Equity Penetration Analysis: From Graph Computation to Intelligent Document Extraction
Equity penetration analysis traces ownership relationships layer by layer up to the ultimate beneficial owners — natural persons, state-owned entities, or listed companies — and computes the effective interest along each path. This article explains the two penetration perspectives (beneficial ownership and control), the cases most prone to miscalculation such as circular shareholding and multi-path aggregation, and the data-quality rules that bound accuracy. It then walks through graph algorithms for both acyclic and cyclic structures with Python examples, the role of intelligent document extraction powered by large language models, and the dimensions worth testing when evaluating a new model version. It closes with typical application scenarios and compliance controls, showing how to build a layered, auditable system that turns scattered shareholding data into trustworthy, traceable results.
1. Introduction: Why Equity Penetration Matters
In equity investment, supply-chain onboarding, financing due diligence, and compliance review, “who are the shareholders” is rarely obvious at a glance. Typical ownership structures are often nested across multiple layers: a natural person first establishes a holding platform (commonly structured as a limited partnership in practice), which then holds regional companies, holding companies, and finally the project company, tier by tier. Offshore structures may stack several additional intermediate entities. Looking only at the direct shareholder list reveals just the “surface shareholders.” Penetration analysis (also called equity penetration or ownership-structure penetration) traces ownership relationships upward layer by layer until it reaches the ultimate parties — natural persons, state-owned entities, or listed companies — and computes the interest ratio along each path.
In practice, penetration analysis serves three main needs:
- Investment decisions: after penetration, if the majority of interests ultimately belong to a few parties, any assumption about investment diversification must be reconsidered.
- Compliance review: identifying the ultimate beneficial owner and the actual controller for related-party identification, anti-money-laundering screening, and conflict-of-interest checks.
- Group governance: understanding the true control relationships among group members to support consolidated reporting, internal guarantees, and resource allocation.
It is worth emphasizing that penetration analysis is essentially a calculation with defined criteria and premises — an approximation, not a determination of legal title. Results are only comparable once the calculation criteria and data sources are defined.
2. Core Concepts: Two Penetration Perspectives
2.1 Beneficial Ownership Penetration
Beneficial ownership penetration answers “where does the money ultimately flow.” The approach is to multiply the shareholding ratios along each path level by level, then sum all paths that lead to the same ultimate party. Let p(A,B) be the ratio at which entity A directly holds B. Then the penetrated ratio P(A,T) of A in target company T is:
P(A,T) = Σ p(A,v₁) × p(v₁,v₂) × … × p(vₖ,T)over paths A→v₁→…→vₖ→T
A typical example:
| Intermediate / target entity | Direct shareholder | Shareholding ratio |
|---|---|---|
| Project company P | Partnership L | 60% |
| Project company P | Founder A (direct) | 10% |
| Partnership L | Founder A | 40% |
The penetrated ratio of Founder A in project company P is 40% × 60% + 10% = 34%.
2.2 Control Penetration
Control penetration answers “who calls the shots.” In the simple case where voting rights align with capital-contribution ratios, control is computed the same way as beneficial ownership. In practice, however, the two often diverge. Common factors include:
- Dual-class shares: some offshore structures allow super-voting shares, so a founder may control a majority of votes with a relatively small equity stake.
- Acting-in-concert arrangements and voting proxies: several shareholders exercise voting rights collectively through agreements or delegation.
- Partnership management rules: a limited partnership is typically run by the general partner, whose subscribed capital may be minimal yet who effectively controls voting and operations.
As a result, the same target can yield two very different sets of conclusions. For example, a founder may indirectly control a majority of the voting rights in a company through a holding platform, while the penetrated beneficial ownership is less than 20%. A report should present economic interest and control relationships separately, rather than merging them into a single “shareholding ratio.”
