Reputational Herding

What it is

Two very different mechanisms produce the same behaviour, and keeping them apart is the point of this page.

Rational herding is a claim about what a decision-maker can learn. The copier infers that other people may know something, and acting on that inference is correct. Reputational herding is a claim about what a decision-maker is paid for. The inference is irrelevant. The incentive does the work, and the copier may well believe the crowd is wrong.

Michelle Baddeley states the incentive more plainly than any model does: a trader who loses a million pounds while his peers are also losing a million pounds will probably keep his job, and a trader who loses a million pounds while others lose nothing will almost certainly be fired. Relative evaluation makes imitation cheap insurance. The premium is paid by whoever owns the money.

Why relative evaluation is the active ingredient

Nothing here requires a market or a bonus. It requires only that performance be scored against a comparison group rather than against an outcome. Once that is true, the cost of a wrong answer depends on how many other people gave it, and the rational move for the agent diverges from the right answer for whoever hired him.

This is why the construct belongs next to Externalities rather than inside the information story. The copier is not making an error at all. He is optimising against the thing he is actually measured on, and the loss lands somewhere else. It is also why the behaviour is hard to legislate away by improving information: better data does not change a payoff structure in which conformity is insured and independence is not.

What distinguishes it in the evidence

The two mechanisms are separable in principle and awkward to separate in practice, because both predict imitation. Three features of a situation point towards the reputational version.

The agent is an intermediary acting on somebody else's behalf, so there is a gap between whose money is at stake and whose reputation is. The evaluation is relative rather than absolute, which is the defining feature. And the imitation persists in situations where the agent plausibly has better information than the crowd he is joining, which is exactly where an informational account runs out.

Dasgupta, Prat and Verardo's finding matches all three. Institutions traded in the direction of past institutional trades over three to five quarters, and the stocks they persistently bought underperformed over the following 12 to 30 months, while persistently sold stocks did the reverse. Their reported explanation is a simple model of career-concerned trading by delegated portfolio managers rather than a story about information.

In effect

Institutional holdings over twenty-one years

The research case is the one above, and it is the only field evidence this site holds for the construct. Quarterly United States institutional holdings from 1983 to 2004, with the persistence of buying and selling over three to five quarters predicting returns over the following 12 to 30 months in the opposite direction. The authors report that their results are of similar magnitude to, but distinct from, other known asset pricing anomalies.

Two limits travel with it and they are not decorative. The abstract was read in full at the London School of Economics record for the 2007 working-paper version, and the published 2011 text was not read, so nothing beyond the abstract is reported here and the published page range comes from search results rather than from a reading of the paper. More importantly, the authors flag that single-quarter institutional herding predicts short-term returns positively. The horizons disagree. Anybody using this study as a clean indictment of imitation in markets is using half of it.

Agreeing with the room in a meeting

Away from finance, the same structure appears wherever an opinion is scored socially rather than by outcome. In a meeting where the forecast nobody will be blamed for is the one everybody gave, the cost of dissent is borne entirely by the dissenter and the benefit of being right accrues slowly, long after the decision is made and often to somebody else.

This is worth separating from ordinary social pressure, which is the subject of Normative Social Influence. Normative influence is about standing: what agreeing does for your place in the group. Reputational herding is narrower and more mechanical. It survives in a room full of people who dislike each other, because it runs on how performance is assessed rather than on how anybody feels. No evidence on this page measures the meeting case, and it is offered as the structure recognised rather than as a finding.

An analytics team reporting the industry benchmark

In commerce the construct shows up in how performance gets defined before anybody measures anything. A team whose target is the category average has been given a relative benchmark, and a relative benchmark rewards tracking the category rather than beating it. The safest recommendation is the one every competitor is also making, because a shortfall shared across the market reads as conditions and a shortfall alone reads as a mistake.

The practical consequence is a reporting one rather than a psychological one. If the measure is relative, expect imitation, and expect it from people who are reasoning carefully. Changing the behaviour means changing the measure, which is the connection to Counterfactual: an absolute benchmark requires somebody to specify what would otherwise have happened, and a relative benchmark lets everybody avoid the question.

What it does not say

It does not say that fund managers, analysts or anybody else are cowards, careerists or acting in bad faith. The whole content of the construct is that the behaviour follows from the incentive, and that an agent facing relative evaluation can arrive at imitation by careful reasoning. Nothing on this page is an allegation about anybody's conduct, and the authors' own account is a model of incentives rather than a claim about character.

It does not establish that reputational herding rather than informational herding explains any particular episode. Both predict the same behaviour, and the field evidence here identifies the mechanism by the authors' modelling rather than by an experiment that separates the two.

It does not supply a magnitude that can be quoted. A figure of roughly a one per cent fall in future returns for a one-standard-deviation rise in persistence circulated in secondary summaries of the working paper and was not verified at source, so it is not carried here.

And the construct's own origin is not verified on this site. The literature attributes reputational herding to a 1990 paper by David S. Scharfstein and Jeremy C. Stein, which this vault has not read in any form. That attribution is recorded from reference lists, which is why the originator confidence on the underlying note is medium rather than high, and why the evidence here is graded mixed.

Who developed it

The construct is standardly attributed to David S. Scharfstein and Jeremy C. Stein in 1990, a paper recorded here from reference lists and not read. The field evidence and the explanatory model this page relies on are Amil Dasgupta's, with Andrea Prat and Michela Verardo, and the clearest statement of the incentive is Michelle Baddeley's.

---