Equity Penetration Analysis: Methods and Practices for Identifying the Actual Controller

    Equity penetration analysis answers what a first-level cap table cannot: who ultimately controls a company, how much economic interest each natural person actually holds, and whether nominally independent shareholders are secretly related. This article lays out a reusable, end-to-end practice framework: how to fix analytical conventions before computing anything, how to model equity relationships as a weighted directed graph, how to compute economic interest by multiplying ownership ratios along each path, and how to determine control by evaluating each layer with "or" logic. It also covers exception handling for limited partnerships, asset management products, circular holdings and potential dilution; how to use LLMs for document extraction rather than for arithmetic or legal judgment; how to build a domain-specific evaluation set; and how to make results traceable with automated validation rules. A pitfalls checklist at the end summarizes the mistakes that most often produce wrong conclusions.

    1. Introduction: What Comes After “Owns 40%”?

    In a due diligence exercise, the target company (hereafter “the Target”) has three direct shareholders: an investment company, a limited partnership, and a holding platform co-founded by two entrepreneurs. Looking only at the first layer, you can hardly answer the core questions: Who actually controls the Target? How large is a given individual’s real economic interest? And do hidden relationships exist behind these three shareholders?

    Equity penetration is the set of methods that answers these questions. It follows the ownership chain downward layer by layer until it reaches the ultimate natural persons, state-owned entities, listed companies, or other clearly attributable interest holders, thereby reconstructing each ultimate party’s economic interest ratio and control relationship. Penetration results typically support:

    • Determination of control in investment and M&A;
    • Screening for related-party transactions, horizontal competition, and undisclosed related parties;
    • Verification of the actual controller and acting-in-concert relationships in listed-company disclosure;
    • Look-through identification of financial-institution customers and their beneficial owners;
    • Locating asset leads in judicial enforcement and debt recovery.

    Note that penetration analysis is a job of calculation plus evidence gathering, not “lookup plus listing”. Given the same equity data, different conventions and different rules can produce opposite conclusions. This article walks through convention definition, data modeling, algorithm implementation, complex-structure handling, result validation, and production deployment to provide a reusable practice framework.

    2. Fix the Conventions First: Three Questions That Determine the Outcome

    Three questions must be answered before you write any code.

    Question 1: How far do you penetrate? Typical stopping points include:

    • Natural persons;
    • State-owned investing entities;
    • Listed and public companies;
    • Offshore entities (when disclosure is limited, penetrate to that layer and note the limitation).

    Question 2: Are you computing economic interest or control? The two use different algorithms and must not be conflated. Economic interest (who gets how much) is the sum of the products of ownership ratios across layers. Control (who calls the shots) requires judging at each layer whether control is established, then applying “or” logic. Limited partnerships, dual-class structures, and voting-rights proxies often cause a pronounced separation between economic interest and control.

    Question 3: What is the base date? Equity changes over time. Multiple capital increases, share transfers, and buybacks in a company’s history mean that penetration must be bound to a base date, with historical versions preserved, so that the control structure “as of then” can be reconstructed.

    Once the conventions are set, they should be frozen as written rules covering the scope of stopping entities, control determination thresholds, and the agreed treatment of special entity types, and output alongside every calculation result. This is the prerequisite for penetration conclusions that can be recomputed and audited.

    3. Modeling: Representing Equity Relationships as a Weighted Directed Graph

    3.1 Graph model design

    The most natural abstraction for penetration analysis is a weighted directed graph. Nodes are legal entities (natural persons, companies, partnerships, asset management products, etc.). Edges are ownership relationships, directed from “shareholder → investee”, with the ownership ratio as weight.

    Data element Key fields Notes
    Entity node Entity ID, name, business registration number / ID document number, type, place of registration, incorporation date Entity ID is globally unique and is the basis for later deduplication and linkage
    Ownership edge Source, target, ownership ratio, capital contribution, subscribed/paid-in, holding nature, start/end date, data source Ratios are uniformly stored as decimals, e.g. 40% as 0.4
    Control edge Source, target, control mechanism, supporting document Mechanisms include holding control, GP control, acting in concert, voting-rights proxy, etc.
    Evidence Source type, document name, disclosure date, source excerpt Provides a traceable evidence record for every edge

    3.2 Entity normalization is the key to data quality

    The most common problem in real data is not “missing data” but “the same entity treated as several entities”. Company names mix full-width and half-width brackets, differ in spacing, and appear as abbreviations, former names, or are confused with branches.

    An effective practice is to use the business registration number (for example, a unified social credit code) or a natural person’s government ID number as the unique identifier. When it is missing, first normalize the name for candidate matching, then manually confirm entities that appear to be the same. Once confirmed, merge the nodes in the graph and register all aliases under that node.

