Equity Penetration in Practice: Graph Modeling, Algorithms, and Engineering Implementation

I. Equity Penetration: An Underestimated Graph Computing Problem

Equity penetration refers to the process of starting from a target enterprise and tracing the “who holds whom” ownership relationships layer by layer upwards until reaching terminal nodes that no longer extend further—such as natural persons, listed companies, or state-owned entities. This reveals the complete ownership chain and ultimate beneficial owner. Conversely, downward penetration enumerates all entities actually controlled by a specific subject. This process widely serves scenarios like beneficial owner identification in Anti-Money Laundering (AML) and Know Your Customer (KYC) due diligence, group affiliated party and guarantee circle identification, pre-investment due diligence, audit trail verification, and state-owned asset supervision.

Many misunderstand this as merely “looking up business registration information.” From an engineering perspective, however, it is a serious graph computing problem with five major challenges:

1. Deep Hierarchies: Real corporate group ownership chains often span a dozen or more layers. Manual, layer-by-layer verification is entirely unscalable, making programmatic traversal mandatory.
2. Complex Relationships: Cross-shareholding, circular shareholding, proxy holdings, and voting right delegations are prevalent. Naive “layer-by-layer” searching will fail to converge or even terminate.
3. Poor Data Quality: Public information often contains non-standardized names, multiple spellings for the same entity, and unclosed historical changes. The level of data governance directly determines the accuracy of penetration results.
4. Massive Scale: The number of market entities nationwide reaches hundreds of millions, and the number of equity relationship edges is an order of magnitude higher. Performance optimization is not optional; it is a fundamental requirement.
5. Strict Timeliness: Events like equity transfers, pledges, and freezes occur frequently. Yesterday’s conclusion may be invalid today. Systems must support incremental updates and point-in-time queries.

II. Data Modeling: Building a Trustworthy Equity Graph

2.1 Entity and Relationship Design

The skeleton of the graph is straightforward: nodes are enterprise entities and natural persons (extensible to overseas entities or branches as needed), and edges are shareholding relationships (pointing from the shareholder to the held enterprise). The real complexity lies in the edge properties: shareholding ratio, subscribed capital, relationship valid and invalid timestamps, data source, and confidence level. None can be omitted. Illustrated from a relational perspective:

-- Illustration: Equity relationship table (corresponds to an edge type in a graph database)
CREATE TABLE equity_relation (
  shareholder_id BIGINT      NOT NULL,
  target_id      BIGINT      NOT NULL,
  ratio          DECIMAL(9,6),
  valid_from     DATE,
  valid_to       DATE,
  source_type    VARCHAR(32),   -- Announcement disclosure / Registry filing / Inferred
  confidence     DECIMAL(3,2)
);

2.2 Bi-Temporal Modeling

valid_from / valid_to represents business time, combined with the ingestion time to represent system time. This combination supports “penetration at any historical point in time.” For example, in due diligence, answering “what was the equity structure on the transaction base date?” requires pruning the edge set by business time rather than using the latest snapshot. Without this design, historical retrospective requirements cannot be met.

2.3 Data Governance

• Entity Alignment: Natural persons with the same name must be distinguished using identity identifiers; enterprises before and after renaming must be mapped back to the same entity using a unified code. In practice, a pipeline of “rule-based pre-filtering + similarity scoring + manual final review” is typically employed.
• Confidence Labeling: Ratios disclosed in announcements and ratios derived from inference must be graded differently. The evidence level must be included in the final penetration results for downstream consumption based on risk appetite.
• Change Data Capture (CDC): Shareholder and ratio changes should form an event stream rather than overwriting existing records. Otherwise, once historical information is overwritten, audit and retrospective capabilities are completely lost.

III. Core Algorithms: Reachability, Ratios, and Cycles

3.1 Reachability Penetration

Determining “who is on the equity chain of a subject” is essentially a multi-source search on a graph. The key here is that termination rules must be explicitly defined—traversal stops upon reaching a natural person, a listed entity, or a state-owned entity. Otherwise, the search will not converge in cyclic structures.

3.2 Comprehensive Shareholding Calculation

Direct shareholding and comprehensive (or effective) shareholding are distinct concepts. The latter is approximately the sum of the products of ratios across all shareholding paths. A conceptual implementation is shown below:

from collections import deque
def effective_stake(edges, target, owner, max_depth=15):
    """
    edges: {company: [(shareholder, ratio), ...]}
    Returns the approximate effective stake of 'owner' in 'target' 
    (sum of path multiplications, with cycle pruning)
    """
    total, queue = 0.0, deque([(target, 1.0, {target}, 0)])
    while queue:
        node, acc, path, depth = queue.popleft()
        if depth >= max_depth:
            continue
        for holder, ratio in edges.get(node, []):
            step = acc * ratio
            if holder == owner:
                total += step
            elif holder not in path:  # Cycle pruning to avoid infinite loops in cross-shareholding
                queue.append((holder, step, path | {holder}, depth + 1))
    return round(total, 6)

3.3 Limitations of Approximate Methods and Strict Solutions

Path summation can double-count contributions on cycles in the presence of cross-shareholding, leading to inflated values. A strict approach models the effective shareholding of each entity as a system of linear equations: s = d + R·s, where d is the external direct shareholding vector and R is the shareholding ratio matrix. Solving (I − R)s = d yields the equilibrium solution. Special handling is required for non-invertible matrices caused by pure circular shareholding. A common engineering strategy is to use the approximate method for rapid estimation first, and then switch to solving the linear equations for subgraphs where cycles are detected.

