Research Note

When Risk Language Meets Crowded Ownership: A Research Guide to 10-Ks, 13Fs and Downside Risk

Open in an AI assistant with a suggested prompt

Preview prompt

Summarize the Ionitsa research note titled "When Risk Language Meets Crowded Ownership: A Research Guide to 10-Ks, 13Fs and Downside Risk" for a technical reader.
Cover the problem or research question, implementation or method, evidence or results, and limitations.
Separate facts stated on the page from your own assessment, note anything unclear or unverified, and avoid promotional language.

Primary source: https://ionitsa.com/research/disclosure-crowding-downside-risk-guide.md
Canonical page: https://ionitsa.com/research/disclosure-crowding-downside-risk-guide/

A practical research design for testing whether changes in corporate risk disclosures become more consequential when a stock sits inside a crowded institutional ownership network.

Markets · Published 4 August 2026 · Updated 4 August 2026 · 12 min read

10-KForm 13FTextual AnalysisInstitutional CrowdingNetwork AnalysisDownside Risk

A company can tell investors that its risks have changed. A portfolio manager can still ignore the disclosure. A crowded ownership structure can then turn a normal repricing into a disorderly exit.

Those are three different objects: information, attention and market structure. Most studies isolate one of them. This guide explains how to combine them into a research design that is useful even if the final result is a null.

The central question is:

Do changes in a company’s 10-K risk-factor language predict more severe downside outcomes when the stock is held by a concentrated, overlapping network of institutional investors?

This is a research-design guide, not a claim that the interaction produces tradable alpha. The reusable implementation lives in the investment-management-toolkit.

Why combine disclosures and holdings?

A textual signal and a crowding signal describe different parts of the same possible event.

A disclosure change says management altered the description of the risks.
A holdings network says which investors may be exposed to the same securities.
A downside outcome says whether the information and positioning mattered in the market.

Language alone does not reveal who owns the stock. Ownership alone does not reveal whether the underlying risk changed. The interaction asks whether a risk signal becomes more economically important when the holder base is capable of transmitting selling pressure.

Research on 10-K changes already provides a serious benchmark. Cohen, Malloy and Nguyen’s Lazy Prices finds that active changes in mandatory filings contain information about future firm outcomes and returns. Risk-factor additions and removals have also been studied in relation to the variance risk premium.

The ownership side has its own literature. Coval and Stafford show how forced mutual-fund transactions can create price pressure in commonly held securities in Asset Fire Sales. Network research models common holdings as potential contagion channels; one example is the statistically validated portfolio-overlap network.

The useful contribution is therefore not “text predicts returns” or “crowding is risky.” It is a disciplined test of whether the two mechanisms reinforce each other.

Turn the idea into falsifiable hypotheses

A research project becomes clearer when each claim can fail.

Hypothesis 1: disclosure novelty and uncertainty

Firms with larger year-over-year changes in Item 1A experience greater subsequent uncertainty than firms with stable disclosures, after controlling for ordinary firm characteristics and the amount of text.

Possible outcomes include:

idiosyncratic realised volatility;
option-implied volatility or variance risk premium;
downside semivariance;
maximum drawdown;
abnormal trading volume;
credit-spread changes; and
future adverse corporate events.

Hypothesis 2: crowding and fragile liquidity

Stocks with concentrated ownership or high exposure to overlapping institutional portfolios experience more severe downside moves during negative shocks.

Hypothesis 3: the interaction

The association between disclosure novelty and downside outcomes is stronger when ownership is more crowded.

A simple panel specification is

Y_{i,t+h} =\alpha_i+\gamma_t +\beta_1N_{i,t} +\beta_2C_{i,t} +\beta_3N_{i,t}C_{i,t} +\boldsymbol{\delta}^{\mathsf T}\boldsymbol{X}_{i,t} +\varepsilon_{i,t+h},

where $N$ is disclosure novelty, $C$ is crowding, $Y$ is a future downside outcome and $\boldsymbol{X}$ contains controls. Firm fixed effects $\alpha_i$ remove persistent issuer characteristics; time fixed effects $\gamma_t$ remove market-wide conditions.

The coefficient of interest is $\beta_3$ . It should be defined before looking at the results. Trying many text measures, crowding measures and horizons and then reporting the strongest interaction would turn a clean hypothesis into a search for a lucky t-statistic.

Build a point-in-time data model

The biggest risk in this study is not the language model. It is the clock.

Every observation needs at least two dates:

Field	Meaning
`period_end`	The fiscal or holdings date described by the record
`filed_at`	When the filing became public
`available_at`	When the processed feature could realistically enter the strategy
`outcome_start`	First market observation allowed after availability

A December year-end 10-K filed in February was not known in December. A Form 13F describes quarter-end holdings but may be filed weeks later. Joining either signal to returns from the period end creates look-ahead bias.

