NAICS-SIC Crosswalk for Private Equity Screening

Executive Summary

A NAICS-SIC crosswalk for private equity screening is a controlled translation between classifications, followed by a separate mandate decision. NAICS is the federal statistical standard for classifying business establishments ^[1]. An SEC SIC field instead indicates a filer's type of business ^[2]. Vendor company codes and establishment records may therefore describe different units even when each field is valid.

The workflow has three transformations. First, record the scheme, vintage, granularity, unit, source, observation date, and activity description. Second, map SIC through the official 2002 NAICS bridge ^[3], then update NAICS vintages separately. Third, apply the mandate after classifying the relationship. NAICS numbers do not inherit SIC numbering ^[4], so prefix substitution is indefensible.

Automation should be narrow. An exact mapping links one class to one class with the same activity coverage ^[5]. Splits require actual-activity review. Merges may pass only when all contributing activities receive the same mandate treatment. No-match and conflicting records remain unresolved. This matters at scale: BLS estimated 2017-to-2022 split changes affected 92,500 establishments and 1.1 million jobs ^[6]. The federal comparison identifies 217 part-of-industry entries ^[7].

Report the share of input records ending as exact, ambiguous, activity-reviewed, or unresolved. This is input-record coverage, not total-market coverage. The workbook below preserves source URL, retrieval date, relationship type, mandate action, review flag, and transformation history. Docket retains a source, supporting excerpt, and collection date for each researched answer (Source: docket.capital); that evidence process can support activity review, but it does not replace official concordances or the fund's mandate.

Introduction and Background

Industry codes look deterministic because they are short. The apparent precision is misleading. The Standard Industrial Classification (SIC) system uses four-digit industry codes, while NAICS provides greater detail through a six-digit structure ^[8]. More detail does not mean that every old class has one obvious new home.

The underlying unit matters more than the number of digits. NAICS is explicitly an establishment classification ^[9]. Private equity screens usually operate on companies or groups, which can encompass multiple activities. A single code can consequently be useful routing evidence, but it cannot establish an enterprise's full activity set or revenue mix.

The systems also serve different operational contexts. The SEC's SIC list says the codes indicate a filer's type of business ^[2]. IRS Schedule C uses six-digit principal-business codes based on NAICS ^[10]. Federal procurement assigns a NAICS code according to the principal purpose of a particular supply or service being acquired ^[11]. These fields should not be combined into one timeless company truth.

This report treats the industry-classification crosswalk as an M&A screening control. It answers a narrower and more useful question: when can a mapping safely drive an acquisition screen, when must it trigger activity review, and when should the record remain unclassified?

Classification Units, Vintages, and Source Semantics

Begin with the statistical unit

Classification follows an establishment's primary activity ^[12]. Where ideal production-cost and investment measures are unavailable, revenue, shipments, or employment can act as proxies. None of this warrants inferring a portfolio company's revenue mix from a single code.

Data teams should preserve a classification_unit field with controlled values such as establishment, enterprise, registrant, tax-return activity, contract, vendor-company, and unknown. The correct response to unknown is not to assume enterprise. It is to hold the record for source-semantic review.

Different official series demonstrate why this field is necessary:

• County Business Patterns: establishes geography from an establishment's physical location ^[13]. Its establishment values are perturbed before table creation ^[14], so published aggregates are not firm-level ground truth.

• Quarterly Census of Employment and Wages: attempts to assign separate codes where an employer conducts different activities at different establishments ^[15]. State workforce agencies verify the industry, location, and ownership classifications of virtually all establishments on a three-year cycle ^[16].

• Statistics of U.S. Businesses: can classify an enterprise into the two-digit sector where it paid the largest payroll share ^[17]. That is an enterprise summary rule, not evidence that every site has the same activity.

• EDGAR financial-statement data: defines SIC as a four-digit SEC assignment as of the filing date ^[18]. The data documentation warns that fields derived from submissions do not necessarily represent the most current assignments ^[19].

