How the score works, and how well

Salenovo ranks newly incorporated companies by how likely they are to become real trading businesses, classifies what they are, and verifies websites and premises. This page says exactly how each part is built, what it was tested on, what the results were, what's deliberately left out, and what it can't do. In the week of the latest index, 15,020 companies were incorporated in the UK. Every number below is from a held-out test unless it says otherwise.

What happens to new companies

"Real" here means active on the register with non-dormant accounts filed. To measure it without the usual blind spot (Companies House's bulk file only lists companies that still exist), every company incorporated in every month of 2022, 2023 and 2024 was traced through the register's advanced search, dissolved companies included: 2,502,259 companies, reconciled to the register's own daily incorporation counts to within 0.5% in every month (registered overseas entities, which aren't companies, excluded). Companies are compared at the same age, because a company formed in 2024 has had less time to fail than one formed in 2022.

About a third of incorporations become trading companies; more than half are gone within three years. Very few fail in the first year (96% of companies are still registered at twelve months), and then a third of every cohort disappears in the second year, almost all of it at exactly seventeen months of age: the confirmation statement falls due at twelve months, and a company that never files it is struck off by the registrar about five months later, in a weekly batch processed on Tuesdays. These aren't businesses failing; they're companies that were never started, being removed. Companies that traded first and then closed are a minority of dissolutions. Since October 2024 the registrar has also been removing companies under its new powers: where a registered office turns out not to be a real address, the company is moved to the Companies House default address and struck off if none is supplied within 28 days. About one dissolution in eleven in this cohort happened that way, and a quarter of all dissolutions at the peak in early 2025, mostly online-retail and wholesale companies; those companies no longer carry their original address, so they are missing from the council figures on this site, which therefore understate death rates slightly. A further 18% of the survivors carry property, holding-company or residents'-management SIC codes: they file accounts, so they count as "trading", but they aren't customers for most people who buy new-company data. That's why there's a classification layer as well as a score.

Two shifts in the base rate matter for reading the rest of this page. Since the incorporation fee rose from £12 to £50 on 1 May 2024, fewer companies have been formed each month: 15% fewer in May to August 2024 than in the same months of 2023, and about a quarter fewer than in the four months before the change, which included the April rush. The ones that are formed do better: 38% of companies formed in the four months before the change were dissolved within eighteen months, against 32% of those formed in the four months after, and the share that go on to trade rose from about a third to about two in five. What the fee rise did has the detail and the caveats. The held-out results below were tested on companies formed in early 2024, before the change, so the band percentages are a slightly conservative guide for companies formed since; the bands are being recalibrated on the full cohort. Second, the share of new companies that go on to trade varies from about a quarter to about half between councils, and most of that is not the sector mix: within every major sector, companies formed in the most deprived tenth of England are about 40% less likely to reach trading than those formed in the least deprived tenth (a third less likely once the sector mix is adjusted for). The public council pages show each area's own figures.

Survival, and the official statistics

The Office for National Statistics publishes the official figures on business births, deaths and survival each November (Business Demography, UK; the 2024 edition was released on 20 November 2025). Our cohort should agree with it wherever the two measure the same thing, and it does. But they don't measure the same thing, and the difference is worth stating plainly because it explains why our survival figures are both higher and lower than the official ones.

ONS counts businesses: enterprises that registered for VAT or PAYE, which for most means a business that reached about £90,000 of turnover or took on an employee. We count companies: everything incorporated at Companies House. Roughly 800,000 companies are formed a year and ONS records about 317,000 business births, so around six in ten incorporations never become a business on the official definition. That's the same third-to-two-fifths that our "trading" figure measures from the other direction. The two survival curves cross for exactly that reason. Ours starts higher because a company that will never trade is still registered at twelve months; it ends lower because that company is struck off in year two and we count it. The ONS cohort has already passed the hurdle ours measures, so it loses less early, but when one of its businesses dies it was a real closure. Neither figure is wrong; they answer different questions, and anyone comparing them should say which one they are asking.

Where the two do measure the same geography they agree closely, which is the check that matters for the council figures on this site:

Where the council rankings differ it is for known reasons: a handful of ONS council figures are distorted by unusual registrations (the ONS bulletin itself flags a transport-and-storage anomaly), and London does worse on our measure than on ONS's because its own-address formations include companies that never trade and never appear in the ONS count at all.