    1. def normalize_name(name: str) -> str:
    2. # Full-width to half-width, uppercase, strip whitespace and punctuation: used for candidate matching
    3. import unicodedata
    4. s = unicodedata.normalize("NFKC", name).upper()
    5. return "".join(ch for ch in s if ch.isalnum())
    6. def resolve_entity(raw_name, registry):
    7. code = registry.lookup_code(raw_name) # Exact match on ID number / registration code
    8. if code:
    9. return code
    10. key = normalize_name(raw_name)
    11. candidates = registry.search_alias(key) # Alias / former-name index
    12. if len(candidates) == 1:
    13. return candidates[0]
    14. return registry.queue_manual_review(raw_name, candidates) # Ambiguous -> human review

    4. Algorithm 1: Multiplicative Penetration of Economic Interest

    The core rule for economic interest is: an indirect holding ratio equals the product of the ownership ratios along the path, and when multiple paths point to the same entity, the interests can be added.

    The example below shows a sample structure; ratios are in parentheses. A and B are limited liability companies, M is a management company, P is a limited partnership, and T is the Target.

    1. Zhang San (60%) ──▶ A ──────(40%)─────┐
    2. Li Si (40%) ──────▶ A │
    3. Li Si (100%) ─────▶ B ──────(30%)─────┼──▶ T (target company)
    4. Zhang San (100%) ─▶ M ──────(1%)──┐ │
    5. Zhang San (20%) ──────────────────┴─▶ P ──(25%)──┘
    6. Other LPs (79%) ──────────────────────▶ P

    List each path from Zhang San to T:

    • Zhang San → A → T: 0.60 × 0.40 = 0.24
    • Zhang San → P → T: 0.20 × 0.25 = 0.05
    • Zhang San → M → P → T: 1.00 × 0.01 × 0.25 = 0.0025

    Zhang San’s total economic interest is 0.2925, or 29.25%. By the same method, Li Si holds 16% + 30% = 46%, the other limited partners hold 19.75%, and other minority shareholders hold 5%.

    Zhang San holds about 29.25% and Li Si 46%. But control does not fall where those ratios suggest — which is exactly the subject of the next section.

    Edge cases need dedicated handling

    Circular holdings. If A holds 20% of B and B holds 10% of A, naive recursion loops forever, so cycles must be detected explicitly. Path traversal should record the nodes on the current path and prune as soon as a visited node is encountered, while flagging the cycle separately for manual review.

    Performance on large graphs. When node and edge counts are large, strategies such as memoization, on-demand computation per target shareholder, and a maximum penetration depth help. Batch workloads can spin up elastic compute instances on demand through container services on Baidu AI Cloud and release them automatically after the run, avoiding idle resources.

    Illustrative code:

    1. def calc_equity(graph, shareholder_id, target_id, max_depth=10):
    2. """Compute the penetrated economic interest of a shareholder in a target entity.
    3. Edge weights in graph are fractional ownership ratios."""
    4. memo = {}
    5. def dfs(node, depth):
    6. if node == target_id:
    7. return 1.0
    8. if depth >= max_depth or node in memo:
    9. return memo.get(node, 0.0)
    10. total = 0.0
    11. for edge in graph.out_edges(node): # Edge carries weight, ratio and effective period
    12. if not edge.is_effective: # Only holdings effective as of the base date
    13. continue
    14. if edge.to in graph.path_nodes: # Guard against circular holdings
    15. continue
    16. total += edge.weight * dfs(edge.to, depth + 1)
    17. memo[node] = total
    18. return total
    19. graph.path_nodes = set()
    20. return dfs(shareholder_id, 0)

    Cross-period versions. Every calculation should take a base date and filter out ownership edges that were not yet effective or had already terminated on that date. An immutable versioning model is recommended: each equity change writes a new version record and keeps the old one instead of overwriting it, so results for any historical point in time can be recomputed.

    5. Algorithm 2: Layer-by-Layer Determination of Control Paths

    5.1 From “multiply” to “or”

    Economic interest multiplies along a path; control is combined with “or”. At each layer, first judge whether control is established: either a direct holding above one half, or less than one half combined with dominant influence. If established, the determination continues downward; if any layer fails, that control chain ends.

    In the example, the general partner M holds only a 1% interest in P. On economic interest alone, this line is negligible. But a limited partnership is usually managed by its general partner, and if the partnership agreement also provides that M represents P externally in exercising shareholder rights, then M controls the voting rights attached to P’s 25% stake in T, and M in turn is 100% held by Zhang San.

