FP&A Work in 2026: Evidence from 101 Current Vacancies

FP&A Work in 2026: Evidence from 101 Current Vacancies

The complete archive - a searchable PDF and the accepted-vacancy dataset - is preserved at Zenodo DOI 10.5281/zenodo.22083631.

Executive summary

This report examines what employers ask FP&A professionals to do, using a point-in-time corpus of 101 publicly discoverable English-language vacancies retrieved on 24 August 2026. The accepted corpus contains 101 unique source URLs, 93 employer or advertiser labels and six public applicant-tracking-system (ATS) source families. The evidence is sufficiently diverse for a bounded analysis of recurring work signals: 86 employer or advertiser labels appear once, six appear twice and one appears three times. It is not a census of the labour market and must not be read as a market-size estimate.

The clearest result is that communication is not peripheral to FP&A work. Executive communication was coded in 95 vacancies (94.1%). It appeared together with forecasting in 73 vacancies (72.3%) and with management reporting in 72 (71.3%). The corpus therefore portrays FP&A as a decision-communication practice as well as an analytical one.

The operating core is also visible. Forecasting appeared in 78 vacancies (77.2%), management reporting in 76 (75.2%), financial modelling in 64 (63.4%), budgeting in 56 (55.4%), and KPI and performance analysis in 55 (54.5%). Forecasting and management reporting co-occurred in 68 vacancies (67.3%); forecasting and budgeting in 54 (53.5%); and financial modelling and forecasting in 57 (56.4%). These are non-exclusive observations, so their percentages are not intended to sum to 100%.

Other capabilities were present in smaller but meaningful groups: business partnering in 38 vacancies (37.6%), data quality and automation in 24 (23.8%), variance analysis in 23 (22.8%), and scenario analysis in 14 (13.9%). The co-occurrence evidence suggests that these are connected to the core cycle rather than isolated specialties. For example, 35 vacancies (34.7%) combined business partnering with executive communication, 21 (20.8%) combined data quality and automation with forecasting, 22 (21.8%) combined variance analysis with management reporting, and 13 (12.9%) combined scenario analysis with forecasting.

The practical implication is a connected view of FP&A work: translate a business question into drivers and assumptions; build or update a model; prepare a budget, forecast or scenario; compare expected and observed performance; and communicate implications to authorized decision-makers. Professional learning designed from this evidence should develop the whole planning-to-performance chain, including data controls, review discipline and responsible use of AI-assisted analysis. It should not promise forecast accuracy, financial outcomes, employment, promotion, salary, accreditation or employer recognition.

1. Research question and scope

The research question is:

Which work capabilities recur in a bounded sample of current public vacancies for FP&A and closely equivalent planning, forecasting, analysis, reporting or finance-business-partner roles?

“Current” refers only to public discoverability at the evidence-retrieval point, 24 August 2026. A vacancy may have changed or closed since that date. “FP&A” is used descriptively as the common abbreviation for financial planning and analysis; this report does not claim or reproduce any professional certification, examination syllabus, proprietary competency model or vendor method.

The study addresses the content of advertised work. It does not estimate the number of jobs available, candidate supply, salary levels, hiring probability, the time professionals spend on each activity, the proficiency expected for a capability, or the causal value of a skill. It does not provide accounting, audit, investment, tax, legal or jurisdiction-specific reporting advice.

2. Reproducible methodology

2.1 Sampling frame

The study uses the accepted vacancy corpus recorded in source S1. It is a purposive, point-in-time sample of publicly discoverable English-language vacancy pages. Twelve search-evidence batches supplied candidate records to the corpus-validation process. A row was accepted when readable public vacancy evidence exposed FP&A or closely equivalent planning, forecasting, analysis, reporting or finance-business-partner work.

Each accepted row preserves the following fields:

• employer or advertiser label;
• role title;
• jurisdiction or location as exposed by the source;
• retrieval date;
• public ATS source family;
• source URL;
• a short, necessary supporting excerpt;
• non-exclusive observed-skill codes; and
• an acceptance reason.

