Beneficial Ownership Penetration: From Data Modeling to Algorithm Implementation

.### 1. What Problems Does Ownership Penetration Solve?

Ownership penetration refers to the process of starting from a target enterprise and tracing up the shareholder chain layer by layer to identify ultimate natural person shareholders, actual controllers, or specific entity types (such as state-owned capital investors). It calculates the equity ratio held by these subjects through both direct and indirect means. Fundamentally, it serves as the technical foundation for corporate relationship analysis and beneficial owner identification. Typical application scenarios include:

• Financial Risk Control and AML: Complying with Anti-Money Laundering regulatory frameworks, financial institutions need to penetrate and identify ultimate beneficial owners during account opening, credit approval, and transaction monitoring.
• Investment and M&A Due Diligence: Identifying hidden related parties, nominee structures, and circular shareholding of target enterprises to assess control ownership and avoid misjudgments where seemingly unrelated entities are actually controlled by the same party.
• Supply Chain and Compliance Review: Determining whether suppliers or partners have conflicts of interest, providing a data basis for related-party transaction identification.
• Group Management and Consolidation: Determining actual control relationships through penetration to support the judgment of consolidation scope and control hierarchy.

2. Why is this Technically Challenging?

If there were only a single layer of shareholder relationships, “penetration” would simply be a table lookup. However, real-world corporate equity networks quickly turn this into a complex problem combining graph computation and data governance.

1. Structural Complexity and Multi-path Overlap: Shareholding chains often feature multi-level nesting. If Entity A holds 80% of Company B, and B holds 60% of Company C, A’s indirect holding in C is 80% × 60% = 48%. When the same subject holds equity in a target enterprise through multiple paths of varying lengths, the results of all paths must be summed. Missing any single path will lead to an underestimation of the ratio.
2. Circular Shareholding: When a closed loop like A → B → C → A occurs, naive recursive traversal will either loop infinitely or accumulate ratios repeatedly. Explicit cycle detection and pruning must be implemented at the algorithmic level.
3. Control Does Not Equal Ownership Ratio: Through arrangements like voting rights entrustment, concerted action agreements, and dual-class shares, a party with a lower ownership ratio might still constitute actual control. A single “ratio threshold” rule is prone to misjudgments and requires multi-criteria evaluation.
4. Inconsistent Data Quality: Enterprise names may have abbreviations, former names, or typos. There are time lags in registration information updates. Capital contribution ratios and voting rights ratios might be separated. Disambiguating natural persons with the same name is highly challenging. Arguably, half of the engineering effort in penetration analysis is spent on data governance.
5. Criteria Must Be Predefined: At which layer to stop penetrating (natural person, entity without shareholders, or state-owned entity), whether to calculate “beneficial rights” or “voting rights”, subscribed versus paid-in capital, and the precision of the ratio—these must be clearly defined as rules during the requirements phase. Otherwise, the results will be neither comparable nor auditable.

3. Data Modeling: Abstracting Equity Relationships into a Directed Graph

The most natural engineering modeling approach is a graph model, with core elements as follows:

• Nodes: Two types of subjects—enterprises and natural persons. Node attributes include a unique identifier (for enterprises, prioritizing a unified social credit code), a standard name, and subject type.
• Directed Edges: Pointing from “shareholder” to “invested enterprise”. Edge attributes include direct shareholding ratio, subscribed capital amount, shareholding relationship type, data source, and update time.
• Auxiliary Tables: Name alias mapping tables (including former names and abbreviations) and historical change records, used for entity alignment and audit tracing.

The directional convention must be globally unified. In practice, the “Shareholder → Invested Enterprise” direction is recommended: upward penetration traverses against the edge direction, while downward checking of a subject’s controlled group traverses in the forward direction. Both queries can reuse the same data. A sample data structure:

{
  "nodes": [
    { "id": "C001", "type": "enterprise", "name": "XYZ Tech Co., Ltd." },
    { "id": "P001", "type": "person",    "name": "John Doe" }
  ],
  "edges": [
    { "from": "C001", "to": "E001", "ratio": 0.6 },
    { "from": "P001", "to": "C001", "ratio": 0.8 }
  ]
}

For storage selection, small-scale scenarios can be supported by a relational database with an adjacency list. When the scale reaches tens of millions of entities and requires high-frequency multi-hop queries, a graph database or distributed graph computation based on general big data components is more appropriate. The key is not to let storage selection hold your business logic hostage.

4. Core Algorithms: Traversal, Ratio Calculation, and Cycle Handling

4.1 Path-Based Traversal Method

Starting from the target enterprise, perform a Breadth-First Search (BFS), accumulating ratios layer by layer, and converge the result when encountering a natural person or triggering a termination condition. The simplified logic is as follows:

def penetrate(target, graph, max_depth=10):
    # result[subject_id] = accumulated shareholding ratio of this subject to the target
    result = {}
    # Queue elements: (current_node, accumulated_ratio, visited_path, remaining_depth)
    queue = [(target, 1.0, {target}, max_depth)]
    while queue:
        node, ratio, visited, depth = queue.pop(0)
        for holder, r in graph.get_shareholders(node):
            new_ratio = ratio * r
            if holder.type == "person":            # Termination condition: natural person
                result[holder.id] = result.get(holder.id, 0) + new_ratio
                continue
            if holder in visited or depth == 0:     # Pruning: cycle detected or depth limit exceeded
                continue
            queue.append((holder.id, new_ratio,
                          visited | {holder.id}, depth - 1))
    return result

The advantage of this method is its intuitiveness and strong interpretability. Every conclusion can retain a complete path detail, naturally meeting audit requirements. The disadvantage is that the number of paths grows exponentially with depth and branching factor, so it must be used in conjunction with a depth limit, visited deduplication, and path merging strategies.

