Herding

Definition

Herding is following the crowd’s action over one’s own information — and the theory’s uncomfortable point is that it is often INDIVIDUALLY RATIONAL: information cascades (Banerjee 1992 — when predecessors’ actions carry information, copying can beat one’s own noisy signal) and career risk (Scharfstein-Stein 1990 — professionals are graded relative to peers, so failing conventionally beats failing alone) both produce herds without stupidity. The platform therefore treats herding as STRUCTURE to measure — crowding — rather than a character flaw to lecture about.

How it works / structure

• The two rational engines: cascades (each actor rationally weights the crowd’s revealed information, eventually drowning private signals — fragile by construction, since the crowd’s information stops accumulating once copying starts) and reputational herding (deviating and failing is career-fatal; herding and failing is “the market”; the payoff matrix herds professionals harder than retail).
• The measurable footprints: crowded institutional positions (sent-13f-holdings hotel names), one-sided flows (sent-fund-flows), positioning extremes (sent-cot-reports), attention concentration (sent-news-social), and correlation spikes within factor/theme cohorts (port-correlation-budgets — herded books fail together).
• The fragility mechanics: cascade-built prices embed little independent information — small contrary news can reverse the whole chain (the theoretical basis for why crowded trades unwind violently rather than smoothly); documented in the quant unwind (2007) and degrossing episodes.
• Engine-executable form: crowding composites per name/ theme (ownership concentration + flow percentile + attention percentile), with the platform’s standing rule that crowding scales EXIT-liquidity risk, not thesis validity — a crowded thesis can be right and still be dangerous.

When it applies

Position risk assessment (crowding as a sizing haircut — the thesis may be shared for good reasons; the exit will be shared for bad ones); regime reading (herd-driven trends persist beyond fundamentals then break discontinuously — both halves matter); contrarian setups (extremes + catalyst, never extremes alone — sent-cot-reports discipline); platform-internal diversity monitoring (agent strategies converging onto the same trade is the in-house version).

Risk profile & failure modes

• Contrarian romanticism: fading every crowd loses — most crowds are right for most of the move (momentum’s documented profitability IS partly herd persistence, strategy-momentum); the exploitable moment is extreme + fragility + catalyst.
• Crowding blindness in “independent” analysis: shared data, shared tools, and shared training produce herds of independent-feeling conclusions — diversity of PROCESS, not of self-image, is the check.
• Unwind correlation: crowded positions convert single-name theses into factor exposure at the worst time (port-correlation-budgets stress correlations).
• Career-risk import: agents graded against benchmarks inherit reputational herding mechanically — grading design is a herding parameter.

Evidence & limits

Banerjee and Scharfstein-Stein are the founding theory; empirical herding measures (LSV herding statistics, institutional clustering) document moderate herding that intensifies in stress; unwind episodes (2007, 2021) are documented. Crowding MEASUREMENT is reliable; timing herd breaks is not — the platform uses crowding for sizing and liquidity, not for entries.

Falsifiable-thesis examples

Illustrations only, not signals:

• “Top-crowding-composite names will show higher downside beta than their sector in the next 5%+ index drawdown” — falsified by the episode cross-section.
• “This platform’s active strategies hold pairwise position overlap below 30% (internal diversity check)” — falsified by the overlap audit.

Cross-references

• The footprint gauges: sent-13f-holdings, sent-fund-flows, sent-news-social, sent-cot-reports
• The portfolio consequence: port-correlation-budgets
• The persistence it feeds: strategy-momentum; the extrapolation companion: bias-recency