1. Introduction

The aim is a target list of about 1,500 firms with 8–300 staff (10–250 at the core), drawn from all 205,551 active companies registered in the West Midlands. The client is a regional quasi-public growth company that consults with firms to raise gross GVA, so firms are not excluded for serving local markets.

Headcount is the deciding filter, and it is hard to get at scale:

• Registers don't carry it. Companies House records names, addresses, industry codes and accounts types, but no staff numbers.
• The figure exists, but in the accounts. Every company must state its average number of employees in the notes to its accounts (Companies Act 2006, s.411). That's a monthly headcount of people on contracts of employment, averaged over the year.
• The accounts come in three forms: machine-readable files for most firms, scanned paper or PDF documents for about one firm in 30, and nothing yet for companies less than about 21 months old.
• The no-cost rule. Commercial databases sell estimates, but the brief ruled out anything that costs money.

This paper describes how each form was handled, the problems met and how they were fixed, how accuracy was checked, and what the resulting dataset shows. The selection of the final 1,500 and the growth projections follow in a separate section.

2. Data sources

Every input is free and official, so the whole dataset can be rebuilt and checked by anyone.

Three other routes were assessed and rejected:

• A public web-crawl company dataset turned out to be a 2019 snapshot, older than every new firm.
• Automated LinkedIn tools break LinkedIn's terms, cost money, and count profiles rather than payroll.
• Web searches found no headcount for any of 10 sampled new firms.

3. Methodology

Each company went through up to four stages, from the cheapest and most reliable source to the hardest, and only the companies still unresolved moved on to the next stage. One program carries out every stage (section 3.1); section 3.6 shows how its output was verified, and section 3.7 sets out the growth-forecast model built on it.

3.1 The program

Every stage is carried out by one program, kept in a private Git repository (wm-growth-model) so that any figure in this paper can be rebuilt from free public files. It has two halves: data collectors written in Node.js 24 with Windows PowerShell, and a forecasting package in Python 3.12 (section 3.7).

Five design rules apply throughout:

• Point in time. Every value is stored with the date it became public, so a test set in the past can only see what was known then.
• Deterministic. Random choices use fixed seeds. The national sample of companies uses a stable hash of the company number (FNV-1a), so every machine draws the same sample.
• Resumable. Each bulk file's output is written to a temporary file and renamed only when complete, with a record of its counts. Finished files are skipped on re-runs.
• Tested. 16 unit tests on synthetic filings run on every change, on a Windows build server. A regression test compares new output with the published results before anything is replaced.
• No data or personal data in the code. Inputs are re-downloaded or read from the project folder. The API key is held in an environment variable, and officer names are never stored.

3.2 Reading electronic accounts at scale

About three-quarters of UK accounts are filed as inline XBRL (iXBRL): web pages in which every figure is tagged with a concept, a reporting period and a unit. Companies House publishes them in monthly zip archives of 1.6–4.1 GB, and in daily files kept for 60 days.

Remote access. A zip file's index (its central directory) sits at the end of the file. The reader fetches the last 64 KB with an HTTP range request and finds the end-of-directory record. For files over 4 GiB it follows the ZIP64 locator. It then downloads only the index, about 20–50 MB per monthly file.

Each entry's name carries the company number and balance-sheet date, for example Prod224_2512_09564242_20250430.html. Entries are then read in one of two ways:

• Selected companies. One range request per entry. The 15 Oct run used this, moving about 1–1.5 GB instead of about 31 GB.
• Whole files. Contiguous 32 MB blocks, cut into entries at the offsets recorded in the index. This mode builds the national history.

Entries are decompressed (DEFLATE) and parsed in parallel worker threads.

Fact model. Each filing is reduced to numeric facts. A fact is a concept, a context, a unit, a value and a stated precision.

• Context: gives the period, as a start and end date for flows or an instant for balances, plus any dimension members (for example creditors due within one year, a staff category, or group versus company).
• Unit: separates pounds from other currencies and from pure counts.
• Plain XBRL: files without inline tagging are read the same way, from their elements' context and unit references.

