Herding and Social Proof: The Crowd as Evidence

How a cascade forms, and why it stops carrying information

An informational cascade is a rational process with an irrational outcome. Each person has private information and also observes what the people before them did. When the first few actions agree, the next person weights the group’s behaviour more heavily than their own signal, and acts with the group; the person after that sees an even stronger majority and defers for the same reason. No one is being stupid — each is using the most informative evidence available. The result is that after a small number of decisions the group’s behaviour stops reflecting information, because everyone is following everyone, and the cascade persists until something external disturbs it. That mechanism is what makes markets hard in a specific way. In most settings social proof is useful: a crowded restaurant is weak evidence that the food is good. In a market the thing being observed is the price, and the price is the crowd’s action, so the inference runs in a circle — the evidence that people are buying is that the price is rising, and the evidence that the price should rise is that people are buying. Nothing in that loop is a fact about the business. A thesis assembled from it will be right while the loop runs and has no content when it stops, which is why the reversal is so violent: an empty thesis cannot support a price, so the same feedback that added to the price subtracts from it. Herding is also not confined to retail crowds, and treating it as a failure of intelligence gets the incentive wrong. Professional investors are measured against peers and against benchmarks, are hired and fired on relative performance, and are safer holding what everyone holds. A career-risking bet on a view the crowd rejects is punished if it is early, which means the correct description of the behaviour is that it is individually rational and collectively destabilising — the same structure as the cascade, with career risk as the private signal being suppressed. The five stages of a mania — 1. Displacement: Something real changes: a technology, a rate cut, a new market · 2. Boom: Early buyers profit visibly and the reasoning spreads · 3. Expansion and euphoria: Credit and leverage expand into rising collateral; the price becomes the argument ← · 4. Insider profit-taking: Best-informed holders sell into the strongest flow · 5. Revulsion: A small trigger forces sales, margin calls feed the fall, the cascade runs in reverse ← The stages are descriptive rather than predictive — knowing the pattern does not date the peak. What it does is identify the moment the thesis has stopped being about the business, which is the stage at which new buyers are buying the loop rather than the asset.

Writing a thesis that does not need the crowd

The countermeasure is not contrarianism. The crowd is right most of the time in ordinary conditions — that is what a price is for — and a trader who treats consensus as a contrary indicator simply inverts the error. What has to go is the specific move of using the crowd’s action as the argument. The test is whether the thesis survives the removal of everyone else’s participation: write down what the business or the instrument is worth, on what evidence, and what would have to be true for the current price to be justified. If the answer is “the price has been going up”, the position has no thesis and the loop is the whole of it. Three habits do most of the work. First, attach every claim to a fact that is not the price — a margin, a unit volume, a cash flow, a spread, a policy path — so there is something to check that the crowd does not produce. Second, write the falsifier in advance, which is what stops a cascade from becoming self-confirming: the specific observation that would end the thesis, dated, so a later review can see whether it happened. Third, watch the composition of the buying rather than the fact of it — leverage, retail participation, and insider sales are all observable, and they describe how the flow is funded, which is what decides how it behaves when it reverses. The uncomfortable part is that all three are cheap to write and hard to write in the middle of a cascade, because the cascade’s defining property is that the evidence feels overwhelming precisely when it is emptiest. That is why the work belongs in the plan and in advance, and why it pairs with the outside view from the advanced rung: classify the situation, find how similar situations actually resolved, and only then read the story that is being told about this one. • Attach each claim to a fact that is not the price. • Write the falsifier, dated, before the position exists. • Observe how the flow is funded — leverage, retail share, insider sales — not merely that it exists. • Ask whether the position still makes sense if nobody else holds it. • Treat consensus as a fact about the price, not as an argument about the asset. A cascade is not detectable from the inside by intensity of feeling: the strongest conviction is produced by the latest stage. The outside view is the only reliable instrument, because it is the only one that does not read the current story for evidence.