How the score is built

The score uses only information that exists on the day a company is formed or arrives in its first weeks, all of it from free public sources: the Companies House register (company type, SIC codes, the registered address and what else is registered there, how many directors were appointed at incorporation and how long the most experienced of them has been a UK director, whether any director has a string of dissolved companies behind them, whether a person with significant control was notified), a curated table of the 500 largest formation addresses with the historical outcomes of companies formed there, whether the company's name resolves as a .co.uk, .com or .uk domain, and whether it appears on the ICO's register of data-protection fee payers. The model is a blend of a logistic regression on 18 curated features and a small gradient-boosted model, trained on 20,000 companies from the 2022 and 2023 cohorts and tested on 10,000 from 2024: a genuine time-based test across a cohort shift, not a random split.

What is deliberately excluded

The two strongest single predictors in the data are the age of the youngest director and whether any director is UK-resident. They aren't used. Age is a protected characteristic and residency and nationality are proxies for one; a commercial lead product that marked companies down for them would be unfair and a legal risk. Leaving them out costs about 0.01 of AUC, measured once on the 2022 to 2024 training cohort. That figure is now a fixed historical result rather than something we re-check each year, because on 20 September 2026 we stopped collecting age, nationality and residency at all and deleted what we held: a field we have decided never to use is a field worth not having. Director names, dates of birth, nationalities and home addresses are never collected, never used as features and never shown to customers.

How well it works

AUC is the standard measure of a ranker: 0.5 is a coin toss, 1.0 is perfect. The columns on the right are what a buyer actually sees (how real the top of the list is, and how empty the bottom is) against a base rate of 37.5% real in the test cohort.

The ICO row is the best number on this page and the least useful one in practice. Only about 19% of companies that become real ever appear on the ICO register at all; the median registration comes 247 days after incorporation, and fewer than 2% are on it within 90 days. A company in this week's list is therefore ranked by the row above it (AUC 0.716), and that's the model the bands below are built from.

The bands customers see are quintiles of the score among companies formed since the fee rise: a band-A company is in the top fifth of post-fee formations. The share of each band that went on to trade, measured on the September 2024 cohort with each half scored by a model that never saw it:

Recalibrated 16 September 2026. The earlier bands (A 60%, E 9%) were measured on companies formed in March 2024, before the fee rise. Since then the base rate has risen from about a third to about two in five and the spread between bands has narrowed: the day-zero signals (name, address, directors, whether a domain resolves) separate post-fee companies less sharply, because the companies they used to catch are no longer being formed. The model is now trained with the post-fee cohort weighted three times and a fee-era indicator; the ranking is 98% correlated with the previous version and the change is in the calibration. A freshly incorporated week shows fewer band-A companies than a fifth, because the domain and ICO signals that lift a score arrive over the first year. One rule overrides the model: a company whose registered office is the Companies House default address is placed in band E whatever the score.

Leaving one cohort month out at a time and testing on it gives AUCs of 0.73 to 0.79 for every pre-fee month and 0.72 for the post-fee month: no seasonal fragility, one genuine regime change. Companies House filings in the first six months add almost nothing (0.716 at day 0 and 0.727 at day 180), because three-quarters of new companies file nothing at all in that time; the score sharpens over the first year as external signals arrive, not from the register. The production scorer, which also uses the ICO signal and is refit on all 30,000, shows in-sample bands running from 7% (E) to 70% (A); the held-out figures above are the ones to quote.

What this means in plain terms. The score is a good ranker and a poor oracle. A band-A company is about 1.7 times as likely to trade as a random incorporation, about two in three against about two in five, and a band-E company is about a quarter as likely; the bottom third of a raw feed can be dropped at a small cost in missed businesses. It isn't a confident yes-or-no label for any individual company, and it is presented as a band, not a verdict.

How long each group lasted, on companies the model had never seen

The table above shows how often each group went on to trade. This one shows how long they lasted. The 30,000 companies the model is built from were ranked out of sample (in five groups, each ranked by a model built from the other four), grouped A to E at today's cut-off points, and followed to their real closing dates in our full record of companies, including the ones that closed. Survival figures use the Kaplan-Meier method, with 95% confidence ranges from resampling. We check them again every month. This edition was measured in October 2026. Companies formed in 2022 and 2023, followed for three years:

The ranges in brackets show how sure each figure is (a 95% confidence range). Months registered is the average number of the first 36 months a company in the group stayed registered. Months trading counts the same months, but only for companies that went on to trade.

The research sets out the method in full, with the two-year check on companies formed since the 2024 fee rise, and what the figures do not show.

The signals, on 30,000 companies

Two things stand out. Companies formed through an accountant or a mid-sized agent are nearly twice as likely to become real as companies registered at a home address, and that same signal is how we flag "already has an accountant". Director experience matters in both directions. Vague or generic SIC codes, which a third of new companies carry, tell you almost nothing about whether the company is real; they cost you what it does, not whether it exists.