3. Technical Challenges: Three Cases Prone to Miscalculation
Circular shareholding: A holds B, B holds C, and C holds A again. Recursion expands the cycle infinitely, so naive multiplication either enters an infinite loop or produces meaningless results. Here the beneficial-ownership ratios satisfy a linear system x = Mx + d, where M[i][j] is the ratio at which j holds i, and d[i] is the ratio at which an entity outside the group directly holds i. This requires numerical solving.
Multi-path aggregation: the same ultimate party often holds shares through multiple paths, and all of them must be summed. For example, A first sets up partnership L1, then co-founds L2 with others; the two partnerships hold 30% and 20% of company P respectively, and A’s shares in L1 and L2 are 50% and 20%. Then A’s penetrated ratio in P is 50% × 30% + 20% × 20% = 19%. Taking only a single path is one of the most common sources of error.
Entity disambiguation: similar names, mixed abbreviations, renamings, and relocations can all cause “mistaken identity.” The Unified Social Credit Code is the most reliable basis for disambiguation, but natural persons have no such code and must be carefully matched against identity information, with a manual review step in place.
4. Data Layer: Sources and Quality Set the Upper Bound on Results
4.1 Data Sources
- Authoritative registration data: public channels such as business registration records offer high accuracy but can lag behind updates, making them suitable as a baseline.
- Announcements and periodic reports: public disclosures by listed entities provide relatively timely equity-change information, but need to be parsed.
- Internal business systems: shareholder information in contracts, account-opening records, and customer master data can fill gaps left by registration data.
- External supplementary materials: articles of association, partnership agreements, and voting-arrangement documents from due diligence materials.
In practice, a common combination is “authoritative data as the baseline, external materials to fill the gaps.” Each data source should carry a timestamp, and a single penetration result should be based on the same point in time, avoiding the splicing of data from different periods.
4.2 Data Quality Rules
- Ratio closure check: the direct shareholder ratios under the same held entity should sum to approximately 100%; a clear deviation indicates missing or duplicated data.
- Direction check: whether the hierarchy level at which a node appears is consistent with its shareholding ratio.
- Missing and conflicting values: treating missing values as non-participating in multiplication (equivalent to treating them as 0) systematically underestimates interests, so confidence should be annotated in the report; conflicting values are resolved by source priority with an audit trail.
- Change tracking: retain every revision record so historical results can be recomputed and traced.
5. Algorithm Layer: From Path Enumeration to Graph Computation
5.1 Acyclic Structures: Memoized Recursion
An acyclic ownership structure is essentially a directed acyclic graph. Edges represent “holder → held entity,” with the ratio as the weight. The example below shows the core idea in Python:
from collections import defaultdictclass EquityGraph:def __init__(self):self.holds = defaultdict(list) # holder -> [(held entity, ratio)]self.parents = defaultdict(set) # held entity -> {direct holders}def add_edge(self, holder, target, ratio):assert 0 <= ratio <= 1self.holds[holder].append((target, ratio))self.parents[target].add(holder)def effective_ratio(g, entity, root, memo, seen):"""Penetrated ratio of `entity` relative to `root`; memo caches subgraph results, seen prevents infinite loops caused by cycles"""if entity == root:return 1.0if entity in memo:return memo[entity]if entity in seen:raise ValueError("Cycle detected; switch to the linear-solving method in 5.2")total = 0.0for holder in g.parents.get(entity, ()):direct = sum(r for t, r in g.holds[holder] if t == entity)total += direct * effective_ratio(g, holder, root, memo, seen | {entity})memo[entity] = totalreturn total# Example: A -> Partnership L (40%) -> Project company P (60%), A -> P (10%)g = EquityGraph()g.add_edge("Founder A", "Partnership L", 0.40)g.add_edge("Partnership L", "Project company P", 0.60)g.add_edge("Founder A", "Project company P", 0.10)print(f"{effective_ratio(g, 'Project company P', 'Founder A', {}, set()):.2%}") # 34.00%
This uses a “bottom-up, parent-backtracking” style. Compared with enumerating paths downward from the founder, memoization computes each node’s ratio only once and naturally aggregates all multi-paths.