The row-level evidence is retained in S1. This report uses aggregate counts and does not reproduce vacancy descriptions. Source and employer names are identifiers only; all names and trademarks remain the property of their respective owners.

2.2 Inclusion, validation and deduplication

The corpus was validated against six deterministic checks recorded in S2:

1. at least 100 suitable rows;
2. all required fields present;
3. unique source URLs;
4. no employer or advertiser label contributing 10% or more;
5. multiple public ATS source families; and
6. short evidence excerpts rather than bulk job descriptions.

All six checks passed. The resulting corpus contains 101 rows and 101 unique URLs. URL uniqueness is the documented deduplication rule; no broader claim is made that differently worded or differently hosted vacancies could never refer to related hiring activity.

The corpus contains 93 unique employer or advertiser labels. Of these, 86 labels contribute one vacancy, six contribute two, and one contributes three. The largest label therefore contributes 3 of 101 vacancies, or 3.0% after rounding to one decimal place.

2.3 Coding and calculations

The analysis counts ten descriptive skill labels already stored in the observed_skills field of S1. Codes are non-exclusive: one vacancy can carry several labels. A label’s percentage is calculated as:

coded vacancies / 101 × 100

Percentages are reported to one decimal place. Co-occurrence means that both named labels appear in the same vacancy row; it does not establish that one capability causes, ranks above or depends on the other. For the workload-breadth summary, the number of codes per row was counted directly. The observed range is 1 to 10, the arithmetic mean is 5.18, and the median is 6.

Two different percentage denominators are used and are always identified. A corpus share divides a count by all 101 accepted vacancies. For example, 73 vacancies containing both executive communication and forecasting represent 72.3% of the full corpus. A conditional share divides that same overlap by a named subset. Because 78 vacancies contain forecasting, the 73-vacancy overlap represents 93.6% of forecasting-coded vacancies. Conditional shares help describe how tightly codes travel together within this sample; they do not estimate a probability for the wider labour market.

An absent code should be interpreted carefully. If six of 101 vacancies were not coded for executive communication, the result means only that the retained evidence did not support that label in those six rows. It does not prove that communication is absent from the actual jobs. Public advertisements differ in length, terminology and detail, so positive code observations are stronger evidence than conclusions drawn from non-observation.

The analysis can be reproduced by importing S1 as a tab-delimited file, confirming 101 data rows and 101 distinct source_url values, grouping once by source_family, splitting observed_skills on semicolon-space, and counting each label or named label pair by row. The machine-readable fact and claim ledger accompanying this draft records the values, formulae, evidence references and caveats.

2.4 Source coverage

Public ATS source family	Vacancies	Share of corpus
Greenhouse employer ATS	77	76.2%
SmartRecruiters employer ATS	11	10.9%
jobs.ashbyhq.com public ATS	5	5.0%
Lever employer ATS	4	4.0%
apply.workable.com public ATS	3	3.0%
www.careers-page.com public ATS	1	1.0%
Total	101	100.0%

The six-family coverage reduces dependence on a single employer or advertiser. It does not remove platform-selection bias: 77 vacancies (76.2%) come from Greenhouse employer ATS pages. Findings should therefore be treated as evidence of recurring patterns in this corpus, not as precise estimates for all employers or regions.

3. Findings

3.1 Communication is the most prevalent coded work signal

Executive communication appears in 95 of 101 vacancies (94.1%), making it the most frequently observed label by a wide margin. This is not merely a stand-alone communication signal. It co-occurs with forecasting in 73 vacancies (72.3% of the full corpus) and with management reporting in 72 (71.3%).

Communication relationship	Vacancies	Share of corpus
Executive communication	95	94.1%
Executive communication and forecasting	73	72.3%
Executive communication and management reporting	72	71.3%
Business partnering and executive communication	35	34.7%

Among the 78 forecasting-coded vacancies, 73 also carry executive communication, a conditional share of 93.6%. Among the 76 management-reporting-coded vacancies, 72 also carry executive communication, a conditional share of 94.7%. Among the 38 business-partnering-coded vacancies, 35 also carry executive communication, a conditional share of 92.1%. These conditional results make the integration visible without changing the full-corpus denominator used in the main prevalence table.