The SEC’s EDGAR APIs provide filing histories and identifiers without an API key. The SEC also publishes flattened quarterly Form 13F datasets, but explicitly warns that they are derived from as-filed submissions and are not a substitute for checking the filings themselves.

A robust raw layer should preserve:

filing accession number
form type and amendment flag
CIK and manager identifier
period end and acceptance timestamp
original filing URL
retrieval timestamp and content hash
raw document or structured row
parser version

Amendments should not silently overwrite the original state. A restated filing is a new event with its own availability time.

Extract the disclosure you actually intend to study

“The 10-K” is not one homogeneous document. Item 1A is the risk-factor section. Extraction should begin with filing structure and headings, not an arbitrary character range.

A defensible extraction pipeline should:

retrieve the filing identified by accession number;
remove tables, navigation and repeated headers without deleting sentences;
locate Item 1A using multiple heading patterns;
stop at the next valid item boundary;
retain the original and cleaned text;
record extraction confidence and failure reasons; and
manually review a stratified sample of filings and years.

False boundaries can create artificial novelty. A parser that captures half of Item 1A in one year and all of it in the next year will report a dramatic change that came from code, not management.

The study should also distinguish genuine annual reports from amendments and transition reports. If the research universe includes non-US issuers, form and disclosure conventions need separate treatment.

Start with transparent text measures

Sophisticated embeddings should not be the first model. Begin with a baseline that can be inspected.

For document-term vectors $\boldsymbol{x}_{i,t}$ and $\boldsymbol{x}_{i,t-1}$ , cosine novelty is

N_{i,t} =1- \frac{\boldsymbol{x}_{i,t}^{\mathsf T}\boldsymbol{x}_{i,t-1}} {\lVert\boldsymbol{x}_{i,t}\rVert_2 \lVert\boldsymbol{x}_{i,t-1}\rVert_2}.

A practical baseline stack is:

Feature	What it captures	Main failure mode
TF-IDF cosine novelty	Broad language change	Formatting and boilerplate
Added/removed sentence share	Direction of revision	Sentence-boundary errors
Financial negative-word share	Tone	Context and negation
Section length change	Disclosure expansion	Regulatory or template effects
Topic-share change	Type of risk	Topic instability
Embedding distance	Semantic change	Model drift and weak interpretability

Keep additions and removals separate. “We face a material refinancing risk” and the deletion of that sentence can have the same absolute distance but opposite interpretations.

The text model must be fit point in time. A TF-IDF vocabulary built from the full 2006–2026 corpus leaks future language into early observations. The same rule applies to topic models and embedding fine-tuning.

Represent Form 13F as a bipartite network

Let $W_{m,i,t}$ be manager $m$ ’s portfolio weight in security $i$ at time $t$ . The holdings data form a bipartite graph:

manager ── portfolio weight ── security

Two managers’ portfolio overlap can be measured with cosine similarity:

O_{m,n,t} =\frac{\sum_i W_{m,i,t}W_{n,i,t}} {\sqrt{\sum_i W_{m,i,t}^2}\sqrt{\sum_i W_{n,i,t}^2}}.

At the security level, ownership concentration can be measured by a Herfindahl index. If $s_{m,i,t}$ is manager $m$ ’s share of the observed institutional exposure to security $i$ ,

HHI_{i,t}=\sum_m s_{m,i,t}^2.

Useful crowding features include:

number of reporting managers;
ownership HHI;
effective manager count $1/HHI$ ;
aggregate portfolio weight across managers;
average overlap of the security’s owners;
concentration among the largest managers;
quarter-over-quarter ownership change; and
a security’s centrality in the holdings network.

These measures are not interchangeable. A security can have many owners but be dominated by three large funds. Another can have concentrated ownership among funds whose remaining portfolios barely overlap.

What 13F does not reveal

Form 13F is not a complete live position book. It is delayed, primarily covers reportable long positions, omits many derivatives and shorts, and can contain amendments or confidential-treatment effects. The reported market value is not an execution timestamp.

The guide’s use of 13F is therefore structural: it estimates the disclosed ownership network as of the latest public filing. It should not be described as real-time crowding.

Join the signals without inventing contemporaneity

A defensible event clock looks like this:

latest public 13F snapshot
          ↓
10-K accepted by EDGAR
          ↓
text processed after acceptance
          ↓
signal becomes available
          ↓
outcome window begins next tradable period

For each 10-K event, attach the latest 13F snapshot that was public before the 10-K signal became available. Do not attach the holdings report with the closest quarter end if it had not yet been filed.

Standardise novelty and crowding within date or industry-date groups before forming the interaction. This prevents a secular rise in document length or institutional ownership from masquerading as a cross-sectional signal.