The ICO figure needs a caveat: the ICO publishes only current fee payers, so companies that registered and later dissolved have vanished from it, and the 88% is inflated. Salenovo has downloaded the register daily since 9 September 2026 and dates every registration it first sees, which removes that bias for everything from then on.

Classification: what the company is

Alongside the score, each company is classified as an operating business, a property vehicle, a holding or investment company, a dormant or shelf company, or a non-profit, and given a sector, from its name, SIC codes, formation channel and director history. Two independent classifiers are run (a rule set and a language model) and compared: they agree on entity type for 96% of companies and on sector for 85%. Where they disagree the company is marked for review rather than given either answer.

Verified websites and contacts

For every company with a distinctive name we check whether the name resolves as a .co.uk, .com or .uk domain. Resolving sites are fetched (homepage, contact page and a terms or privacy page, respecting robots.txt) and judged against the company: most name-matching domains turn out to be parked pages, placeholders, or an older business with a similar name, and those are excluded. A site is marked confirmed when the company's own registered number appears on it; verified when a model judges it the same business from name, location and content, with a conservative rule that a low-confidence or different-host match is only "possible" and isn't shown. Only generic mailboxes (info@, hello@, enquiries@ and the like) shown on the company's own site are kept; a mailbox's domain is checked for a mail exchanger; personal mailboxes are never kept. In the Essex pilot, 498 of 6,938 companies formed since June 2026 (7.2%) had a verified live site when checked; the share rises as companies age, so every company is re-checked at 7, 14, 28 and 56 days, and again whenever a new lead to its website turns up (the name in other shapes, the trading names and listings we hold, and for the companies most likely to trade, a web search). Each row carries the page it was verified from and the date.

Registered food premises

Every food business must register with its local authority before trading, and the Food Standards Agency publishes those registrations daily with name, address, business type and inspection status. We match them to new companies by name and postcode at three levels of strictness: same postcode with agreeing names; same postcode district with identical or very close names; same postcode area only for distinctive names unique in that area. A list of generic words (café, kitchen, spice, express…) stops "The Coffee Shop" matching anything beyond its own postcode. A match is shown with the premises address, the FSA's rating or "awaiting inspection", the rating date and the date of the FSA extract, as the FSA's terms require.

Geography and the public pages

Postcodes are mapped to local authorities, regions and travel-to-work areas with the ONS Postcode Directory. The expected-to-trade figure on the public Formation Index pages isn't the paid score: it is a transparent estimate in which each company takes the historical trading rate of its mass-formation address (an address with 100 or more registrations in a single calendar year, 287 of them from 2022 to 2024; the Companies House default address is never counted as one; what these addresses are and why companies use them) or otherwise of its SIC section, summed over the area. The rates are measured on companies formed in the months after the May 2024 fee rise, since every company on these pages was formed after it; an address's own rate is blended with its 2022 to 2024 history where its post-fee count is small. On a typical week this puts about 41% of new companies in the expected-to-trade figure, up from 36% under the pre-fee rates used until 16 September 2026. Year-earlier comparisons on those pages start from the only thing the current register can show (companies incorporated then and still on it) and correct it: each week's surviving count is divided by the share of companies the outcome cohort says is still registered at that age (96% at a year, about 70% at eighteen months, when the confirmation-statement strike-offs have gone through), with a separate curve for companies formed since the May 2024 fee rise and a multiplier for each area's own death rate. Reconstructing 2024's monthly formation counts from the September 2026 register this way lands within about 2% of the true figures. Until 16 September 2026 the raw surviving count was shown with a label, which understated year-earlier formations by a third to a half; the raw count is still published in each area's JSON as last_year_on_register. From June 2027 every area also has a clean 52-week baseline from our own daily collection, which has recorded every UK incorporation since 1 June 2026.

Running accuracy

The tests above are on companies formed from 2022 to 2024 whose outcomes are known. The real test is what happens to the companies we score from now on, and that takes time to know: strike-offs become visible from about 15 months, first accounts fall due at about 21 months. Each scored month will be reported here as it matures, as the share of companies in each band that has traded, dissolved or stayed dormant.

Formations by trade

The figures in What Britain starts now are counts from the Companies House advanced search, read on 10 October 2026: for each of the 731 SIC codes, the companies registering that code by year of incorporation (2019 to 2025, and January to September of 2025 and 2026), the companies formed in 2022 that were dissolved by 31 December 2025, and those formed in 2024 that were dissolved by 30 June 2026. The search takes one code per query; it doesn't support asking for companies with any of several codes, so a company with two codes is counted under both, and the trade totals add up to more than the number of companies.