The defensible interpretation is that employers commonly describe FP&A analysis together with the need to explain it. The data do not reveal the quality, audience, channel or frequency of that communication. They do support treating decision-ready narrative, transparent assumptions and clear management messages as part of the work rather than as optional presentation polish.

3.2 Forecasting and reporting anchor the recurring operating cycle

Forecasting is present in 78 vacancies (77.2%) and management reporting in 76 (75.2%). They co-occur in 68 vacancies (67.3%). Budgeting appears in 56 (55.4%), and 54 vacancies (53.5%) combine budgeting with forecasting.

Planning and performance signal	Vacancies	Share of corpus
Forecasting	78	77.2%
Management reporting	76	75.2%
Budgeting	56	55.4%
KPI and performance analysis	55	54.5%
Forecasting and management reporting	68	67.3%
Forecasting and budgeting	54	53.5%

The pattern is consistent with a recurring planning-to-performance cycle. A forward-looking view is prepared or updated; performance is monitored; and the result is reported for management use. Because the sample is coded from vacancy evidence rather than observed workplace behaviour, the result should be expressed as an advertised-work pattern, not a universal workflow followed by every organization.

The conditional view reinforces that interpretation. Of the 78 forecasting-coded vacancies, 68 also include management reporting (87.2%). Of the 56 budgeting-coded vacancies, 54 also include forecasting (96.4%). These denominators are narrower than the full corpus, and therefore the percentages are higher than the corresponding 67.3% and 53.5% corpus shares. Both views are useful: corpus shares show prevalence across all accepted vacancies, while conditional shares show association within a named coded subset.

3.3 Modelling is usually connected to forward-looking work

Financial modelling appears in 64 vacancies (63.4%). It co-occurs with forecasting in 57 (56.4% of all vacancies). The evidence therefore supports modelling as a means of structuring a forecast or planning decision, not simply as production of a standalone spreadsheet.

Within the 64 financial-modelling-coded vacancies, 57 also contain forecasting, a conditional share of 89.1%. This does not prove that all modelling work is forecasting work. It does show that the two labels are closely connected in the retained evidence.

This finding does not identify a required software product, model architecture or technical standard. The source evidence is compatible with tool-agnostic professional learning focused on drivers, assumptions, checks, sensitivity, version control and explainability. It does not justify reproducing a vendor workflow or teaching interface-specific clicks.

3.4 Business partnering connects analysis with organizational action

Business partnering appears in 38 vacancies (37.6%). In 35 vacancies (34.7%), it appears together with executive communication. Although it is less prevalent than forecasting or reporting, its co-occurrence pattern supports a relational interpretation of FP&A: professionals work with budget owners and other decision participants, not only with financial data.

The result should not be inflated into a claim that every FP&A role has equal decision authority. Vacancy language can describe influence, support or responsibility without proving final authority. Sound practice therefore keeps assumptions, approvals and consequential decisions with a named authorized human.

3.5 Data quality and automation support, rather than replace, judgment

Data quality and automation appear in 24 vacancies (23.8%). Twenty-one vacancies (20.8%) combine that label with forecasting. The sample shows that data reliability, repeatability and automation are relevant to a substantial minority of advertised roles, while communication, forecasting and reporting remain more prevalent.

The overlap represents 87.5% of the 24 data-quality-and-automation-coded vacancies. That conditional rate is descriptive of a small subset and should be read with its denominator visible. It suggests a relationship worth teaching and testing, not a universal technology requirement.

This evidence supports incorporating source checks, reconciliation, traceability and controlled automation into professional practice. It does not support a claim that automation can own assumptions or decisions. AI-assisted analysis should remain reviewable: inputs, constraints, transformations, uncertainties and approvals need an auditable human-controlled record.

3.6 Variance and scenario work are narrower signals tied to the core cycle

Variance analysis appears in 23 vacancies (22.8%), and 22 (21.8%) combine it with management reporting. Scenario analysis appears in 14 (13.9%), and 13 (12.9%) combine it with forecasting.

