Business Buyer School Research · Methods · Version 2026.1

Methodology: The Industry Resilience Study

How 152 industries were chosen, measured and scored, and what the study cannot show.

Business Buyer School calculation. Drawdowns, recoveries, benchmarks, scores and ranks are calculated by Business Buyer School from Census Bureau BDS and BLS QCEW data. Official statistics are labelled as such.

Version 2026.1 published

October 5, 2026

Analysis prepared

October 4, 2026

Great Recession data

Census Bureau Business Dynamics Statistics, 2023 release (March 2007 to March 2023)

COVID data

BLS Quarterly Census of Employment and Wages, monthly employment, 2018 to 2025 (2025 provisional)

Short version · Full study · Data · How to cite

Version 2026.1. Run date 2026-10-04. Calculation versions: universe-v1, recession-metrics-v1, recession-score-v1, redteam-v1. Every number in the study is reproduced from the raw public files by the study's scripts; the published datasets carry the version, calculation version and run date on every row.

1. Question

How did employment and establishment counts in acquisition-relevant U.S. industries behave through the Great Recession (baseline 2007) and the COVID shock (baseline 2019), and how quickly did they recover? This is a measure of industry resilience. It is not a measure of business failure, profitability, owner earnings, valuation, or the likelihood that any single business survives.

2. Sources (primary only)

SBA, IRS, ABS, SUSB and CBP data are not used in this study and are not acquired yet.

3. Industry classification and NAICS treatment

Level: 4-digit NAICS. The BDS series is NAICS 2017 vintage-consistent and is the master classification. BLS QCEW uses the NAICS vintage of the data year (2017 for 2017 to 2021, 2022 from 2022). QCEW 2018 to 2021 therefore matches BDS codes directly.

Crosswalk architecture (fields: from_naics, from_naics_version, to_naics, to_naics_version, relationship, allocation_method, confidence, notes): 4-digit relationships derived from the Census 6-digit concordances. 2012 to 2017: 308 one-to-one, 3 many-to-one, 1 one-to-many, 1 many-to-many. 2017 to 2022: 280 one-to-one, 25 many-to-one, 4 one-to-many, 46 many-to-many. Relationships are judged on code structure only. No employment weights are used, so no allocation is ever performed: non one-to-one cases are labelled no defensible mapping for splicing.

Rule applied: a QCEW COVID series is joined across the 2021/2022 change only for 1:1 codes. For every other code the QCEW series ends at 2021Q4 and, if employment has not recovered by then, recovery is read from BDS March 2022 / March 2023 (labelled BDS_March2022_fallback or BDS_March2023_fallback) or reported as not recovered. Result: 125 of 152 ranked industries use a spliced series; 3 use a BDS fallback for recovery. The three (4413, 4533 and 5151) all show 24 months to recover: their monthly series ends at 2021Q4 before employment recovered, so recovery is read from the March 2022 annual snapshot. Read 24 as recovered by March 2022, not as an exact month.

4. Industry universe (selection is structural and outcome blind)

Starting set: all 288 BDS 4-digit codes. Rules, applied in order, none reads a drawdown, recovery or rank value:

• U1 Not acquirable as an operating small business: agriculture, forestry, fishing and hunting (BDS and QCEW omit most farm and fishing operators); mining; utilities; management of companies; regulated finance (central bank, depository and nondepository credit, securities, insurance carriers, funds) and intangible-asset lessors. Kept from finance: insurance agencies (5242), investment advice and related (5239), credit-related brokers (5223).
• U2 Nonprofit-, government-, religious- or public-service-dominated, or licensed-gaming and franchise-asset industries: hospitals, K-12 and colleges, religious/civic/advocacy/grantmaking organizations, social assistance except child day care, nursing and residential care for intellectual disability and continuing care, museums, performing arts, sports franchises, casinos, urban transit and interurban bus.
• U3 Infrastructure, common carrier, telecom, port, rail, pipeline and network industries.
• U4 Size screen: mean employees per firm above 75 in 2007 or 2019 (dominated by large multi-establishment operators).
• U5 Owner-practitioner industries: mean 3.5 or fewer employees per firm in both years (no transferable payroll business).
• U6 Data-sufficiency floor: fewer than 500 establishments or 5,000 employees in any year 2007 to 2023.