Preserve code vintage and granularity

NAICS is revised every five years, in years ending in 2 and 7 ^[20]. Federal statistical establishment data for reference years beginning on or after January 1, 2022 were directed to use 2022 NAICS ^[21]. BLS likewise states that QCEW data from 2022 forward use NAICS 2022 ^[22].

A code without a vintage is incomplete. So is a two-digit sector presented as if it were a six-digit national industry. The import contract should require:

• Scheme: SIC, NAICS, or a vendor taxonomy.
• Vintage: 1987 SIC, 2002 NAICS, 2017 NAICS, 2022 NAICS, or explicit unknown.
• Granularity: two through six digits for NAICS, four digits for SIC.
• Unit: establishment, enterprise, registrant, contract, or unknown.
• Assignment source: agency, vendor, self-report, analyst, or derived mapping.
• Observation date: when the source asserted the code.
• Activity evidence: source description, target-site wording, and relevant establishment context.

The hierarchy has five levels from the two-digit sector through the six-digit national industry ^[23]. A mandate written at four digits should not silently become a six-digit rule, and a six-digit exclusion should not automatically exclude its entire sector.

A Version-Aware NAICS-SIC Crosswalk Protocol

Step 1: normalize, but do not reinterpret

Normalize formatting before mapping: preserve leading zeros, remove display punctuation, and reject nonconforming lengths. Keep the original field unchanged beside the normalized value. Do not pad an unknown three-digit value into a plausible four-digit SIC code. Do not promote a broad NAICS code to six digits.

There is no central government agency that assigns, monitors, or approves every establishment's NAICS code ^[24]. That makes the origin of the code part of the evidence, not administrative decoration.

Step 2: map SIC to the bridge vintage

Census describes concordances as detailed descriptions of direct relationships between classification systems ^[25]. For legacy SIC data, the official files connect 1987 SIC and 2002 NAICS in both directions. Direction matters because a forward mapping and reverse mapping need not yield a one-to-one round trip.

Use the official files as relationship tables:

• SIC to NAICS: start with the Census 1987_sic_to_2002_naics.xls workbook listed in the official archive ^[3].
• NAICS to SIC: use the separately published 2002_naics_to_1987_sic.xls workbook ^[26].
• No result: store unresolved_no_official_destination; do not select the nearest wording. The reverse file includes a blank-SIC, “Null Set for U.S.” record for NAICS 112130 ^[26].

Step 3: update the NAICS vintage separately

Once SIC has been bridged to 2002 NAICS, traverse each later vintage as a distinct transformation. This prevents a 1987-to-2022 jump from hiding whether ambiguity arose in the SIC bridge or in a later NAICS revision. The full Census 2022-to-2017 workbook should drive the 2017/2022 relationship layer. BLS distinguishes the full concordance from a changes-only file ^[27].

Do not treat the manual's appendixes as a complete lookup. The manual says those tables do not cover all economic activities ^[28]. A code absent from a changes appendix may simply be unchanged.

Step 4: classify the relationship before the mandate

Table 1 converts classification relationships into automation decisions. It also places Docket honestly as an evidence and review layer, not as an official classification authority.

The key distinction is scope. Statistics Canada separates real changes that alter classification scope from virtual changes that do not ^[31]. A title change can be automated after verification. A split cannot be resolved from code structure alone.

Building the Audit Workbook

Required fields and downloadable source files

The workbook should be assembled from the official Census downloads linked above, not from a commercial converter. Because this publication delivers article and metadata files only, Table 2 is a workbook-ready schema that can be copied to CSV or a spreadsheet. Each relationship row retains a downloadable official source file in source_url.

Table 2 provides the required fields plus controls needed for reproducibility.

The table should use one row per source-destination relationship. A three-way split therefore produces three rows with one transform_id, not one cell containing three codes. This makes ambiguity countable and prevents a spreadsheet lookup from returning only the first candidate.

The provenance fields are substantive controls. NIST defines provenance as the chronology of origin, development, ownership, location, and changes ^[32]. GAO's data-reliability framework evaluates accuracy, completeness, and applicability for the purpose at hand ^[33]. For PE screening, a valid government code may be accurate for its source yet inapplicable to an enterprise-level acquisition rule.