The toolkit exposes the building blocks explicitly:

from investment_research_toolkit import (
    build_disclosure_crowding_panel,
    disclosure_novelty,
    manager_portfolio_weights,
    portfolio_overlap,
    security_crowding,
    tfidf_from_counts,
)

tfidf = tfidf_from_counts(document_term_counts)
novelty = disclosure_novelty(tfidf.loc["current"], tfidf.loc["previous"])

weights = manager_portfolio_weights(holdings)
manager_network = portfolio_overlap(weights)
crowding = security_crowding(weights)

The package does not download filings or claim an empirical result. It keeps the feature calculations reusable while the study-specific data contract remains visible.

Choose outcomes that match the mechanism

If the proposed mechanism is fragile liquidity, a one-year raw return is a blunt outcome. Use a hierarchy:

Market reaction

abnormal return from the next close;
downside semivariance over 5, 20 and 60 trading days;
maximum drawdown;
realised idiosyncratic volatility;
abnormal volume and illiquidity; and
option-implied downside skew where available.

Ownership response

next public quarter’s change in institutional ownership;
breadth of manager selling;
persistence of the disclosed ownership network; and
changes in overlap among remaining holders.

Fundamental outcome

earnings deterioration;
credit-rating action;
CDS or bond-spread widening;
distress or bankruptcy event; and
subsequent risk-factor additions.

The return window must begin after the signal is available. Longer outcomes should use overlapping-window inference or non-overlapping event rules.

Controls and identification

The interaction can correlate with familiar firm risks. At minimum, test controls for:

size and book-to-market;
momentum and recent volatility;
liquidity and trading volume;
leverage and profitability;
analyst coverage;
institutional ownership level;
document length and prior novelty;
industry; and
calendar time.

Cluster standard errors at a level consistent with the residual dependence. A firm panel commonly requires firm clustering; market-wide events can justify time clustering or two-way clustering.

This design remains observational. A significant $\beta_3$ does not prove that crowded owners caused the downside move. Stronger identification could use plausibly exogenous changes in institutional ownership, index membership or manager structure, but every instrument brings its own assumptions.

The right language is “associated with” unless the design genuinely identifies a causal effect.

Validation should be harder than the model

Separate statistical association from strategy evaluation.

Explanatory panel

Freeze the feature definitions.
Estimate the panel specification.
Report coefficient magnitude and confidence intervals.
Test industry and time subsamples.
Show whether the interaction survives the component signals.

Predictive evaluation

Use expanding or rolling chronological training windows.
Fit text vocabularies and scalers inside each training window.
Select model parameters without touching the final test period.
Evaluate calibration, rank information coefficient and tail classification.
Compare with novelty-only, crowding-only and simple characteristic baselines.

Portfolio experiment

Delay trading until after signal availability.
Apply liquidity and borrow constraints.
Neutralise broad market and industry exposures where relevant.
Charge turnover, spread, borrow and market-impact costs.
Report event concentration and capacity.

A statistically interesting tail-risk signal may be useful for risk monitoring even if it cannot support a cost-effective long-short portfolio.

Negative controls that can save the project

Good negative controls reveal whether the result comes from the proposed mechanism.

Compare Item 1A with sections not expected to contain risk changes.
Shuffle manager identities while preserving security ownership totals.
Use future 13F crowding as a deliberate leakage check; it should look suspiciously stronger if the pipeline is vulnerable.
Test formatting-only document changes.
Compare the interaction with size × novelty and liquidity × novelty.
Run placebo event dates before the filing.

If the signal appears before the filing, the model has learned leakage, slow data revision or an omitted event—not a reaction to the disclosure.

What would invalidate the thesis?

Reject or narrow the thesis if:

disclosure novelty loses significance after document-quality controls;
crowding adds nothing beyond institutional ownership and liquidity;
the interaction is confined to microcaps that cannot be traded;
results depend on one text model, horizon or crisis;
the event effect starts before public availability;
amendments materially change the signal after the fact; or
costs and borrow erase the economic effect.

A null interaction is still informative. It says textual risk and ownership structure may be separate monitoring dimensions rather than a combined trading signal.

A minimal credible first study

Start smaller than the final ambition:

US common stocks with clean CIK–security mappings.
Item 1A from annual 10-K filings.
TF-IDF cosine novelty plus added-sentence share.
Latest publicly available 13F ownership HHI and owner overlap.
Next-20-day idiosyncratic downside volatility as the primary outcome.
One pre-registered interaction specification.
A chronological holdout containing a different market regime.

Only after that baseline works should the study add embeddings, dynamic graph models, options or causal designs.

The discipline is deliberate. In alternative-data research, the first victory is not a high Sharpe ratio. It is knowing that every observation could actually have been known when the model says it was known.

For the next modelling layer, see Sparse Factors Without Backtest Theatre, which develops the chronological validation and portfolio controls in more depth.