Expressed conditionally, 22 of the 23 variance-analysis-coded vacancies also include management reporting (95.7%), while 13 of the 14 scenario-analysis-coded vacancies also include forecasting (92.9%). These high rates come from relatively small bases. Reporting both the counts and denominators prevents a conditional percentage from looking more representative than the evidence allows.

The lower frequencies do not make these capabilities unimportant. Frequency can be affected by how much detail an advertisement exposes and by employers’ choice of terminology. The co-occurrence results indicate where these practices sit when named: variance analysis connects actual and expected performance, while scenario analysis extends forecasting under alternative assumptions. The data cannot rank the expertise or time required.

3.7 FP&A vacancies combine multiple capabilities

The 101 vacancies contain between 1 and 10 observed-skill labels. The mean is 5.18 labels per vacancy and the median is 6. This descriptive breadth reinforces the connected-work finding: many advertisements combine several planning, analytical, communication and partnering signals.

This measure reflects the coding scheme and the amount of detail exposed in public vacancy evidence. It is not a validated job-complexity scale, and a role with fewer codes is not necessarily less senior or less demanding.

4. Implications for professional practice and learning

4.1 Build the full decision chain

The strongest learning design would connect the steps that the vacancy evidence repeatedly places together: clarify the management question; define drivers and assumptions; construct a baseline; prepare a budget, forecast or scenario; compare plan, forecast and actual performance; explain material differences; and record a decision-ready management narrative.

Teaching these as disconnected topics would miss the co-occurrence evidence. An integrated planning and performance pack is a more appropriate practice artifact than a collection of unrelated calculations because it makes inputs, logic, review, communication and ownership visible in one chain.

4.2 Treat communication as an analytical control

With executive communication coded in 94.1% of the sample, learners should practice concise explanations of what changed, why it matters, what remains uncertain and which decision is requested. A narrative should reconcile to the underlying numbers and distinguish observation, assumption, estimate and recommendation. This makes communication part of quality assurance, not simply presentation style.

4.3 Use models to expose assumptions

The modelling and forecasting relationship supports practice with driver trees, assumption registers, sensitivities, version comparisons and review checks. The goal is not to imply precision. It is to make the path from assumptions to forecast and decision legible enough for challenge and approval.

4.4 Connect business partnering to decision rights

Business-partnering practice should include mandate definition, stakeholder questions, ownership of inputs, challenge protocols, decision logs and escalation of unresolved assumptions. This develops collaboration without implying that an FP&A practitioner, a model or an AI system owns every consequential decision.

4.5 Place automation inside a controlled workflow

Data quality, automation and AI assistance should be taught as controlled support for tasks such as checking completeness, documenting transformations, generating alternatives for review, and testing whether a narrative reconciles with approved figures. Outputs remain estimates or decision support. Authorized people remain responsible for assumptions, approvals, external reporting and consequential action.

4.6 Translate the evidence into a four-part curriculum spine

The findings support four coherent learning blocks. This is an original educational interpretation of the evidence, not a certification structure or borrowed competency framework.

Learning block	Vacancy evidence informing it	Practical learning result
1. Mandate, drivers and controlled inputs	Business partnering: 38 of 101 (37.6%); data quality and automation: 24 of 101 (23.8%); executive communication: 95 of 101 (94.1%)	Define the decision, stakeholder roles, driver tree, assumptions register, data lineage and approval boundary before building a forecast.
2. Integrated planning, budgeting and forecasting	Forecasting: 78 of 101 (77.2%); financial modelling: 64 of 101 (63.4%); budgeting: 56 of 101 (55.4%); scenario analysis: 14 of 101 (13.9%)	Build a traceable baseline, budget, rolling forecast and bounded scenario set with explicit assumptions, sensitivities and review checks.
3. Performance review and management reporting	Management reporting: 76 of 101 (75.2%); KPI and performance analysis: 55 of 101 (54.5%); variance analysis: 23 of 101 (22.8%)	Reconcile expected and observed performance, construct a variance bridge, identify decision-relevant signals and prepare a controlled review pack.
4. Decision narrative, partnering and responsible AI assistance	Executive communication: 95 of 101 (94.1%); business partnering and executive communication: 35 of 101 (34.7%); data quality and automation with forecasting: 21 of 101 (20.8%)	Communicate implications and uncertainty, document decisions, challenge assumptions, and use AI assistance only through reviewable human-controlled steps.