Value rules. The displayed text is converted by its stated format:

• digit grouping, with point or comma decimals
• "zero-dash" and "fixed-zero" marks read as 0
• worded numbers from "nil" to "twenty"

The value is then the displayed number × 10^scale, negated when the sign attribute is "−". Facts marked nil are dropped, and a repeated concept–context pair is kept once.

Staff figure. The legal disclosure is the average number of employees (Companies Act 2006, s.411). It is resolved by seven rules, in order:

1. Use the concept AverageNumberEmployeesDuringPeriod, or its older UK GAAP equivalents.
2. The current year is the context ending on the balance-sheet date. The prior year is the latest earlier period end. Both are kept, so every filing gives two years.
3. An undimensioned total wins. If only dimensioned figures exist and one dimension separates group from company, the consolidated group figure is used. Group and company are never added together. Staff categories are summed within the chosen entity.
4. A negative scale on a headcount is ignored and the displayed number kept. From about August 2025 some filing software tags headcounts with scale −2, which would turn 14 staff into 0.14 (section 3.6).
5. A whole number equal to the period's year, give or take one, is rejected as a mis-tagged column heading. The exception is when another staff figure in the same filing is within a factor of two, so a firm of 2,025 with 1,950 the year before is kept.
6. If no usable tagged figure remains, a text reader looks for the employee sentence ("the average number of employees during the year was 23 (2024: 21)"). It deletes dates, years and note numbers, then takes the first count.
7. Dormant accounts carry no employee note. They are detected from the dormant flag (EntityDormantTruefalse = true) or a "Dormant Company Accounts" title, and classed as having no staff.

Other facts and fields.

• Full fact set: kept for filings showing 4 or more staff in either year, for every West Midlands company, and for any filing without a staff figure. Smaller filings keep headline totals only: turnover, net assets, current assets, debtors, cash and fixed assets.
• Filing-level fields: entity name, registered number, period dates, accounts type, audit status, accounting standard, postcode and town, and which statutory statements are present (for example the s.477 audit exemption).
• Text flags: a material uncertainty over going concern, a qualified audit opinion, wording about a business combination, and furlough.

Availability date. A filing in a monthly file counts as public from the 1st of the following month, and one in a daily file from the next day. This conservative rule is what lets the forecasting panel replay any past date without seeing later filings (section 3.7).

Coverage and speed.

• Coverage: thirteen months of filings (Sep 2025 to 14 Oct 2026) gave 152,781 filings for 132,374 companies, with no download errors. A second pass over Jan 2024 – Aug 2025, for companies still unresolved, added about 5,800 figures.
• Speed: on a 15 W laptop processor with 11 worker threads, a 72 MB daily file of 9,977 filings was parsed in 11–17 seconds with no errors. Stored output averages about 220 bytes per filing, compressed.

3.3 Joining the company register

Each company's latest accounts were joined to the Companies House register snapshot. This adds accounts type, industry codes, incorporation date, status and registered charges, and assigns a result: a size band, dormant, or unknown with a reason. Every figure carries its source and a confidence level.

3.4 Scanned accounts: retrieval and OCR

For the 6,656 companies with no electronic accounts since January 2024, the Companies House API supplied each one's latest accounts: 6,295 scanned PDFs, 123 electronic documents and 238 with no usable document. The scans are images with no text layer, so they were read with the OCR engine built into Windows, which needed no extra software.

• Finding the right page. The employee note sits 55–95% of the way through a set of accounts, so reading starts at the middle, runs to the end, then works back. It stops once the note is found.
• Rebuilding tables. OCR returns words with positions but loses table structure, so rows and columns were rebuilt from word coordinates.
• Reading the figure. The figure is read from a sentence, a "no employees" statement, a table total, the sum of staff categories, or a figure on the heading line. It is scored high, medium or low confidence.
• Throughput. Up to 13 parallel workers ran at the machine's limit. Any document where the early stop missed the note was re-read in full: 240 documents, 5,152 pages.

3.5 Firms with no published accounts

Companies less than about 21 months old have not yet filed accounts, so no headcount exists for them. Instead they were ranked using free signals and a statistical model, and linked to any established firm under the same ownership (section 7).