    On a voting-rights basis, T belongs to: Zhang San 65% (40% + 25%), Li Si 30%, other minority shareholders 5%. Zhang San’s economic interest is about 29.25% but he controls 65% of the votes, so he should be identified as the actual controller.

    The lesson: looking only at capital contribution ratios and ignoring voting arrangements leads to the wrong conclusion.

    5.2 Joint control and holdings below one half

    Company A is 60% held by Zhang San and 40% by Li Si, and Zhang San controls A alone. If instead each holds 50%, you must check the articles of association, board seat allocation, and any acting-in-concert arrangements to decide whether the company is under joint control or has no controller.

    In practice, voting rights held directly or indirectly below one half usually do not constitute sole control. In that case, consider a combination of factors: the relative advantage of controllable votes over the largest shareholder, how dispersed the other shareholders are, board composition, and actual influence in past votes. Many analytical frameworks treat “30% plus largest-shareholder status plus a dispersed remaining cap table” as one indicative condition for deemed control, but the specific thresholds and criteria should follow the applicable regulatory rules.

    5.3 Acting in concert and voting-rights proxies

    An acting-in-concert relationship means the voting rights of several parties are aggregated; a voting-rights proxy means the delegating party transfers the right to the proxy holder under an agreement. Both should be modeled explicitly in the graph as “control edges” and linked to the agreement text as evidence.

    In practice, pay particular attention to two kinds of hidden arrangements: family relationships, and related-party networks formed by “indirect holdings plus concurrent positions”. Penetrating equity while ignoring personnel relationships frequently misses the real parties acting in concert.

    1. def calc_control(graph, subject_id, target_id):
    2. """Determine the voting-rights ratio of a target entity that a given subject can control"""
    3. total = 0.0
    4. def dfs(node, ratio, depth):
    5. nonlocal total
    6. if node == target_id:
    7. total += ratio
    8. return
    9. if depth >= 10:
    10. return
    11. for edge in graph.control_edges(node): # Control edges: holding / GP control / concert party / proxy
    12. if not edge.is_effective:
    13. continue
    14. # Holding edges propagate by ownership ratio; concert-party edges propagate the combined ratio;
    15. # voting-proxy edges propagate per the agreement
    16. share = edge.combined_ratio if edge.kind == "CONCERT" else edge.weight
    17. if share is None: # Control relationships requiring case-by-case judgment
    18. continue
    19. dfs(edge.to, min(1.0, ratio + share) if edge.kind == "CONCERT" else ratio * share, depth + 1)
    20. dfs(subject_id, 1.0, 0)
    21. return total

    6. Complex Structures: Handling Frequent Exceptions

    1. Limited partnerships: distinguish “property interest (economic interest)” from “managing partner (control)”. A general partner may hold a minimal interest yet control all voting rights, so you must penetrate further into the GP’s own ownership structure.
    2. Asset management products and contractual funds: generally identify the ultimate principal or investor. For products that cannot be penetrated, mark the disclosure layer according to the relevant rules.
    3. Multi-layer nesting and circular capital contributions: use cycle detection to prompt manual judgment; in consolidated-reporting scenarios, cross-holdings usually need to be treated as treasury shares or eliminated.
    4. Hidden acting-in-concert arrangements: combining four types of leads — equity, positions, family ties, and financing (such as equity pledges and loan guarantees) — materially improves coverage.
    5. Potential equity changes: once grants are made, options are exercised, or instruments convert, equity incentives, convertible bonds, and warrants cause dilution. Penetration analysis should distinguish “current interest” from “fully diluted interest” and define a corresponding calculation convention.

    7. Let LLMs Do the Extraction, Not the Conclusions

    The most labor-intensive part of equity penetration is not calculation but information gathering. Business registry records, announcements, audit reports, and partnership agreements are mostly unstructured text, and letting an LLM handle extraction and organization can noticeably improve efficiency. This is a technical option rather than a prescription, and public technical evaluations show that different models have different strengths, so the choice depends on the specific task.

    7.1 A sensible division of labor

    LLMs are good at semantic understanding and cross-format summarization, and can take on:

    • Locating shareholder registers, capital contribution details, director/supervisor/senior management lists, and management clauses in partnership agreements within long documents;
    • Identifying the timeline of equity changes and holding status before and after each transaction;
    • Generating entity-matching candidates for name variants at the semantic level;
    • Producing readable analytical notes for already-computed control relationships.

    What LLMs should not do is arithmetic and legal judgment. Decimal multiplication and summation across multiple nesting layers is exactly where LLM hallucinations peak. The correct split is: the model extracts, the program calculates, and the human makes the qualitative call.