Preserve mandate logic outside the crosswalk

The crosswalk table should never contain the only copy of the investment thesis. Maintain a separate, versioned mandate table with:

• Explicit inclusions: exact codes and descriptions that qualify.
• Explicit exclusions: exact codes and activities that disqualify.
• Family rules: broader two-, three-, or four-digit branches, with exceptions.
• Activity keywords: evidence terms that route ambiguous mappings, never terms that silently override a code.
• Review bands: conditions requiring analyst review, such as mixed activities or unknown units.
• Precedence: explicit exclusions, explicit inclusions, family rules, keywords, then unresolved.
• Effective dates: the period during which each rule applied.
• Rationale: a short statement connecting the rule to the mandate.

Keep the taxonomic relationship and mandate outcome separate. One official mapping can feed different actions for different funds. Likewise, the same fund may change its treatment without any change to NAICS.

For source-sensitive rules, retain the source's effective date. Federal contracting, for example, applies the SBA size standard in force when a solicitation is issued ^[34]. SBA size standards are usually expressed through employee or average annual receipt measures and account for affiliates ^[35]. Those rules are useful context, but an SBA size determination is not a PE revenue or earnings estimate.

Testing and refresh controls

Every transformation should pass four tests before production:

1. Schema test: scheme, vintage, granularity, and unit are present or explicitly unknown.
2. Referential test: the source code and each destination occur in the cited official file.
3. Cardinality test: the stored relationship count equals the official relationship count.
4. Round-trip diagnostic: reverse mapping is compared, but not required to reproduce the input because splits and merges are legitimate.

Refresh on source events, not an invented universal schedule. NAICS has a five-year revision cycle, but vendor assignments and company activities can change on other timelines. The official 2022 decision also eliminated the enterprise-level SIC directive immediately ^[36]. Recommended triggers are a new official vintage, a revised concordance file, a new vendor extract, material activity evidence, and a changed mandate rule.

Reconciling Vendor Codes With Target Activity

A vendor code is a lead for verification. It becomes screening evidence only after the team knows what unit it describes, which vintage it uses, and how it was assigned. The reconciliation process should compare, not overwrite, sources.

An evidence-first reconciliation sequence

1. Retain every original field. Never replace the vendor code in place.
2. Identify the claimed scheme and vintage. If the vendor cannot specify either, mark them unknown.
3. Determine the classified unit. A registrant, legal entity, location, contract, and vendor company profile are not interchangeable.
4. Run the official relationship mapping. Preserve every candidate and its relationship type.
5. Collect activity evidence. Use current product and service pages, filings, facility descriptions, and named operating segments where relevant.
6. Compare establishment evidence. Review the activities of each material location instead of forcing a multi-location target into one company-wide code.
7. Apply the versioned mandate. Use explicit inclusion, exclusion, and review logic.
8. Keep conflicts visible. Do not flatten disagreement into a single “clean” code.

Docket's Audit stage provides a compatible evidence practice: it records conflicts and their resolutions (Source: docket.capital). Its Triage stage checks disqualifying criteria first (Source: docket.capital). In a code-normalization workflow, that means inexpensive exact exclusions can run early, while ambiguous mappings remain reviewable instead of being forced into a score.

Automation bands

Use three operational bands:

• Green, automatic: exact official relationship, known source vintage, aligned unit, destination rule explicit, and no contradictory activity evidence.
• Amber, review: split or partial relationship, multiple establishments, broad source granularity, stale assignment, or disagreement among sources.
• Red, unresolved: unknown scheme or vintage, no official destination, missing activity description, or irreconcilable unit mismatch.

An amber result is not a failed record. It is a correct statement that code translation alone cannot decide the mandate. A red result should remain searchable and measurable so analysts can prioritize remediation. Industry codes also travel into machine-facing systems: Google's Organization structured-data documentation lists naics among identifiers used to disambiguate organizations ^[37]. That makes explicit version and unit metadata important when codes flow between sources.