The mechanical crowd: flows that cannot look at the price

Most of the herding in this lesson is psychological — people inferring information from other people’s behaviour. There is a second kind that is purely mechanical, and it matters more for a modern portfolio because its demand is **price-insensitive by construction**. An index fund tracking a benchmark must buy the constituents in proportion to their weight, on the day they enter the index, whatever the price is. It is not following a crowd or reading a signal; it is executing a published rule. The same applies to a target-date fund rebalancing on a schedule and to a fund tracking a sector whose weight has risen because the sector rose. When a substantial share of the market’s assets are held by vehicles of this kind, a stock can rise because its weight rose, and its weight can rise because it rose. The observable consequence is that index events have become events in themselves. A stock added to a major index is repriced in anticipation of the forced buying, and the trading volume on the effective date is concentrated into the closing auction for reasons the auctions lesson lays out. The literature on index inclusion finds a permanent increase in price and ownership around these events, not merely a temporary blip — which is a statement about the demand curve for shares being downward sloping. When the largest buyers do not care about the price, the price is set by whoever has to care, and the marginal seller can demand more. Two implications follow for the way you read a rally. The first is that **ownership concentration is now a structural fact rather than a vote of confidence**: a stock widely held by passive vehicles is widely held because of its size, not because anyone decided it was good value, and its holders will not sell it because they became pessimistic. The second is that crowding has become measurable in a way it was not, because the flows are published — ownership by index funds, short interest, fund flow data, and the concentration of the largest holders are all observable. That is a genuine improvement over the era in which crowding had to be inferred from price action, and it is worth using: if the thesis requires the flow to continue, check who is doing the flowing and whether they have a choice. Distinguish the two crowds before acting on either. A psychological crowd can change its mind and reverse. A mechanical crowd has a written rule and will keep buying regardless of the price — which makes the flow predictable and the exit, when it comes, unpriced.

The datasets: what each one measures, and the lag that spoils it

Crowding is measurable, and the useful discipline is knowing what each dataset actually contains before treating it as proof. **Short interest** is reported twice a month by exchanges, with a settlement lag of several business days, so it describes a position that has already been taken and may already be gone; its level is more informative than its change, because a squeezed short base is a fact about who still has to buy. The **borrow fee** — the rate a short seller pays to borrow the shares — is the closest thing to a real-time measure of crowding on the short side, because it is set by supply and demand for the loan and moves within a day; a name whose borrow rate has gone from fifty basis points to twenty percent has told you more about crowding than any position report. **Option open interest** by strike is published daily and, read against the underlying price, describes where the market has concentrated its bets and, more importantly, where the dealers who sold those contracts must hedge. None of the three is a forecast; all three describe who is currently exposed and to what. The disclosure datasets are slower and their lag is structural rather than incidental. The **commitments report** from the derivatives regulator gives the breakdown of futures positioning by trader type — producers, swap dealers, managed money — on a weekly basis and only for the futures markets, which makes it a clean read of speculative positioning in commodities, rates and index futures and nothing at all about equities. The **quarterly institutional holdings** filing reveals what large managers owned, but with a reporting lag of up to forty-five days, and it captures long equity positions in U.S.-listed securities only — no shorts, no derivatives, no foreign listings. Using it to infer current positioning is a category error, though using many quarters of it to infer who owns a company structurally is exactly right, which is the use the previous read made of it. **Fund flows** are weekly and real, but they mix the decisions of households, advisers and automated allocation rules, so a large flow can be a view or a rebalancing. Two rules follow from knowing the lags. The first is that **crowding is not a reversal signal**, and the datasets cannot make it one: a crowded position can persist for years, and the correct use of the data is to ask whether your own thesis depends on the crowd continuing to behave as it has — which is the question the cascade read established. The second is that the datasets should be used **against price rather than instead of it**: a name at a new high with a low borrow fee and rising open interest is one where nothing is forced to happen; the same name with a twenty percent borrow fee, heavy call open interest and a rotating holder base is one where the next move may be driven by mechanics rather than by fundamentals, and the distinction is the whole content of the read. The test that keeps it honest is to name what each dataset cannot see — the shorts that report late, the holdings that were closed a month ago, the private funds that never file — and to treat the composite as a rough map of the crowd rather than a census of it. Five crowding measures and how current each one is — Borrow fee: Daily, and the closest to real time — it prices the scarcity of the loan · Short interest: Twice monthly, settled with a lag — the level matters more than the change ← · Option open interest: Daily by strike — a map of concentrated bets and of dealer hedges · Commitments report: Weekly, futures only — clean for commodities and rates, silent on equities ← · Quarterly holdings and fund flows: Lagged by up to 45 days or mixed in composition — structural, not current The composite question is the one the sort lab in this lesson is training: for each observation, what does it prove about who is positioned, and how would I know if it were stale? Attention is not positioning, a survey is not a position, and a filing is a position that may already be closed.

Sources

• Informational cascadesBikhchandani, Hirshleifer & Welch (1992), Journal of Political Economy
• Conformity in an unambiguous taskAsch (1955), Scientific American
• The anatomy of a maniaMinsky (1986), “Stabilizing an Unstable Economy”; Kindleberger, “Manias, Panics and Crashes”