5.2 Cyclic Structures: Iterative Solving of Linear Equations
When circular shareholding exists, solve x = Mx + d with the classic iterative method. In each iteration, every node recomputes using the latest ratios of its holders. For example, suppose M (an entity outside the group) directly holds 70% of A, and the remaining ownership is held indirectly through the cycle A→B→C→A, where A holds B at 60%, B holds C at 50%, and C holds A at 30%:
import mathdef solve_loop(g, nodes, external):"""external: the direct shareholding ratios that entities outside the group hold in each node"""x = [float(v) for v in external]for _ in range(10000):diff = 0.0for i, node in enumerate(nodes):nxt = external[i] + sum(sum(r for t, r in g.holds[holder] if t == node) * x[j]for j, holder in enumerate(nodes))diff = max(diff, abs(nxt - x[i]))x[i] = nxtif diff < 1e-10:breakelse:raise RuntimeError("Iteration did not converge; check the data for abnormal cycles")return dict(zip(nodes, x))# Node order [M, A, B, C], where M is the entity outside the group, external = [1.0, 0.70, 0, 0]result = solve_loop(g, ["M", "A", "B", "C"], [1.0, 0.70, 0.0, 0.0])# Result: M=1.0, A≈76.92%, B≈46.15%, C≈23.08%
The result satisfies A’s ratio = 70% + 30% × C’s ratio. Convergence depends on the magnitude of the product of ratios around the cycle: the smaller the product, the faster the convergence; when the product approaches 1, convergence is slow, and one should switch to a direct-solving method with precision checks.
5.3 Complexity and Engineering Considerations
- Computational complexity: memoized recursion is O(V+E); iterative solving is O(E) per pass, with the total number of passes depending on the convergence condition.
- Incremental computation: corporate equity changes frequently; recomputing only the affected nodes and their descendants is far more efficient than a full recomputation.
- Large-scale scenarios: when the number of entities reaches the tens of millions, general-purpose infrastructure such as graph databases and distributed computing platforms can handle graph construction, path queries, and batch computation, but the criterion checks must be implemented consistently at the business layer.
6. Enhancement Layer: Introducing and Selecting Intelligent Document Extraction
6.1 Why Intelligent Extraction Is Needed
Registration data covers only part of the ownership relationships. Partnership agreements, articles of association, and voting-arrangement documents are often unstructured text. Newer large models keep iterating, with steadily improving long-context understanding, reasoning, and structured-output capabilities, which raises the feasibility of an automated pipeline: “parse the entire agreement → extract ownership relationships → write them into the graph.” But it must be recognized that no matter how strong a model is, it cannot replace manual review. The value of the model is to shift human effort from page-by-page searching to sampling checks and difficult judgments.
6.2 Technical Architecture and General Cloud Product Capabilities
A layered “extract — compare — store” architecture is recommended, which can be implemented with combinations of general-purpose product categories:
- Document parsing: run OCR on scanned documents first; for complex layouts, leverage layout analysis and preserve table structures.
- Object storage: centrally store the original files and extraction results to ensure end-to-end traceability.
- Message queue: carry extraction tasks and failure retries, smoothing out traffic spikes.
- Container platform: deploy extraction services with on-demand elastic scaling.
- Logging and monitoring/alerting: record model versions, latency, and failure reasons to provide data for selection and evaluation.
6.3 Prompts and Output Constraints
Structured extraction should force JSON output and write into the store only after passing business validation:
{"entities": [{"id": "E001", "name": "Founder A", "type": "natural person"}],"relations": [{"holder": "E001", "target": "E002", "ratio": 0.4},{"holder": "E002", "target": "E003", "ratio": 0.6}],"evidence": [{"relation_index": 0, "source_page": 3, "quote": "A subscribes 4 million yuan of capital, 40% of the partnership's total subscribed capital"}],"notes": ["The agreement's annex states that limited partners have no voting rights over partnership affairs"]}
Key points:
- Each relation must carry its source and the original text excerpt, making review and location easy.