Data Analysis and Evidence

The available numbers quantify classification complexity, not the addressable acquisition market. The five-level hierarchy described earlier culminates in national industries. The official 2022 manual counts 20 sectors and 1,012 U.S. industries ^[38]. Those structural counts explain why a family-level rule and a national-industry rule have different precision.

Revision effects are measurable. Using 2019 annual-average QCEW data, BLS estimated that direct-plus conversions affected about 610,000 establishments and 8.1 million jobs ^[39]. It estimated that split changes alone touched about 92,500 establishments and about 1.1 million jobs ^[6]. A separate BLS estimate said approximately 10% of employment would be reclassified into different industries ^[40].

The federal comparison table reports 217 entries marked as part of a 2022 U.S. industry ^[7]. OMB also reported that 79 industries, or 71% of new 2022 industries, were created by merging two or more 2017 industries or parts ^[41]. These figures show that ambiguity is not a rare spreadsheet edge case, but they do not reveal the share of any PE target universe that will be ambiguous.

The correct coverage calculation

For an imported set of N source records, calculate mutually exclusive final outcomes:

• exact_share = exact_count / N
• ambiguous_share = ambiguous_count / N
• activity_reviewed_share = activity_reviewed_count / N
• unresolved_share = unresolved_count / N

The four shares should sum to 100%. Define precedence so a reviewed split lands in activity_reviewed, not both ambiguous and activity_reviewed. Table 3 shows a fictional batch to make the arithmetic concrete.

The fictional example yields 64% exact, 19% ambiguous, 13% activity-reviewed, and 4% unresolved. It is a workflow-quality view, not total-market coverage and not a benchmark. The measurement should be repeated by vendor, source vintage, classified unit, and mandate version. A rising exact share can be good, or it can indicate that ambiguity was incorrectly flattened. Pair it with audit samples and relationship-cardinality tests.

Additional useful measures are:

• Round-trip mismatch rate: records whose reverse mapping returns more or fewer source candidates.
• Unknown-vintage rate: records that cannot enter an official version chain without an assumption.
• Unit-unknown rate: codes for which company, establishment, or other granularity is undocumented.
• Conflict rate: targets with materially different codes across sources after vintage normalization.
• Review yield: amber records resolved by activity evidence divided by amber records reviewed.
• Rule-change impact: records whose mandate action changes under a new rule version while the taxonomy stays constant.

None should be interpreted as market penetration. GAO's purpose-specific reliability framing is the right discipline: applicability depends on the intended use ^[42].

Case Studies and Real-World Examples

Manufacturer with service operations (Hypothetical Example)

A fictional industrial group arrives from Vendor A with one four-digit SIC code. The SIC-to-2002 bridge produces two NAICS candidates, one for manufacturing and another covering a service activity. Vendor B supplies a six-digit 2017 NAICS code, but the target website describes three plants and a national maintenance network.

A naïve first-match lookup selects manufacturing and passes the target. The controlled workflow stores all candidates, records that the company has multiple establishments, and routes the split for activity review. Analysts verify what each establishment does, but they do not infer consolidated revenue mix from the codes. If the mandate includes manufacturers and excludes stand-alone service companies, the result can remain review until evidence establishes which activities define the acquisition thesis.

The logic follows the establishment-versus-enterprise distinction described earlier. Review is required because the acquisition perimeter can contain several operating activities, not because uncertainty is inherently preferable.

E-commerce classification change (Hypothetical Example)

A fictional target carries a 2017 code from an older vendor export. The 2022 relationship table shows that the destination draws whole or partial activities from several prior classes. A string-based update chooses the destination whose title contains “electronic.” The official relationship, however, is a merge with partial contributions, the pattern federal tables explicitly flag when a 2022 industry came from more than one 2017 industry ^[43].

The correct workflow inspects the target's actual retail or wholesale activity, assigns the 2022 code only when evidence supports it, and preserves the original 2017 field. If every source class has the same mandate action, the merge can be automated. If one is excluded, the record remains amber.