4.2 Matrix-Based Iterative Method

Organize direct shareholding ratios into a matrix M (where M[i][j] represents the direct shareholding ratio of subject i in enterprise j). Mathematically, indirect shareholding forms a geometric series: M + M² + M³ + …. When the shareholding ratios in each row are strictly less than 1, the series converges and can be iteratively calculated until the difference between two adjacent rounds is less than a threshold. This approach is more friendly for large-scale data and naturally converges to a reasonable result even with circular shareholding. The drawback is weaker interpretability compared to the path method, requiring an additional path backtracking mechanism to meet the detailed requirements of compliance audits. A common combination in production systems is using the matrix method for bulk pre-computation and the path method for single-point queries and result interpretation.

4.3 Actual Controller Determination

After obtaining the ratio results, determining the actual controller requires overlapping business rules: calculating both beneficial rights and voting rights; incorporating relationships like concerted action agreements and voting rights entrustment as special edges into the graph; and setting thresholds combined with heuristic rules like “single largest holder” and “significant lead” to output candidate controllers and confidence explanations, rather than just giving a binary conclusion.

5. Engineering Implementation: A Usable Penetration Analysis Pipeline

1. Data Ingestion and Sync: Use officially published corporate registration information as the primary source, supplemented by annual reports and other public data. Establish incremental sync and change detection mechanisms, recording the collection time for each data entry to attach a “data freshness” label to the results.
2. Cleaning and Entity Alignment: Use the unified social credit code as the primary key for enterprises; for entities without codes (like natural persons), use composite keys like “Name + ID Type + ID Number”. Establish name normalization rules (removing bracketed notes, unifying full/half-width, mapping abbreviations) and disambiguate subjects with the same name. The accuracy of entity alignment directly determines the upper limit of penetration results.
3. Graph Construction and Servitization: Write cleaned relationships into graph storage, providing two types of capabilities upward—real-time penetration queries with caching and rate limiting, and offline batch pre-computation (pre-generating penetration results for high-frequency subjects, invalidated by change events).
4. Quality Validation: Validate ratio domains (0~100%), detect abnormal shareholder ratio totals, monitor orphan nodes, and alert on sync delays. Expose “broken data” as early as possible rather than leaving it for the business side to discover.
5. Visualization and Reporting: Render equity structure trees/graphs, highlight key paths, and enable one-click result export, allowing business personnel to quickly locate anomalies during the review process.

If deploying in the cloud, you can combine general cloud capabilities like container platforms, object storage, message queues, logging services, and monitoring alerts (such as the corresponding services provided by Baidu AI Cloud) on demand, focusing your in-house development efforts on data governance and algorithm logic.

6. How Can Large Language Models Help?

Recent advancements in LLM capabilities provide new auxiliary tools for penetration analysis, but they must be used in the right places:

• Unstructured Information Extraction: Extracting relationship triplets of “who holds whom and at what ratio” from texts like announcements, annual reports, and legal documents. This is an area where LLMs excel and can significantly broaden data sources.
• Entity Normalization Assistance: Determining the correspondence between entities like “Company Beijing Branch” or “Company (former name)” and standard subjects, aiding entity alignment decisions.
• Result Summarization and Report Generation: Organizing penetration results into natural language conclusions, lowering the understanding barrier for the business side.

However, boundaries must be drawn: deterministic computations like ratio multiplication and path searching must be handled by algorithms. Having an LLM directly “mental math” shareholding ratios is highly error-prone. Relationships extracted by the model must also undergo entity existence validation and ratio domain validation before being written to the graph. Evaluating whether an LLM is suitable for such tasks should not rely solely on public benchmark scores, but rather on testing extraction accuracy, recall, and hallucination rates on an in-house business evaluation set. Good benchmark performance does not equal good task performance—building an evaluation set before model selection is essential to avoid pitfalls.

7. Common Pitfalls and Best Practices

• Do Not Rely on a Single Data Source: Use multi-source cross-validation. In case of conflicts, provide confidence markers rather than silently overwriting.
• Abnormal Ratios are Signals: Results exceeding 100% or showing negative values likely indicate unhandled cycles or data errors. Investigate before storing.
• Explicitly Declare Criteria: Beneficial rights/voting rights, subscribed/paid-in, direct/indirect—the output results must be labeled with the specific criteria used, otherwise subsequent reconciliation is impossible.
• Retain Path Details: Every ratio conclusion should be traceable to a specific path. This is the baseline requirement for auditing and compliance.
• Pragmatic Penetration Depth Limit: In practice, a limit of within ten layers is commonly used, combined with the rule of “terminate when reaching a natural person”, balancing effectiveness and cost.

8. Conclusion

On the surface, ownership penetration is simply “tracing shareholders upward”, but essentially it is an engineering problem of graph traversal and ratio computation, overlaid with three types of challenges: entity alignment, data quality, and business criteria. Abstracting subjects and shareholding relationships into a directed graph, entrusting ratio calculations to deterministic algorithms, utilizing LLMs in appropriate places like information extraction and normalization, and retaining traceable path details for every conclusion—these are the steps required to build a fast, accurate, and audit-compliant penetration analysis capability.