actually is. Each row's position was produced by things that already happened, and every reader of the list sees the same ordering at the same moment. A sorted list is a queue with the people removed. Moving towards the top of it is exactly what the fourth person at the door did: acting on a visible outcome of other people's decisions while knowing nothing whatever about the reasons behind them. The list is more dangerous than a queue precisely because it does not look social, and a reader who would never say they follow crowds will move to the top of a ranking without noticing they have done anything of the kind.

The same shape shows up wherever attention gets concentrated, and the media entry in the Palash log is one more case of it. Of the 96 buys recorded in that invented log, 41 followed a media mention within three days, being 42.7 per cent. Only about 11.0 per cent of the eligible list was mentioned at all in a given week. Notice what that comparison does and does not show. The comparison does not show that anybody was copying anybody. The comparison shows that attention and action moved together. A footprint like that would be left behind by a room full of people examining the same newly noticed thing and by a room full of people watching each other alike. The log records the footprint. The log cannot record which foot made it.

Why is a crowd of copiers one observer repeated?

Return to the arithmetic of what a crowd contains. Sixty people who each examined the thing hold sixty readings, and if their readings are even loosely independent the errors partly cancel and what is left is a considerably better estimate than any one of them had. Cancelling error is the whole reason crowds are ever worth listening to, and the effect is real rather than a folk belief. But it depends entirely on the readings being separate work. Sixty copies of one reading is not sixty readings. Sixty copies are one reading printed sixty times, and printing does not reduce error.

The word indistinguishable earns its place here. Set side by side, the two crowds offer nothing that separates them. Both have sixty members. Both hold the same view. Both will answer the same way if asked what they think. The copiers genuinely do think it. Both will grow at the same rate, express the same confidence and produce the same paperwork. There is no observable feature of a crowd, none at all, that separates sixty examinations from three examinations and fifty-seven copies. Size does not do it, speed does not do it, and confidence does the opposite of doing it. A copied view is often held more firmly than an examined one, for the simple reason that it never met an inconvenient detail.

Which means the question to ask is not about the crowd at all. The question is about each member of the crowd, one at a time, and it is short: what did this person examine? Asked of three or four members, it yields most of what there is to learn. If the answer keeps coming back as some version of everyone else is doing it, the crowd is the right-hand panel, and the sixty standing there are really a three. If two or three of them can say what they read and what they concluded from it, the crowd is the left-hand panel and it is worth something. Neither answer settles what to do. Both settle what is being held.

What can break a cascade once it has formed?

A cascade sounds like a trap and it is not one. A cascade forms fast because it rests on very little, and the same fact makes it fragile: a structure that locked after three weak hints can be undone by one strong reading. Fragility is the reason converged crowds sometimes reverse in a way that looks wildly out of proportion to whatever triggered it. Nothing large broke. Nothing large was holding it up.

Four things break one, and they are worth keeping separate because only some of them are available to a person rather than to a room. A private reading strong enough to outweigh the visible pile breaks it. Somebody with genuine specialist knowledge can stop a cascade dead. Information released to everybody at once breaks it. Such information does not have to travel through the sequence to arrive. Reasons made visible alongside choices break it. The moment a choice arrives with its reason attached, a copied choice stops being able to pass as a reading. And breaking the sequence breaks it. People choosing at the same moment cannot watch each other. Three of those four are changes to how the decision is arranged rather than to how anybody thinks, and that is the practical news.

The household version of the fourth one is the cheapest remedy of all. Three people deciding where to eat will converge on whatever the first person says out loud. If all three write a name on a slip of paper first, three names often appear, and the conversation that follows starts from three readings instead of one. Nothing about anybody's character changed. The order of speaking changed, and that was the whole mechanism.

When is following the crowd simply correct?

Following the crowd is correct when its members examined the thing rather than each other, and when they know more about it than the person deciding does. Saying so without hedging matters. An account that only warns has taught half a thing. The situation is extremely common. Copying covers the medicine, the road crossing, the school in the lane and most of the practical decisions anybody makes in a week. In those cases refusing to copy is not independence, it is a decision to throw away free information in order to feel a particular way about oneself, and it usually costs something.

The error is never copying. The error is copying without ever having asked the question. The question is cheap, answerable in advance, and almost never asked. And it is worth being clear about what a satisfactory answer looks like: not certainty, just a name. If what one member of the crowd examined can be named, or one thing they know that the decider does not, the crowd has passed. If the only answer available is that a lot of people are doing it, then what has been learned is that a lot of people are doing it, and nothing else at all.

How does somebody deciding for other people put this to work?

Devika Rao, the adviser at the invented Palash Advisory Services Private Limited, cannot stop anybody herding and does not try. She can change what the record contains, and the record decides what can be examined a quarter later. Her file note on any decision that arrived with a crowd attached carries three lines and no opinions. First, what did the people being matched examine, named if it can be named and marked not known if it cannot. Second, would this decision be taken if nobody else had taken it, answered yes or no rather than discussed. Third, what would have to be true for this crowd to contain three readings instead of sixty. None of those lines is a judgement, and that is exactly why they survive being read back by somebody who was not in the room.

For a person deciding alone, with no adviser and no committee, the same three lines work unchanged and take about a minute, written on the back of the statement. The second one does most of the work. Of the three it is the only question that can always be answered honestly and immediately: would I be doing this if nobody else were? A no is not a reason to stop. A no is a reason to go back and find the first line. If the first line comes back blank the real reason for the decision is now exactly known, and that is more than 156 of the 240 entries in that invented log can say for themselves.

A last word on what any of this is worth. A lender reading a file, an analyst reading a note and a household reading its own statement are all doing the same thing: separating what is known from what is merely widely held. None of it improves anybody's outcome, and one invented log of 60 people over eight quarters could not support such a claim. The three lines make the crowd in a decision countable instead of atmospheric. Whether any authority requires a reason to be recorded is a separate matter entirely, and the Securities and Exchange Board of India is the place to confirm what applies.

Meera Sundaram, Devika Rao, Palash Advisory Services Private Limited, the Palash decision log, the Palash 100 index, the Vindhya index scheme, the Nilgiri mid-cap scheme, Suvarna Chemicals Limited and Kesari Logistics Limited are invented.
Educational material. Not advice on any investment, tax, budget or market position.