Filing code mistaken for a current company taxonomy (Hypothetical Example)

A fictional public-company carve-out list imports an SEC SIC field and applies it to each subsidiary. That creates two errors: the filing code concerns the registrant's type of business, and the data-set value is tied to the filing date. SEC documentation states both the assignment context ^[18] and the possibility that an assignment is not current ^[19]. The screen should retain that SIC field as registrant evidence, then classify each relevant operating unit from its own activity evidence.

Implications and Future Directions

The immediate implication for origination teams is organizational. Industry normalization should be owned jointly by data operations and mandate owners. Data operations controls provenance, vintages, relationship cardinality, and reproducibility. The deal team controls which activities qualify. Combining both responsibilities in one undocumented lookup formula makes silent mandate drift likely.

The protocol also changes vendor evaluation. A useful vendor field is not merely populated. It identifies its scheme, vintage, unit, assignment method, and observation date. A lower fill rate with explicit unknowns can be more decision-useful than a fully populated column whose semantics are undocumented.

Future NAICS revisions should be treated as migrations. NAICS was designed and maintained for statistical purposes ^[44]. A new version can reflect real scope changes or virtual coding changes. Teams should load the new official relationship file, classify each changed row, replay the mandate rules, and compare outcomes before switching production.

Machine-assisted review can help rank ambiguous records, but it should not invent certainty. A model can extract activity language, compare it with official descriptions, and explain why a destination is plausible. The accepted record should still retain the source passage, model or analyst decision, timestamp, rule version, and competing destinations. Docket records changes so teams can see which criteria applied to each assessment (Source: docket.capital). That pattern is useful for code normalization because a later mandate revision should be replayable without erasing the earlier result.

Frequently Asked Questions (FAQs)

How do teams map SIC codes to NAICS codes?

Start with the official 1987 SIC-to-2002 NAICS concordance, preserve every destination row, classify the mapping relationship, and then traverse later NAICS vintages as separate steps. Do not infer the destination from similar titles or numbers. The reverse direction is a separate official workbook ^[26].

Can a NAICS-to-SIC conversion be fully automated for deal sourcing?

Only exact relationships with known vintage, granularity, unit, and consistent mandate treatment should be automatic. Splits, partial relationships, cross-source conflicts, and no-match rows need activity review or remain unresolved. The official files are directional relationship tables, not assertions about a particular target company.

Which NAICS code should represent an acquisition target?

There may be no single code that represents the entire company. The answer depends on the acquisition perimeter and on the mandate's treatment of material activities. A primary code should not be presented as revenue mix.

What are the main NAICS-SIC crosswalk limitations?

The central limitations are different statistical units, missing source vintages, one-to-many and many-to-one relationships, partial-class mappings, revisions over time, and no-match rows. Numeric resemblance should never substitute for an official relationship row.

How should a PE team preserve investment-mandate logic across code systems?

Store taxonomy relationships separately from mandate rules. Version explicit inclusions, exclusions, family rules, activity keywords, review bands, precedence, and effective dates. Then replay those rules against each translated destination. If contributing activities receive different mandate actions, hold the record for review.

How often should the mapping be refreshed?

Refresh when an official classification or concordance changes, a vendor sends a new extract, evidence shows a target's activities changed, or the mandate changes. The five-year taxonomy cycle described earlier is not a universal schedule for company-code refreshes.

Conclusion

A reliable NAICS-SIC crosswalk for private equity screening is a versioned evidence process, not a lookup widget. The process records what the incoming code means, maps it through official directional concordances, identifies relationship cardinality, checks the target's actual activities, and only then applies the investment mandate.

The safest automation boundary is narrow. Exact relationships with known unit, vintage, and consistent rules can move automatically. Splits, merges with differing mandate treatment, partial relationships, source conflicts, and absent matches require review or remain unclassified. This preserves useful uncertainty instead of converting it into false precision.

The durable artifact is an audit workbook in which every destination has an official source, retrieval date, relationship type, mandate action, review flag, and transformation history. Coverage then means the share of the imported records that are exact, ambiguous, activity-reviewed, or unresolved. It does not mean share of the market. With that distinction intact, industry codes can accelerate target screening without silently rewriting the acquisition thesis.