3.4 Circular Shareholding Detection

Rather than hitting cycles one by one during traversal, it is more efficient to use Strongly Connected Components (SCC) algorithms (like Tarjan or Kosaraju) to identify all mutually holding groups at once. Caching these results as graph metadata for reuse in penetration calculations improves efficiency by an order of magnitude.

3.5 Control Rights vs. Economic Rights

The shareholding ratio answers “how much is distributed,” while control rights must also factor in non-equity relationships such as voting right delegations, concerted action agreements, and board seats. Penetrating the “actual controller” requires applying the same graph reasoning over these edges—which is why modeling must reserve the capability for relationship expansion.

IV. Engineering: From Algorithmic Correctness to System Usability

1. Scale and Performance: For subjects with high penetration frequency, perform offline pre-computation and caching. Queries should hit the cache first, falling back to real-time computation only on misses. Equity change events should drive local incremental recalculations to avoid full graph rebuilds. Regarding storage, native graph database traversal generally outperforms recursive queries in relational databases for multi-layer penetration scenarios, but decisions should be based on write throughput and operational costs, evaluating general capabilities rather than blindly trusting a single solution.
2. Timeliness and Consistency: Change events are distributed to computation and caching modules via message queues to ensure consistent query results across all entry points at the same point in time. Penetration results must attach the computation timestamp and data version number; otherwise, different systems will yield contradictory answers.
3. Explainability and Compliance: Every penetration conclusion must retain the complete path, the source snapshot of each edge, and the confidence level. During regulatory inquiries and audit evidence collection, “how the conclusion was reached” is often more important than the conclusion itself, continuously driving investment in data governance.
4. Presentation Layer: A penetration graph with dozens of layers and thousands of nodes cannot be rendered all at once. Common practices include collapsing by hierarchy, filtering by ratio thresholds for aggregated display, supporting progressive drill-down expansion, and linking the evidence panel with the graph view.

V. The Auxiliary Role of LLMs in Equity Penetration

Large Language Models (LLMs) do not replace graph computing, but they significantly reduce the cost of “turning the world into a graph.” There are four main application areas:

1. Unstructured Information Extraction: Extracting “shareholder - ratio - event” triplets from announcements, annual reports, legal documents, and news. Automating this previously manual pipeline provides the greatest value.
2. Entity Alignment Assistance: Models generate candidate matches and provide matching rationales, which are then confirmed by rules and human review, drastically reducing alignment costs.
3. Report Generation: Assembling penetration results into clearly structured due diligence summaries. Note that the generated content should only “reorganize verified facts” without introducing new conclusions.
4. Numerical Hallucination Governance: This is the baseline. Any ratio or amount output by the model must be written back to a structured database and verified against the source text before being committed. Extraction results carry confidence scores, and low-confidence results are routed to manual review. The reason is straightforward: in penetration scenarios, a one-percentage-point error could alter the determination of the beneficial owner.

Regarding model selection and evaluation, several practical insights are worth noting:

• Layered Usage: Delegate complex extraction and reasoning to flagship models, and lightweight tasks like batch cleaning, classification, and formatting to smaller models, routing based on cost and latency. A model tier covering high, medium, and low loads is far more economical than “using the most powerful model for everything.”
• Take Benchmark Scores with a Grain of Salt: Public benchmarks often systematically diverge from real-world business performance. The correct approach is to build a proprietary evaluation set (labeled samples and regression test cases) and monitor business metrics such as extraction precision, recall, and numerical consistency, rather than vague benchmark rankings.
• Version and Documentation Consistency: Before production integration, verify that version release cadence, API documentation, and actual behavior are synchronized. Be cautious of “released but undocumented” versions, or models with long periods of no minor updates followed by sudden major version jumps—evaluate the risk of behavioral drift. Production integration must lock the model version and retain a rollback path.

VI. Phased Implementation Roadmap

• Phase 1 (Data Foundation): Complete the collection and cleaning of public data sources, store raw artifacts in object storage, and finalize entity master data and relationship repositories. This phase determines the system’s ceiling—it is better to proceed slowly and ensure alignment and confidence are solidly established.
• Phase 2 (Graph and Algorithms): Complete graph modeling, reachability penetration, effective shareholding calculation, circular shareholding detection, and incremental updates to form a verifiable algorithmic core.
• Phase 3 (Productization and Intelligence): Implement penetration visualization, evidence chain panels, report generation, and integrate the LLM extraction pipeline with a human-in-the-loop review process.

At the infrastructure level, general-purpose cloud services like object storage, message queues, graph databases, and logging/monitoring/alerting can cover most requirements. Leveraging such general services provided by platforms like Baidu AI Cloud can significantly compress self-build and operational costs, allowing teams to focus their energy on data governance and algorithms—the areas that truly create differentiation.

VII. Conclusion

On the surface, equity penetration is about “finding relationships.” At its core, it is a comprehensive test of data engineering, graph algorithms, and system engineering capabilities: graph quality determines the ceiling, algorithmic rigor determines correctness, engineering capability determines usability, and LLMs determine the efficiency of information acquisition. However, key numerical values must always be verified against structured data and source evidence. By sequentially refining these phases, equity penetration can evolve from a one-time manual check into an auditable, scalable, and continuously iterable core risk control asset.