Result: 167 included, 121 excluded (U1 25, U2 23, U3 15, U4 42, U5 1, U6 15). Every decision and reason is in the industry universe file. Judgment calls a reader may dispute: keeping 5239 and 5223 (investment advice and credit brokers are bought by individuals), keeping hotels (7211), restaurants (7225) and radio/TV broadcasting (5151), and excluding school-bus (4854) and nursing facilities (6231). Sensitivity to these is limited because each is one row of 152.

Granularity limit. BDS and QCEW national tables stop at 4-digit NAICS for these years. HVAC, plumbing and electrical are one industry (2382). Landscaping, janitorial and pest control are one industry (5617). Veterinary services sit inside 5419. Self-storage sits inside 5311. Auto collision repair sits inside 8111. The report profiles the 4-digit group and says so.

5. Confidence tiers and minimum-data rules

Computed from BDS 2007 to 2023 series minimums (noise-infused percentages are unstable in small cells):

• High: minimum annual establishments 3,000 or more and minimum employment 30,000 or more, no downgrade flag.
• Medium: minimum establishments 1,000 or more and minimum employment 10,000 or more, or a High industry capped for a data flag (BDS 2023 establishment re-match affecting 4451 and 4541; no QCEW COVID series).
• Low: passes the 500 / 5,000 floor only. Low-confidence industries are scored on the same scale but receive no rank and are reported separately.
• Excluded: fails U1 to U6.

Result: High 118, Medium 34, Low 15 (unranked). Flags that do not change the tier but are shown: large_firm_skew_40_75_emp_per_firm (chain-dominated industries; an industry mean is a crude proxy), 2017to2022_<relationship> (QCEW splice not possible). Suppression: BDS 4-digit cells carry a 'D' flag in the entry, exit and firm-death columns for six industries (5152, 5211, 5232, 4861, 4869, 6223; 32 to 70 cells each), all of them excluded by rules U1 to U3, so no included industry has a suppressed cell in any column; the columns Report 1 uses (employment, establishments, firms) have none for any industry. QCEW national 4-digit private rows carry no disclosure suppression. The validator asserts both.

6. Formulas

All metrics are calculated per 4-digit industry.

Great Recession (Census BDS). Baseline = March 2007 value. Trough = minimum March value over 2008 to 2012. Employment drawdown = min(0, trough / baseline − 1). Establishment drawdown likewise. Recovery year = first year after the trough year with the series at or above baseline; recovery time = recovery year − 2007. A drawdown of 0 means the industry never fell below baseline in the window and its recovery time is 0 (recorded as no decline observed). If no year through 2023 reaches baseline, the industry is not recovered during observation period.

COVID employment (BLS QCEW). Baseline = mean of the 12 monthly 2019 employment values. Trough = minimum monthly value from March 2020 to December 2021. Drawdown = min(0, trough / baseline − 1). Recovery = first month after the trough with employment at or above baseline; recovery time = months from March 2020 (0 if there was no decline). Not recovered if no month through the end of the usable series.

COVID establishments (BLS QCEW). Baseline = mean of the four 2019 quarterly counts. Trough = minimum quarter from 2020Q1 to 2021Q4. Recovery months = 3 × quarters from 2020Q1 to the first quarter at or above baseline.

Long-run volatility. Standard deviation of annual percentage employment change (BDS, March to March) over 1991 to 2019, excluding changes into 2008 to 2012, so the Great Recession is not counted twice.

Context columns (not scored). Employment CAGR 2002 to 2007 and 2014 to 2019 (pre-shock trend), employment change 2007 to 2023, establishment change 2007 to 2023.