3.6 Verification and re-processing

The new reader was tested against the old one before any output changed. The test exposed a scale defect that had dropped 1,751 firms out of the 8–300 band. The affected filings were re-read and every output rebuilt, which moved the band from 10,673 to 12,417 firms.

Unit tests. Sixteen synthetic filings each isolate one case:

• Periods and values: prior-year figures; scale, sign, zero-dash and nil values; dimensions
• Staff-figure rules: group versus company; a negative scale on a headcount; dormant accounts; a year heading tagged as staff; a genuine 2,025-staff firm
• Formats and storage: plain XBRL; untagged text; the size pre-filter; slicing entries out of zip blocks

They contain no real company data and run on every change.

Regression test. The 14 Oct 2026 daily file (9,977 filings) was re-parsed with the new reader. Each of its 382 West Midlands filings was compared with the published figure, and every difference was classified:

Scope of the defect. The old figures held 9,466 fractional headcounts. They start in filings from about August 2025, which points to a change in one software package. A real 100, 200 or 300 would have shown as a whole 1, 2 or 3, so every such figure from July 2025 on was re-read as well. In all, 75,937 filings across 26 bulk files were re-read, with no errors.

Result. 9,563 figures changed:

• 9,560 scale corrections: 1,751 firms moved into the band.
• 5 year headings tagged as staff: 2 were resolved from the note text and 3 have no usable figure.
• 7 firms left the band:
  • 4 genuinely employ 814 to 1,310 staff, as their accounts notes confirm.
  • 1 has a single employee.
  • 2 give no usable figure.

Check. Eight corrected firms were drawn at random with a fixed seed. All eight matched the employee note in their accounts: 14, 96, 27, 24, 20, 272, 8 and 9 staff.

Rebuild. The original merge steps were re-run on corrected copies of the stored figures: register join, ownership and sponsor signals, the ranking model, the final CSVs and the zip.

• The original figures and files were kept unchanged.
• Each corrected firm's row says what was changed and what it read as before.
• A separate file lists every firm whose figure or result moved.

3.7 The growth-forecast model

The 1,500 firms are chosen by each firm's forecast chance of hitting a size-based growth target over the next 12 months. The design below was fixed before any model results were seen, so it can be tested honestly. It is being built in stages, and an extra component is kept only if it beats simpler rules on years the model never saw.

Success test. Let B be the firm's financial year ending between 16 Oct 2026 and 15 Oct 2027, and T the year after it. Both are annual averages, so comparing T with B measures growth over a window that, across firms, is centred on the next 12 months (±6 months). For any one firm the horizon runs from about a year (October year-ends) to about 2.5 years (September year-ends), and the outcomes are known only by mid-2029.

With n the average staff in B, the target rate is:

• r(n) = 20% when n < 50
• r(n) = 20% − 13% × ln(n/49) ÷ ln(250/49) when 50 ≤ n ≤ 250, giving about 16.6% at 75, 14.3% at 100, 11.1% at 150 and 8.8% at 200
• r(n) = 7% above 250

A firm hits its target if either route is met:

• Staff route: staff in T ≥ (1 + r) × staff in B, and at least 2 more staff when n < 50.
• Revenue route: real turnover in T ≥ (1 + r) × real turnover in B. This route applies only where turnover is published for both years and both periods run 50–54 weeks. Turnover is converted to constant prices with the GDP deflator.

The remaining outcomes are classed as follows:

• Closure: dissolution, liquidation, administration, or a pre-pack into a successor company is a miss.
• Paper accounts: accounts filed only on paper are read by OCR rather than counted as misses.
• Late accounts: accounts still missing three months after the deadline count as misses. A sensitivity run excludes them.
• Acquisitions: growth from buying another company is deducted where it can be detected.

Sensitivity variants are fixed in advance:

• a rebound-proof target, (1 + r) × the larger of B and the latest reported figure
• stepped targets of 20, 15, 10 and 7%
• nominal revenue
• no +2 rule

Point-in-time panel. The unit is a company in a financial year, drawn from every UK electronic filer from January 2017 to October 2026. Prior-year comparatives extend this back to financial years 2015–16. The national scope matters because the region alone has too few larger firms to learn from: only 77 West Midlands companies have 251–300 staff.