Each block should produce an observable artifact rather than end with content recall. Appropriate artifacts include a planning mandate, driver tree, assumptions register, baseline, operating plan, budget, rolling forecast, scenario set, variance bridge, performance review, management narrative, decision log and responsible AI-use record. Combined, they form an integrated planning and performance pack that can be checked for internal consistency.

Assessment should follow the same chain. A learner should be asked to explain the business question, identify missing or unreliable inputs, connect assumptions to drivers, update a forecast under stated constraints, reconcile a performance bridge, distinguish observation from estimate, and present a decision request with uncertainty visible. Quality criteria should include traceability, arithmetic reconciliation, version discipline, clarity, proportionality of scenarios, appropriate escalation and explicit human approval. Software speed or visual polish should not compensate for an unsupported assumption or unreconciled figure.

4.7 Avoid false precision in learning claims

The evidence justifies curriculum relevance, not a promise that completion will produce a job or improve a forecast by a stated amount. Learning outcomes should therefore use observable verbs such as define, construct, reconcile, test, explain, document and review. They should avoid guarantee, master, certify or predict unless a separate verified standard supports that wording. The same discipline applies to AI practice: a learner may use an AI system to propose checks or alternative narratives, but must verify inputs, reject invented facts, reconcile every quantitative statement and record the authorized human decision.

5. Practical application: an evidence-aligned FP&A workflow

The following application illustrates how the aggregate findings can be translated into professional practice without inventing an employer case or reproducing a vacancy description. Consider a fictional organization preparing a quarterly reforecast after operational assumptions change. The exercise uses synthetic records and makes no claim about a real company, sector or financial outcome.

Step 1 — establish the mandate. The practitioner records the decision to be supported, the forecast horizon, required review date, responsible budget owners, authorized approver and boundaries on use. This operationalizes business partnering and executive communication before analysis begins. A clear mandate also prevents the model from expanding into unrelated valuation, investment or external-reporting advice.

Step 2 — map drivers and evidence. The practitioner constructs an original driver tree linking operating activities to financial effects. Each driver receives a source, owner, unit, update frequency and confidence note. Missing, stale or conflicting inputs are recorded rather than silently replaced. Automation may flag anomalies or assemble a reconciliation queue, but it does not decide which assumption is approved.

Step 3 — register assumptions and versions. The assumptions register separates observed values, management assumptions, modelled estimates and unresolved questions. Every material revision records who changed it, why, when and which forecast version it affects. This makes later variance explanations and management reporting reproducible.

Step 4 — prepare the baseline and forecast. The practitioner checks opening values, calculation logic, sign conventions, time periods, totals and cross-schedule consistency where applicable. The forecast is then updated from approved drivers. A model is accepted for review only when another person can trace a material output back to its input and assumption.

Step 5 — construct bounded scenarios. Alternative scenarios vary a small, disclosed set of decision-relevant assumptions. They are not presented as predictions or probabilities unless those interpretations have separate evidence. The scenario set explains what changes, what stays fixed, which risks are exposed and which management action each scenario could inform.

Step 6 — bridge performance. The practitioner reconciles the prior plan or forecast with the updated view and, when available, actual performance. The bridge distinguishes timing, volume, price, mix, efficiency and assumption effects only where the organization’s synthetic case data support those categories. Residuals remain visible and are investigated; they are not hidden inside a convenient narrative.

Step 7 — communicate the decision. The management narrative states the question, observed change, main drivers, uncertainty, scenario implications and requested decision. Every quantitative statement reconciles to the approved analysis. Charts or dashboards may support the message, but the decision logic must remain understandable without relying on interface-specific demonstrations.

Step 8 — close the review loop. Review comments, approved actions, owners and next-update triggers enter a decision log. If AI assistance was used, an accompanying record identifies the input scope, prompt purpose, constraints, generated suggestions, human edits, rejected material and final checks. The record makes clear that the authorized human—not the tool—owns the approval and consequential decision.