    7.2 Constrain the output, not the wording

    Instructions to the model should stress three points: every conclusion must cite the source text; fields that cannot be confirmed are filled as “unknown” and guessing is not allowed; output follows a fixed structured format (such as a JSON Schema) so that programs can validate it.

    For an extraction result like “Company A holds 40% of Company B”, the program side should immediately cross-check it against public registry information and flag a discrepancy as an anomaly.

    7.3 Build your own evaluation set

    Judging whether an LLM suits equity penetration work should not rely on generic leaderboard scores; build a dataset that matches the business:

    • Sample coverage: direct holdings, multi-layer nesting, circular holdings, limited partnerships, acting in concert, asset management products, equity-change timelines;
    • Annotation granularity: node level (is the entity correct), edge level (are the ratio and dates correct), conclusion level (is the controller determination correct);
    • Metric design: for extraction, field-level precision and recall; for conclusions, the agreement rate with actual controller determinations; hallucination rate tracked separately as the share of “output with no support in the source text”;
    • Stability testing: whether results are consistent across repeated runs of the same document and across different prompts;
    • Version regression: rerun the fixed set every time you switch models or upgrade versions, rather than switching “by feel”.

    Sample evaluation entry:

    1. [
    2. {
    3. "case_id": "LP-001",
    4. "doc_type": "partnership agreement",
    5. "structure": ["GP holds 1% interest, LP A holds 79% interest, LP B holds 20% interest"],
    6. "expected": {
    7. "equity": {"LP A": "0.1975", "LP B": "0.05", "GP": "0.0025"},
    8. "control": "GP is the managing partner and controls the 25% of voting rights in the target company held by the partnership",
    9. "controller": "actual controller of the GP"
    10. }
    11. }
    12. ]

    8. Production Deployment: Traceability Matters More Than Speed

    8.1 The full lifecycle of a penetration result

    1. Data collection: public channels such as business registry filings, listed-company announcements, regulatory disclosures, and bond prospectuses;
    2. Text parsing: structured extraction, layout reconstruction, long-document chunking;
    3. Entity normalization: entity deduplication and alias merging;
    4. Graph storage: nodes and edges persisted, with versions managed by base date;
    5. Penetration computation: economic interest penetration and control path determination;
    6. Result output: controller determination, control chains, evidence chains, analytical notes;
    7. Manual review and archiving.

    8.2 Automated validation rules

    After computation, the program should automatically run the following checks:

    • Interest conservation: with no circular holdings, the sum of path products from any node does not exceed 1;
    • Share conservation: each entity’s interests sum to 1 (rounding tolerance allowed);
    • Ratio range: 0 < ownership ratio < 1;
    • Time logic: an ownership start date is not earlier than the entity’s incorporation date;
    • Structural anomalies: cycles, self-holdings, and isolated nodes listed separately.

    If any rule fails, the item should be treated as “to be verified” rather than accepted as-is.

    8.3 Infrastructure choices

    • A large equity graph is best stored in a graph database product that supports multi-hop queries and path retrieval;
    • Batch workloads can schedule elastic compute on demand through container services on Baidu AI Cloud;
    • Object storage keeps raw documents and parsing results to meet audit trail requirements;
    • A message queue decouples collection from computation, making retries and load smoothing easier;
    • Logging, monitoring, and alerting services track job health and failure retries.

    8.4 Designing the outputs

    • Equity structure chart;
    • Controller determination conclusion and control chain paths;
    • Layer-by-layer interest breakdown table;
    • Evidence chain (the source and source excerpt of every edge);
    • Analytical notes.

    Conclusion, path, and evidence are all indispensable — only together do penetration results become reviewable.

    9. Common Pitfalls Checklist

    • Counting only direct holdings and ignoring indirect ones;
    • Adding indirect ratios directly instead of multiplying along each path;
    • Substituting economic interest ratios for control ratios;
    • Overlooking the difference between GP and LP in a limited partnership;
    • Missing parties acting in concert and voting-rights proxies;
    • Skipping entity normalization, so one entity is split into several;
    • Ignoring the base date and answering a historical question with the current structure;
    • Double-counting the same underlying interest across multiple paths.

    10. Conclusion

    Equity penetration is a rule-driven and data-driven exercise. Conventions come before algorithms, data quality sets the ceiling, and traceability determines whether a conclusion can be trusted. Introducing LLMs into the process can significantly improve extraction efficiency, but the final controller determination should still be computed by the program and characterized by the human. Only by fixing this method into a pipeline of “data collection — entity normalization — graph computation — conclusion output — manual review” can you reliably produce defensible penetration results in regulatory reviews, investment and financing due diligence, and risk screening.