Equity Penetration Explained: From Graph Modeling to Beneficial Owner Identification

1. Why Equity Penetration Matters: From “Who Holds Shares” to “Who Benefits”

In scenarios such as account opening compliance, anti-money laundering (AML), supplier onboarding, and group governance, a frequently asked question is: who actually calls the shots behind this company? Business registration data typically discloses only direct shareholders. When a shareholder is itself a company, direct ownership data reveals only the “first layer.” Equity penetration solves exactly this problem: by tracing upward through multi-layer ownership chains, converting ownership ratios layer by layer, and ultimately reconstructing the real control chain and aggregate equity held by natural persons or source entities in the target company.

Typical applications include:

1. Compliance: Identifying the ultimate beneficial owner (UBO) who holds at or above a certain threshold—the commonly cited industry standard is 25%—to satisfy AML and due diligence requirements.
2. Risk control: Detecting hidden related-party credit exposure, circular guarantees, and related-party transactions disguised as arm’s-length deals.
3. Due diligence: Assessing whether multiple counterparties have undisclosed relationships with one another.
4. Group management: Clarifying complex group ownership hierarchies to support consolidated reporting and management decisions.

2. Data Modeling: Turning Equity Relationships into a Graph

Equity penetration is inherently a graph problem in terms of data structure. The minimal viable model requires just one relationship table:

equity_edge(holder_id, company_id, ratio, contribution_type, data_date)

Nodes fall into two categories: company nodes and natural-person nodes. Edges represent ownership relationships, and the weight is the ownership ratio. In practice, three additional categories of information are typically needed:

• Position edges: Legal representative, director, supervisor, senior management, and similar roles, enabling “equity + position” dual-dimension relationship analysis.
• Control type: Distinguishing equity control, contractual control, concerted-action arrangements, and other mechanisms, since some control relationships are not reflected in ownership ratios.
• Time dimension: Retaining ownership change history to support point-in-time snapshot penetration rather than only the current state.

What often determines project success is not the algorithm but data governance. Common issues in practice include multiple name variants for the same entity, duplicate names among natural persons, and conflicts between historical and latest registration records. Entity resolution must come first—establishing unified master data for companies and natural persons, then attaching ownership relationships to that master data. Otherwise, penetration results suffer from hard errors such as the same shareholder being counted as two different people.

3. Core Algorithm: Three Principles of Penetration Calculation

3.1 Multiplication Principle: Calculating Indirect Ownership

The industry-standard approach is path multiplication: if A holds 60% of B and B holds 50% of C, then A’s indirect ownership of C is 0.6 × 0.5 = 30%. If A also directly holds 20% of C, A’s aggregate equity in C is 50%. It is important to note that “aggregate equity” and “actual control” are two distinct concepts: equity conversion uses multiplication, while control determination typically examines whether each layer exceeds 50%, whether the entity is the largest shareholder, or whether contractual arrangements exist. The two must not be conflated.

3.2 Cycle Handling: Cross-Holdings Must Not Cause Infinite Loops

Real data contains cross-holdings where A holds B and B holds A, and potentially longer cycles. Traversal must perform path-level deduplication: using the “current search path” as the unit of deduplication, truncating as soon as a node appears in the current path to avoid infinite loops and duplicate ratio accumulation. For equity within a cycle, the common business practice is approximate treatment or separate annotation rather than simply discarding it.

3.3 Threshold Pruning and Beneficial Owner Determination

Full-graph traversal is expensive, so threshold pruning is widely adopted in engineering: when the cumulative ratio of a path falls below a certain floor (e.g., 1%), expansion stops, because further penetration contributes negligibly to the final result. UBO determination typically uses the common standard of “aggregate equity of at least 25%.” When this is not met, business rules provide a fallback—for example, taking the largest shareholder or actual controller as the concluding result, with the reason for the fallback explicitly noted.

4. Code in Practice: From Recursive Functions to Recursive SQL

Below is pseudocode demonstrating the core penetration logic (multi-level indirect ownership with cycle protection):

# Pseudocode: multi-level indirect ownership calculation
function calc_equity(entity, path):
    if entity in path:                 # Cycle protection: truncate if already in current path
        return {}
    path = path + {entity}
    result = {}
    for (company, ratio) in edges_from(entity):
        result[company] += ratio       # Direct ownership
        for (sub, sub_ratio) in calc_equity(company, path):
            result[sub] += ratio * sub_ratio   # Indirect ownership via path multiplication
    return result

If the data resides in a relational database, a recursive CTE can perform upward penetration (finding all upstream shareholders of a company and their converted ratios):

WITH RECURSIVE up AS (
    SELECT :company_id AS holder_id, 1.0 AS path_ratio
    UNION ALL
    SELECT e.holder_id, up.path_ratio * e.ratio
    FROM up
    JOIN equity_edge e ON e.company_id = up.holder_id
    WHERE up.path_ratio * e.ratio >= 0.01   -- Threshold pruning: stop below 1%
)
SELECT holder_id, SUM(path_ratio) AS total_ratio
FROM up
GROUP BY holder_id
HAVING SUM(path_ratio) >= 0.25;             -- UBO determination threshold

Both code snippets are intentionally kept short. A real system would also need path recording (for result explainability), multi-version data switching, concurrency control, and caching.

5. Four Key Considerations for Production Deployment

1. Performance: Equity graphs of large groups can reach millions or even tens of millions of nodes. Common techniques include threshold pruning, on-demand lazy loading, caching of frequently accessed subgraphs, and an architecture combining “offline full pre-computation + online incremental updates.”
2. Data timeliness: Registration data has inherent delays. Every ownership edge should carry a data date, and conclusions should be annotated with “based on data as of a certain point in time” to prevent misinterpretation.
3. Configurable rules: Different regulatory and business contexts use different UBO determination criteria (25%, 50%, largest shareholder, etc.). Thresholds and fallback rules must be configuration items, not hard-coded.
4. Explainable results: Risk and compliance personnel need to know “why this person was determined to be the UBO.” Each conclusion should therefore include the complete ownership path and ratios at each layer to support manual review and visualization.

If building a graph platform in-house is costly, general-purpose cloud graph database and knowledge graph products can be used to quickly establish the foundation (Baidu AI Cloud, for example, offers such general capabilities), allowing teams to focus their efforts on data governance and business rules.

6. Going Further: From Equity Penetration to Enterprise Knowledge Graphs

The natural destination of equity penetration is the enterprise knowledge graph. Once equity, positions, guarantees, bidding, litigation, and other relationship types are unified in a single graph, the analytical perspective can be upgraded from “a single equity dimension” to “multi-dimensional relationship risk mining”—for example, identifying that “two seemingly unrelated companies actually share the same ultimate controller.”

The development of large language models has further lowered the barrier to use: through a combination of “graph retrieval + LLM,” business users can ask questions in natural language, such as “map out this company’s ultimate beneficial owners and ownership paths,” and the system automatically performs graph queries, summarizes results, and generates a draft risk report. But the division of labor must be emphasized: the LLM handles intent understanding and language organization, while ownership ratio calculation and UBO rule determination must be performed by deterministic algorithms—only then are results auditable and reviewable.

Conclusion

Equity penetration may appear to be nothing more than a “multi-table recursive query,” but achieving accuracy, explainability, and evolvability tests a combination of data modeling, graph algorithms, and engineering capabilities. Getting the model, algorithm, and governance right first—before pursuing advanced applications with knowledge graphs and LLMs—is the prudent path to production.