This workflow is evidence-aligned because it connects the sample’s most frequent signals rather than treating them as isolated topics. It is still an educational model, not a claim that all 101 employers use the same process.

6. Limitations

This analysis has material limitations:

• Purposive sample: The 101 vacancies are a selected evidence corpus, not a random sample or a census. Percentages describe this corpus only.
• Point-in-time status: Sources were retrieved on 24 August 2026. Vacancies can change, expire or become unavailable.
• English-language and public-access bias: Public discoverability, search indexing and ATS accessibility shape inclusion. Non-English and non-public hiring channels are not represented.
• ATS concentration: Greenhouse accounts for 77 vacancies (76.2%). Six source families are present, but platform coverage is not balanced.
• Incomplete geography: Thirty-four vacancies (33.7%) have location recorded as “Other or unspecified.” The corpus does not support a reliable regional comparison.
• Advertised work, not observed practice: Vacancy text reflects recruitment communication. It does not prove how work is actually performed after hiring.
• Non-exclusive coding: Skill percentages overlap and sum to more than 100%. Code frequency does not measure proficiency, seniority, time allocation or business impact.
• Short evidence excerpts: Rights-safe excerpts reduce reproduction of source text, but they also limit contextual interpretation.
• No outcome inference: The corpus does not support salary, employment, promotion, forecast-accuracy, financial-performance, revenue, accreditation or employer-recognition claims.
• No regulated advice: Findings are general professional-education evidence and do not interpret accounting standards or provide accounting, audit, investment, tax, legal or external-reporting advice.

7. Conclusion

Across 101 current public vacancies, the dominant picture of FP&A work is integrated rather than narrowly technical. Employers in this sample most often connect executive communication with forecasting, management reporting, modelling, budgeting and performance analysis. Business partnering, data quality, automation, variance analysis and scenario analysis appear less often, but their co-occurrences place them inside the same planning-to-performance system.

The evidence supports professional development centered on an operating cycle: drivers and assumptions become plans and forecasts; results are compared with expectations; implications are communicated; and decisions and revisions are recorded under human authority. The evidence does not justify promises about individual careers, organizational outcomes or forecast accuracy. Its proper use is to make learning priorities more transparent, current and reviewable.

8. Rights and responsible-use statement

This report contains original analysis of aggregate, coded observations. It does not reproduce full vacancy descriptions, third-party templates, certification structures, exam content or proprietary vendor methods. The row-level evidence ledger retains source URLs and short necessary excerpts for verification. Employer, advertiser, platform and product names are used only for descriptive source identification; associated rights remain with their owners.

The report is general professional education. Forecasts, scenarios and AI-assisted analyses are estimates and decision support, not guarantees. A named authorized human remains responsible for assumptions, approvals, external reporting and consequential decisions.

9. References and source ledger

ID	Evidence artifact	Role in this report
S1	accepted-vacancies-2026-08-24.tsv	Row-level source ledger: 101 accepted vacancies, public URLs, source families, retrieval dates, short necessary excerpts, observed-skill codes and acceptance reasons.
S2	vacancy-corpus-validation-2026-08-24.json	Deterministic corpus counts, source-family counts, skill counts, validation checks and limitations.
S3	professional-learning-intent-v1.md	Demand-triangulation and learning-intent context; used only for scope and interpretation, not to add unverified vacancy counts.
S4	portfolio-legal-prequalification-v1.md	Portfolio differentiation, legal boundaries, credential exclusions and rights-safe language.
S5	gates/gate-evaluation-2026-08-24.json	Conditional-go record, evidence summary and mandatory scope exclusions.

All five artifacts are located in the durable evidence bundle research/fpa-professional-2026/. S1 is the authoritative row-level ledger for the vacancy analysis; S2 is the authoritative validation summary. The accompanying JSON fact and claim ledger records SHA-256 fingerprints for all five inputs so a later reviewer can confirm that the draft was evaluated against the same evidence versions.