Benchmarks (calculated the same way). (a) Total U.S. private sector: BDS national table for the Great Recession; QCEW total private for COVID. (b) The included universe: sum of all 167 included industries. (c) The industry's whole NAICS sector: BDS sector table; QCEW sector series (ending 2021Q4). Total private: Great Recession employment -7.1% decline, COVID -16.3% decline with 20 months to recover. Included universe: -7.9% and -19.9%.

7. Normalization and the composite score

Each of seven components is converted to a mid-rank percentile (0 to 100, 100 is best) within the 152 ranked industries. Percentiles are used because drawdown distributions are bounded at 0 with many ties and a few extreme values; a min-max scale would let one outlier compress everyone else. Ties share the mid-rank. Not recovered is assigned the worst value (tied). Final ranks follow the Model A score: industries with equal scores share a rank and the next rank is skipped. Two industries, 4441 and 7211, tie at rank 66, so no industry holds rank 67; 152 ranked industries use 151 distinct rank values and the last rank is 152. Lower-is-better metrics (recovery time, volatility) are inverted.

Model A (balanced, the pre-specified weights): Great Recession employment drawdown 25, Great Recession establishment drawdown 15, Great Recession recovery speed 15, COVID employment drawdown 20, COVID establishment drawdown 10, COVID recovery speed 10, long-run employment volatility 5. Recovery speed uses employment. Composite = weighted mean of the seven percentiles, shown to one decimal.

Low-confidence industries are scored against the 152-industry distribution but not ranked.

8. Sensitivity analysis (weights were tested, not assumed)

Results (computed in the study's sensitivity step): minimum pairwise Spearman rank correlation across all six models = 0.935 (lowest pair: drawdown-heavy against recovery-heavy). Top-20 overlap with Model A: B 19, C 17, D 17, E 18, F 17. A 3,000-draw Dirichlet perturbation of Model A weights gave each industry a probability of landing in the top quartile.

Stability classes: 29 stable top quartile (rank within the top quarter in all six models), 25 stable bottom quartile, 10 method-sensitive (rank range above a quarter of the list), 88 moderate. Because rank correlations stay high, the composite is reported prominently; because 10 industries and a middle band move with the weights, the report also shows component metrics next to every rank, flags stability class, and builds findings from stable results.

9. Interpretation rules

Official statistic: "U.S. Census Bureau reports...". Business Buyer School calculation: drawdowns, recoveries, benchmarks. Business Buyer School score: the 0 to 100 composite and rank. Business Buyer School estimate: none are used in this study. Establishment exit is not business failure and no failure claim is made. No SBA, IRS, profit, SDE, EBITDA or valuation figure is used.