Snapshots are taken as of 15 April and 15 October each year. Each one uses only what was public before its date:

• filings, by availability date
• API events, by filing date
• economic series, by publication date
• forecasts, as published at the time

For the remaining fields:

• Feature values: come from each figure's first publication.
• Labels: come from the target-year filing's own current and comparative figures, so both years share one definition.
• Register fields: fields that exist only as today's snapshot (status, industry code, address) are never training inputs. Accounts type, audit status and address are taken from each filing.
• Dissolved firms: included, with their status, closure date and postcode from the Companies House API (which covers dissolutions since 2010), so the panel does not learn from survivors only.
• Paper filers: added from OCR, because paper filing rises sharply with size: 5% of firms with 8–9 staff, 59% at 100–249 and 78% at 250–300. The sample covers every West Midlands paper filer and a national sample of about 3,000 with 50–400 staff.

Inputs.

Directors' sex, nationality and age are excluded.

Models. The models compete as a tournament:

1. Benchmarks: random choice within quota; last year's growth repeated; the 3-year trend; the sector average; a plain logistic regression; the earlier points-based score.
2. Challenger: a penalised logistic regression with splines on about 30 inputs.
3. Main model: gradient-boosted trees (LightGBM) fitted directly to the binary label.
   • The gap since the latest accounts is an input.
   • Monotone constraints are used only where the direction is certain.
   • Tuning is nested and time-ordered, on a grid fixed in advance.
   • The model is refitted 20 times on resampled data and averaged.
4. Add-ons, each kept only if it beats the main model by a pre-registered margin. The nowcast, the events layer and the closure score are tested in v1; the other three only after v1 ships, to limit the comparisons made on six vintages of data:
   • a base-year staff nowcast
   • an API events layer
   • a whole-number headcount model for firms under 50
   • pooled sector × area effects
   • a closure-risk score
   • the paper-filer sample
5. Calibration: logit p = a + b · logit(score), plus a West Midlands offset and size-band offsets. It is fitted on out-of-time predictions from folds other than the one being scored.
6. Economy: a scenario shift on the logit for each of about eight sector groups, estimated from how yearly hit rates moved with sector employment growth and shrunk toward one national value. The history is short, so this is a judgement overlay: it moves the headline probability and its range, and never changes the ranking. Base, downside and upside scenarios for 2027 come from the OBR forecast of 28 Oct 2026, the Bank of England, its Decision Maker Panel and the regional PMI.
7. Reasons: each firm's main reasons come from SHAP values.

Selection.

1. Exclusions: charities and not-for-profits, dormant firms, and firms that aren't active. The other exclusions are switches.
2. Groups: one entry per group.
3. Quotas: places are shared in proportion to the eligible pool across six size bands (8–9, 10–19, 20–49, 50–99, 100–249, 250–300).
4. Ranking: within each band, firms are ranked by calibrated chance of a hit, in 5-point tiers, then by expected added gross GVA. Each firm must beat its band's base rate, and each council area keeps at least 75% of its share of places (section 9).

Firms with 8–9 staff also need dated evidence of growth: one hard item or two soft ones.

• Hard items:
  • Sustained growth: up in 2 of the last 3 years, by 2 or more in net, and not a rebound.
  • Young and scaling: 6 years old or less, and never fallen.
  • Money growth: real turnover up 20%, or debtors or current assets up 25% with positive net assets.
• Soft items: cash raised, a new charge, a new director, or a move to commercial premises.

Gross GVA is operating profit plus staff costs plus depreciation and amortisation, where published. Otherwise it is staff × West Midlands GVA per job for the sector, adjusted for the firm's pay.

Validation.

• Test dates: the development dates are Oct 2020, Apr 2021, Oct 2021, Apr 2022 and Oct 2022. A lockbox of Apr and Oct 2023 is opened once, at the end.
• Embargo: training uses only snapshots whose outcomes were known at each test date (about a 36-month embargo). This mirrors the live forecast, which trains on snapshots up to October 2023.
• Geographic holdout: the model is trained outside the region and tested inside it.

