Directed Herding
Research Instrument.

The Directed Herding Research Instrument is an empirical research platform built to study how institutional investors, official information events, and retail attention interact around directed herding — the coordinated tilt of holdings and attention toward specific issuers following identifiable public information.

The instrument does not attempt to prove imitation, intent, manipulation, or causation. It integrates independent public data streams into a single reproducible workspace so that non-causal candidate patterns can be surfaced, documented, and moved into reviewable hypotheses and formal empirical tests.

Every signal is a description of what is publicly observable — not proof of coordination, not a trading recommendation.

Institutional 13F holdings, Federal Reserve and BLS releases, and Wikimedia pageview attention proxies each live in separate systems on separate cadences. Researchers routinely rebuild the same extraction, normalization, and alignment code just to ask a single question.

The instrument standardizes ingestion, storage, alignment, and audit for these sources so that a behavioral-finance question can be asked once and reproduced later. It shortens the distance between an observation and a defensible empirical test, while making every candidate signal traceable to the underlying source record and retrieval timestamp.

Public sources are collected by Supabase Edge Functions on a scheduled cadence and written into a normalized PostgreSQL research database. Analytical routines run inside Supabase to compute overlap metrics, synchronized-change classes, and attention baselines. A read-only research API exposes vetted views to the production interface, which is hosted on Cloudflare Workers. Source code is managed on GitHub; Supabase is the backend, automation, and analytics layer, while Cloudflare Workers hosts the production frontend.

Institutional holdings are drawn from SEC EDGAR Form 13F filings. 13F is a quarterly disclosure reported with regulatory lag, so the weekly automation only polls for new or amended filings and records their acceptance timestamps; no new holdings appear between quarterly filing windows regardless of how often the job runs.

Official information events are collected from Federal Reserve releases (FOMC statements, H.15 rate updates) and Bureau of Labor Statistics releases (CPI, employment). Federal Reserve records are populated in the event registry. BLS collection is configured and scheduled, but no validated BLS observations are currently present in the database. Retail-attention context is drawn from Wikimedia Pageviews as a public firm-level attention proxy — not trading data and not a measure of retail order flow.

Scheduler success (that a job ran on time) and ingestion validation (that the resulting records passed schema, provenance, and content checks) are tracked as separate signals. A green scheduler run does not by itself imply new validated data.

The operational universe contains ten validated 13F filing managers selected under documented governance criteria: market relevance, filing continuity, breadth of holdings, strategy diversity, data quality of prior filings, and balanced representation of active and passive management. This ten-manager set is operational and defensible for platform testing.

Every analysis uses a period-specific common-period denominator rather than assuming all ten managers are always present. For the March 31, 2026 comparison, nine managers qualify: Vanguard's latest loaded report period is December 31, 2025, so it is not eligible for the March 31, 2026 common period. The set of eligible managers is recomputed per period and recorded alongside each result.

The weighted top-10 ranking fields are not yet fully populated. Final publication use will require documented candidate-universe construction, scoring evidence, and explicit inclusion and exclusion decisions before the sample can be presented as a ranked selection.

Ten validated managers in the universe — but each period's analysis reports against the common-period eligible subset for that period.

For each period, the descriptive herding signal is computed on the common-period eligible managers for that period. Portfolio similarity is measured with the Jaccard index over issuer sets, and holdings changes are categorized into transparent classes — coordinated accumulation, coordinated reduction, coordinated entry, and coordinated exit — using explicit thresholds recorded in the methodology log.

Three descriptive quantities drive the signal. Participation is defined as dominant-direction managers divided by eligible managers. Consensus is defined as dominant-direction managers divided by directionally classified managers. The base signal is participation × consensus. Every threshold, window, and rule change is versioned in the methodology log so results remain reproducible.

These outputs are candidate synchronizations — descriptions of what is publicly observable across the eligible sample in a given period. They are not causal proof of herding, imitation, intent, coordination, or manipulation.

The active attention pilot uses Wikimedia Pageviews on issuer articles as a public firm-level attention proxy. It is explicitly not a validated retail-trading proxy: it does not measure retail order flow, execution, sentiment, or intent.

Coverage is currently partial across the security universe. Issuer-to-article mappings are generated by automated procedures and still require human validation before they can be treated as authoritative. Rolling baselines and deviation thresholds are recorded in the methodology log.

Event and attention context are not yet temporally linked to the March 2026 institutional episodes in the database. Alignment across holdings, events, and attention on a common time axis remains a required build step before those layers can be jointly interpreted.

The production interface presents candidates in a screening console, then supports the researcher through a structured workflow. Each candidate can be promoted into a hypothesis with stated preconditions and expected observations. Hypotheses can be linked to experiments that run defined queries against the research database and record their outputs alongside the methodology version used.

Findings are consolidated into research briefs that include the underlying candidate, the linked hypothesis, the experiment results, the applicable limitations, and the exact source records consulted. Every brief is reproducible from the stored methodology log.

The instrument is a research environment, not a trading system. Every surfaced pattern is a descriptive candidate synchronization — not proof of imitation, intent, manipulation, coordination, causation, or investment advice. Only non-causal language is permitted in the interface (associated with, followed by, coincided with, aligned with, diverged from).

Sampling, ingestion cadences, thresholds, and analytical choices are documented in a versioned methodology log. Changes to the eligible-manager logic, denominators, or thresholds require an explicit log entry so historical results can be interpreted under the rules that produced them.

The platform is research-ready for descriptive screening and hypothesis generation across the ten-manager validated universe using period-specific common-period denominators. It is not yet sufficient for causal claims or for publication as a finished empirical result.

Remaining requirements before publication use include the peer-reviewed literature review, null-model calibration and permutation testing, verified event and attention linkage on a common time axis, metadata enrichment, greater historical depth, and robustness checks across specifications.