10. Red-team review (questions, answers, and what was done)

1. Is the industry universe biased? The rules are structural and written before any outcome was read, but they remove many consumer-facing hospitality, health facility and manufacturing industries by size or sector, and keep 4-digit groups that mix trades. Disclosed in Section 4. The universe is "industries where small-business acquisition is plausible," not "every small business."
2. Are the recession periods fair? Baseline 2007 (March) is a pre-shock peak for most industries but construction employment peaked earlier. Variants tested: baseline = peak of 2005 to 2007 (composite Spearman with Model A 0.993, top-25 overlap 24 of 25); baseline 2006 (0.981, 22 of 25); trough window 2008 to 2011 (0.996, 25 of 25).
3. Does annual data conceal COVID? Yes, which is why COVID uses monthly QCEW. BDS March 2020 is pre-shock and is not used for COVID drawdown. Baseline variants (Feb 2020, 2019Q4) give composite Spearman 0.996 and 0.999 with the used baseline.
4. Are employment and establishments enough to call something resilient? No. They show how many jobs and employer locations existed, not margins, owner earnings or whether a given business survived. The report says this in its buyer section.
5. Could an industry rank highly because it was already declining? The data point the other way: growing industries tended to score higher (Spearman 0.40 with 2002 to 2007 employment growth, 0.39 with 2014 to 2019). No industry that lost more than 25% of employment from 2007 to 2023 appears in the top 25 (0 do). A trend-adjusted rescore (drawdowns residualised on pre-shock growth) correlates 0.968 with Model A and keeps 20 of 25; the industries that fall most are 4885, 4541, 6116, 4884. Not corrected, disclosed: these are correlations, and part of the score may reflect underlying growth rather than shock absorption.
6. Are NAICS changes distorting results? The BDS series is on one vintage. QCEW is spliced only across one-to-one codes (Section 3). BDS 2023 re-matched establishments for 4451 and 4541; both are capped at Medium.
7. Do very large industries overpower the analysis? The composite is unweighted by size (each industry counts once). Employment-weighted mean Great Recession drawdown is -10.2% against an unweighted median of -11.2%; weighted COVID drawdown -20.6% against -11.4%. Top-25 median 2007 employment 414,880 against 278,958 for all, so the top group is somewhat larger than average, which partly reflects the minimum-size confidence rule.
8. Do tiny industries create unstable percentages? Industries below the 500 / 5,000 floor are excluded; Low-confidence industries are unranked; 8 of the top 25 are Medium confidence (versus 34 of 152 overall) and are marked.
9. Does recovery measurement punish structurally changing industries? Yes: 53 industries (35%) have not regained March 2007 employment. Recovery is only 15% of the score and the stability tests include a recovery-light model (B).
10. Would another reasonable methodology change the top rankings? Slightly: leave-one-component-out top-25 overlap ranges from 18 to 25 of 25; the largest change comes from dropping COVID employment drawdown (18 of 25).
11. Are buyer conclusions stronger than the evidence? The buyer section is limited to what the data show and states what they do not.

Source cross-checks: BDS and QCEW agree on Great Recession employment drawdown across 140 comparable codes with Spearman 0.923 (mean absolute difference 3.0 percentage points; the series differ in timing, March snapshot against annual average, and in NAICS vintage before 2017). BDS and QCEW agree on the March 2019 to March 2021 employment change across 152 industries with Spearman 0.891.

11. Limitations (cannot be fixed with these sources)

• Two shocks of very different kinds; the Great Recession and COVID rankings correlate weakly (Spearman 0.14).
• Employer establishments only. Sole proprietors without payroll are outside BDS and QCEW.
• National data. A local market can behave differently.
• 4-digit groups hide trades inside them (Section 4).
• Industry-wide results are dominated by the industry's size mix; QCEW national files here carry no firm-size breakdown.
• Employment is not revenue or profit. Establishment counts barely fell in COVID (60% of ranked industries had no decline), so they add little signal there.
• BDS is noise-infused; QCEW 2025 is provisional; QCEW and BDS have different timing and coverage.
• Recovery "not recovered" depends on the end of the series, and for industries without a one-to-one 2022 mapping the QCEW series ends in 2021.
• The score is relative (percentiles within this universe). A score of 70 is not a probability or a prediction.

12. Revision and version system

Version string YYYY.N (this package 2026.1). A new data release (BDS 2024, QCEW revisions) triggers a re-run and a version increment; weights, thresholds or universe rules may change only with a calculation_version increment and a changelog entry. Every derived row carries source_version, calculation_version, run_date, naics_version, confidence and flags. Raw source files are never edited and are hashed. Prior versions' files stay available in their own download folder; the published files are listed with checksums on the data page.

How to cite

Business Buyer School (2026). Methodology: The Most Recession-Resistant Small Businesses to Buy: 2026 Industry Resilience Study, version 2026.1. https://businessbuyerschool.com/research/recession-resistant-businesses/methodology/. Published October 5, 2026.

A data license has not been published; short quotations with attribution and a link are welcome.

Take this into your own search

Industry research is context, not a verdict on any one business. These free parts of the school show how to use it.

• Write your Buy Box: set the type of business, size and hard exclusions before you search.
• Cleaning, landscaping and HVAC compared: how the same principles show up in three trades.
• See the curriculum or start free with Business Buyer School.