- For fuzzy expressions such as “less than,” “no less than,” and “approximately,” the model should output a numeric range and flag it for manual confirmation.
- Having the model first restate the key clauses before extracting can reduce omissions and misalignments in long documents.
6.4 How to Evaluate the Highlights of a New Model Version
After a new model version is released, it is not advisable to draw conclusions based on demo effects alone. Small-scale benchmark tests can be run along the following dimensions, comparing across models using the same batch of real documents and labeled answers:
- Long-document understanding: whether it can fully extract key clauses from a several-dozen-page partnership agreement, rather than only capturing the opening part.
- Reasoning ability: whether it can recognize situations that require multi-step reasoning, such as “circular shareholding exists in the penetration chain” or “an acting-in-concert arrangement causes control and beneficial-ownership ratios to diverge.”
- Structured-output stability: whether the JSON is valid, whether fields are complete, and whether numeric units are consistent.
- Hallucination rate: fabricating non-existent shareholders or ratios is the most serious error; focus on the proportion of relations with no supporting source.
- Cost and latency: assess unit price and batch throughput against the workload; for scenarios requiring real-time queries, pay attention to time-to-first-token latency.
- Tool-calling ability: whether it can call tools such as calculators and database queries under controlled conditions to verify numbers, rather than “computing from memory.”
- Transferability: whether it remains stable on your own business samples, avoiding good performance on only one demo set.
At the same time, extraction quality directly affects the correctness of penetration computation:
- Misreading a ratio by one digit causes the penetration result to deviate by a multiple.
- Missing one path systematically underestimates the interest.
- A shareholder relationship “hallucinated” by the model is taken as fact by the algorithm and continues to be computed, and it is hard to detect in the final result.
Therefore, the quality bar for the extraction stage should be higher than that for generative applications.
7. Application Layer: Typical Scenarios and Output Forms
- Investment due diligence: output the complete penetration chain from founder to target company, the beneficial-ownership ratio at each layer, the control judgment, and an evidence index.
- Related-party identification: cluster member companies under the same actual controller to support related-transaction review and credit-concentration management.
- Risk alerts: monitor equity changes and trigger incremental recomputation; raise alerts when the actual controller, beneficiary, or shareholding ratio breaches a threshold.
- Reporting and visualization: use a layered force-directed graph to display the equity network, color nodes by penetration ratio, and separately mark cycles, missing nodes, and low-confidence nodes.
8. Compliance and Risk Control Points
- Data compliance: when handling natural-person shareholder information, follow the principle of data minimization; desensitize or encrypt sensitive information, and define retention periods and scope of use for the data.
- Method transparency: reports must clearly state the penetration levels, ratio criteria, data sources, and time point. Results under different criteria cannot be directly compared.
- Qualified conclusions: model extraction results should be phrased as “based on the available materials” and note any uncovered indirect holdings, nominee holdings, or contractual arrangements; they must not be treated as proof of title.
- Human in the loop: threshold judgments, exception handling, and final report sign-off must be done by humans; model output should not be published externally automatically.
9. Conclusion: Layered Decoupling as an Engineering Practice
Equity penetration analysis may seem like merely “multiplying shareholding ratios,” but in practice it involves multiple layers: data governance, graph algorithms, intelligent extraction, and compliance control. A sound approach is to split the system into five layers, connected by clear data contracts:
- Data layer: multi-source ingestion and entity disambiguation.
- Graph layer: a unified graph model carrying ownership relationships.
- Computation layer: separate implementations for beneficial-ownership penetration and control analysis.
- Enhancement layer: general-purpose large models handle low-barrier tasks such as document extraction.
- Application layer: outputs results for scenarios such as due diligence and risk control.
This way, the system can absorb the efficiency gains from improving general-purpose large-model capabilities, while controlling risk through rule-based validation, manual review, and full traceability — making penetration results truly trustworthy, usable, and auditable.