The primary pass marks are:

1. Hit rate: the shortlist's staff-route hit rate (the primary measure, section 9) beats the best benchmark. On the development origins the pooled 90% block-bootstrap interval must lie above zero and the point estimate must be at least 2 points higher; on the lockbox, treated as one pooled test, the point estimate must be higher.
2. Calibration slope: between 0.8 and 1.2.
3. Overall level: within the scenario range.

Secondary checks are reported with Benjamini–Hochberg correction:

• Calibration by size band, sector, area, firm age, filing format and turnover publication
• Accuracy: AUC (published studies reach 0.62–0.72), Brier and ranked-probability skill, and interval coverage
• Robustness: stability; removing one data source at a time; drift; and a time-travel rebuild of October 2023 inputs

If the model fails, the best simple rule that passes is used, and the report says so.

4. Challenges and solutions

Most of the work went into problems that only showed up once real filings were read. Each fix was tested on the documents that exposed it before being applied to the whole set.

5. Validation

Accuracy was checked by eye against the source documents at every stage. After the final corrections, 36 of 37 verified figures (97%) match the accounts, up from 26 of 31 before them.

Two safeguards apply to the output itself:

• Every figure carries its source and confidence, so a reader can tell a tagged electronic figure from one read off a scan.
• Estimates for unsized firms sit in separate columns and files, never mixed with reported counts.

6. Results

12,417 companies fall in the 8–300 band, about 6% of the active register. 146,095 companies (71%) were sized or classified; most of the rest are too new to have filed accounts.

Within the band, 10,017 sit in the core 10–250 range, 2,323 have 8–9 staff and 77 have 251–300. These counts include the 1,751 firms restored by the scale correction (section 3.6).

Where the staff figures came from:

The scanned accounts matter more than their number suggests. They make up 4% of figures but 14% of the band, largely because PDF filers lean towards fuller accounts formats (full, group, medium, small and subsidiary accounts make up 3,885 of the 6,656). Of the 1,737 in-band figures read from scans, 1,224 are high confidence, 491 medium and 22 low.

7. Firms without published accounts

59,456 companies have no staff figure. 57,671 of them are simply too new: first accounts are due 21 months after incorporation, and as of 1 October 2026 none of these had filed any. The rest are 1,431 scanned accounts with no employee note, 238 with no accounts document, 87 electronic accounts that omit the figure, and 29 unresolved.

They can't be counted, but they can be ranked. Of 38,454 West Midlands firms set up in 2022–24 that have since filed accounts, 3.3% have 8 or more staff. That rate varies sharply by sector: about 15% for restaurants and takeaways and for care homes, 8–11% for home care, cleaning and recruitment, and under 2% in 22 of the 52 sectors with at least 50 young firms.

7.1 Signals

Each company was matched against free signals that point to real staff:

• Ownership: the same person, matched on name and birth month and year, or a parent company that already runs an established firm with staff.
• Visa sponsor licence: the firm is licensed to employ overseas workers.
• Care Quality Commission registration: care homes (with bed numbers) and home-care services, matched by exact company number.
• Food premises: a food business with a hygiene rating trading under the company's name or at its address.
• Other: registered charges (usually bank lending), a commercial-premises address versus an address shared by many companies, and company type.

7.2 Model

A logistic regression with 73 features was trained on the 38,454 young firms to estimate each unsized company's chance of having 8 or more staff.

• Separation: on 7,772 firms held back from training it scores an AUC of 0.83, where 0.5 would be guessing.
• Calibration: its probabilities are honest. The top decile it picks has a 15.6% actual rate against 17.3% predicted, about five times the base rate.

7.3 What it finds

Summing the probabilities, the model expects about 1,729 of the 59,456 to have 8 or more staff. Most of them can't be singled out yet; they are spread thinly across thousands of low-probability firms.

A prioritised supplementary list of 2,471 new firms (1,953 of them ownership-linked) should hold about 268 with 8 or more staff. That's a hit rate of about 11%, against 2.7% for new firms overall.

The new firms are not the core of the target list, but ignoring them would systematically miss the newest businesses. About 10,800 of them must file first accounts by March 2027, and a monthly re-run will replace these estimates with real counts as they do.