Appendix A. Reproducibility protocol

This protocol allows a reviewer with the five fixed evidence artifacts to recreate every numeric statement in the report without visiting or scraping the live vacancy pages again.

A.1 Pin the evidence version

Calculate the SHA-256 fingerprint of S1 through S5 and compare each value with the accompanying JSON ledger. A mismatch means the evidence version has changed and the report should not be treated as reconciled until every calculation is rerun. This step separates reproducibility of the analysis from later changes to live vacancies.

A.2 Load and validate the row structure

Import S1 as UTF-8 tab-delimited data with its header row. Require exactly these nine fields: employer_or_advertiser, role_title, jurisdiction_or_location, retrieval_date, source_family, source_url, minimal_supporting_excerpt, observed_skills and acceptance_reason. Reject an empty required field rather than inferring its value. Confirm 101 data rows, a single retrieval date of 2026-08-24, and 101 distinct source URLs.

A.3 Reproduce coverage checks

Group rows by employer_or_advertiser. Confirm 93 groups and the frequency distribution of 86 groups with one row, six with two and one with three. Divide the largest count, 3, by 101 and round to one decimal place to reproduce 3.0%.

Group rows by source_family. The counts must be 77 Greenhouse, 11 SmartRecruiters, 5 Ashby, 4 Lever, 3 Workable and 1 careers-page.com. Divide each count by 101 and round once at the end. The rounded shares must be 76.2%, 10.9%, 5.0%, 4.0%, 3.0% and 1.0%, summing to 100.0% at the displayed precision.

Count rows whose jurisdiction_or_location is exactly Other or unspecified. The expected count is 34 and the corpus share is 33.7%. Do not redistribute these rows across inferred regions.

A.4 Reproduce skill prevalence

For each row, split observed_skills on semicolon-space and count the presence of each label once in that row. The expected counts, in descending order, are 95 executive communication, 78 forecasting, 76 management reporting, 64 financial modelling, 56 budgeting, 55 KPI and performance analysis, 38 business partnering, 24 data quality and automation, 23 variance analysis and 14 scenario analysis. Divide each by 101 and round to one decimal place.

Count the number of labels per row to reproduce a minimum of 1, maximum of 10, arithmetic mean of 5.18 and median of 6. Do not interpret this descriptive distribution as a seniority or complexity score.

A.5 Reproduce co-occurrences and conditional shares

For each named pair, count a row only when both exact labels are present. Divide by 101 for the corpus share. For a conditional share, divide the overlap by the stated base label count:

Pair	Overlap	Corpus share	Conditional calculation	Conditional share
Executive communication + forecasting	73	72.3%	73 / 78 forecasting-coded vacancies	93.6%
Executive communication + management reporting	72	71.3%	72 / 76 management-reporting-coded vacancies	94.7%
Forecasting + management reporting	68	67.3%	68 / 78 forecasting-coded vacancies	87.2%
Financial modelling + forecasting	57	56.4%	57 / 64 financial-modelling-coded vacancies	89.1%
Forecasting + budgeting	54	53.5%	54 / 56 budgeting-coded vacancies	96.4%
Business partnering + executive communication	35	34.7%	35 / 38 business-partnering-coded vacancies	92.1%
Variance analysis + management reporting	22	21.8%	22 / 23 variance-analysis-coded vacancies	95.7%
Data quality and automation + forecasting	21	20.8%	21 / 24 data-quality-and-automation-coded vacancies	87.5%
Scenario analysis + forecasting	13	12.9%	13 / 14 scenario-analysis-coded vacancies	92.9%

These conditional calculations are directional. For example, 73 divided by 78 answers how many forecasting-coded vacancies also contain executive communication. Reversing the denominator would answer a different question. Any future use must retain the overlap count, the base and the direction together.

A.6 Reconcile before release

Compare all recreated values with S2 and the JSON fact/claim ledger. Confirm that every numeric claim in the report occurs in the ledger, that source hashes still match, and that no code or narrative converts a non-observation into proof of absence. Publication metadata, authorship, review, DOI, public PDF and canonical links require separate verified records and are outside this non-published draft.