8. Limitations

The figures are as good as the accounts behind them, and eight limits apply:

• Headcount is not workforce. The legal figure counts people on contracts of employment, averaged monthly, with part-timers counting as one. It excludes subcontractors, agency staff and workers employed by a sister company. Firms in construction, care and hospitality may therefore operate larger than they report.
• Accounts lag reality. The latest accounts are usually 6–21 months old, so a fast-growing firm may already have moved past its reported band.
• Scanned figures carry some risk. 47 low-confidence figures (22 of them in the band) are flagged for manual checking. Medium-confidence figures were mostly correct in testing but weren't all verified by hand. About 1,400 scanned accounts contain no employee note at all: mostly small charities, clubs and slimmed-down micro-entity accounts.
• Estimates for unsized firms are probabilities, not counts. The model learned from firms that survived long enough to file accounts, and its signals are measured today. Its expected totals are therefore an upper bound.
• Electronic tags can be wrong. The reader catches the tagging errors found so far, a misapplied scale and a year heading in place of the count (section 3.2). A filing that simply states a wrong number is kept as filed: one firm reports 2,026 staff against 2 the year before. The forecasting panel's plausibility checks are designed to catch jumps of this kind.
• The forecast is odds over a long horizon. Each firm's probability covers growth 1 to 2.5 years out, and the first outcomes are known in 2028–29. Published research on spotting high-growth firms in advance reaches AUC of about 0.62–0.72, so a shortlist can roughly double or treble the base hit rate but cannot promise that any one firm will grow.
• The rules change during the target years. Companies House reform makes turnover observable for small firms from 2027, so the either-route probability is an estimate and the staff-route probability is the measure that survives the change (section 9).
• Growth by acquisition can look organic. Small-company accounts don't disclose business combinations, so some 20%+ jumps in the 50–300 band are purchases or staff moved between group companies. Detected cases are flagged; undetected ones remain.

9. Design review and decisions

A review of the forecast design on 16 October 2026, before any model code was written, found 17 weaknesses. Four could have broken the result; each is now built into the design in section 3.7 and in the model configuration. The four decisions the review left open were taken the same day and are recorded here so they can be revisited.

Decisions taken on 16 October 2026

What the review changed

Status. The success test, the no-leakage rule and the selection logic are now code (model/labels.py, model/asof.py, model/select.py), with tests that pin the rate curve, the +2 rule, the 49/50 boundary, deflation, closure and late outcomes, the leakage canary, quota apportionment, one entry per group and the area floor. Since then the rest of the pipeline has been written and tested on synthetic data: the filing-events and accounts-history collectors, the OCR reader for both years and the money items, the economic data collector with publication dates on every value, the panel that joins them per firm per as-of date, and the model stage (benchmarks, the transparent challenger, the LightGBM main model, calibration, the tournament with its pass marks, reasons, the forecast and the selection). Nothing has yet run on real data: the first step is a 100-firm pilot of the collectors, then full collection, then pre-registration of the pass marks before any backtest.

Sources

• Companies Act 2006, section 411: information about employee numbers and costs
• Companies House: accounts data product, monthly files
• Companies House: accounts data product, daily files
• Companies House: historic monthly accounts data
• Companies House: free company data product
• Companies House: persons with significant control snapshot
• Home Office: register of licensed sponsors (workers)
• Care Quality Commission: using CQC data
• Food Standards Agency: food hygiene rating open data
• Companies House: public data API (company profile, filing history, officers, persons with significant control, insolvency)
• XBRL International: Inline XBRL 1.1 specification and transformation registry
• Financial Reporting Council: UK GAAP (FRS 102 and FRS 105) XBRL taxonomies
• HM Treasury: The Green Book, GDP deflators for converting to constant prices
• Office for National Statistics: Business demography, UK: 2024 (high-growth enterprises by region)
• Bank of England: Decision Maker Panel; Office for Budget Responsibility: Economic and fiscal outlook, October 2026
• Coad, A. and Srhoj, S. (2020). Catching gazelles with a lasso: big data techniques for the prediction of high-growth firms. Small Business Economics, 55(3)
• Ke, G. et al. (2017). LightGBM: a highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems 30