Codes per company. Founders choose more codes than they did: 1.30 per company in 2019 and 1.47 in 2025 (896,689 and 1,220,897 code assignments against 687,219 and 829,933 companies). Code assignments grew 36% while companies grew 21%, so 36% is the baseline every trade's growth is read against.

Survival. "Gone" is the share of a year's companies that had dissolved by the date given. For the 2022 intake that's three to four years of exposure: four for a company formed in January 2022, three for one formed in December. Every trade's intake has much the same mix of months, so the ranking of trades is fair; the level is a blend. Most dissolutions come 18 to 30 months after formation, when the first confirmation statement and accounts fall due and the registrar strikes off companies that don't file.

Geography. For twelve high street trades every company formed in 2019 and in 2025 was downloaded and placed by the postcode of its registered office with postcodes.io, which places 97 to 99% of them. An address with ten or more companies of the same trade among the 2019 and 2025 formations counts as nobody's own (formation agents and accountants) and is left out, which also removes the Companies House default address. Regional percentages are quoted only where the 2019 base is 50 or more. The registered office is often the owner's home or accountant, not where the business trades.

Corrections. In the first pass 59 code and period cells were misread: the search says "No results found" or "1 result" rather than a number of results, and the reader took a figure from elsewhere on the page. Each was asked again and is 0 or 1; none is a code the piece quotes. The geography was first built with a reader that mishandled companies carrying more than one code, which placed too few of them; it was rebuilt, and the published figures are the rebuilt ones.

Against the official series. The search index runs about 1% above Companies House's official incorporation statistics in every period checked (687,219 against 678,419 in 2019, 829,933 against 821,780 in 2025), so growth rates are unaffected. The official figure for the first half of 2026 is down 5.7% on the first half of 2025. The index's figure for January to September, down 7.4%, is provisional until the official July to September release, due in late October 2026.

Independent check. Before publication every figure was re-derived from the raw counts in an independent check, with spot-checks against the live register and the company totals compared with the official series. Everything it asked to be fixed was fixed before the piece went out. The check itself is a text file you can download.

Pubs and pub companies

The figures in the pubs piece come from three sources, each counting something different, never added together.

Companies. Every company carrying SIC 56302 (public houses and bars) in the Companies House search index, 95,874 on 9 October 2026, 39,642 of them active, downloaded by monthly slices so no query hit the 5,000-row cap. Licensed restaurants (56101) are the comparator, because they share the trade's rhythm and most of its policy exposure; licensed clubs (56301) are too few to read. The two trades overlap: 9,065 companies carry both codes (9.5% of pub companies), and 12.5% of the pub insolvency entries from 2019 are companies in the restaurant file too, so the comparator isn't independent and any gap between the trades is narrower than it would be. Formations and survival are comparable from 2016: before 30 June 2016 a company gave no trade code when it was formed, so one that closed before its first annual return never got one and isn't in a search by code, and earlier years are undercounted. The established-company series from 2019 aren't affected, because those companies all lived past their first return. 2020 to 2022 are the registrar's pause on striking off and the catch-up after. "Established" means five years old or more on 1 January; rates are per 100 alive on 1 January, and only companies in that stock are counted; every 2026 count is cut at 9 October and annualised by days.

How companies left. From 9,568 filing histories of established pub companies dissolved since 2019: the registrar's label on the final notice and the forms before it, sorted into a directors' application to strike off, a compulsory strike-off by the registrar, an insolvency (creditors' voluntary liquidation, compulsory winding up, administration or a voluntary arrangement) and a solvent liquidation, the final notice deciding first. A capital reduction or a receiver over property isn't an insolvency, and a liquidation with a declaration of solvency anywhere in its history is a solvent one unless a statement of affairs shows it became a creditors' liquidation. Insolvency entries are dated by the first insolvency filing at Companies House, usually within days of the resolution or order; read together, the companies dissolved since 2019 and those in a process on 9 October 2026 catch every entry from 2019 on. Earlier years can't be counted in full from these records, so 2019 is the one year before the pandemic compared with, and a company that went into administration or a voluntary arrangement and came out still trading wasn't read. A first version of this series counted solvent liquidations and capital reductions as going bust; the independent check found it, and these are the corrected figures. Every licensed restaurant company was read for the entries series; the restaurants' split by manner and year of dissolution is from a sample and isn't published.

Why every address. An earlier pass counted only companies at their own address, leaving out addresses shared by ten or more pub companies to drop group head offices and agents. But a company in liquidation takes its liquidator's address, so most of those crowded addresses were insolvency practitioners' and the filter removed most insolvencies. The published series count every address. For the same reason no regional or council figure for insolvencies is given: the address a company had before its liquidator's isn't held.

The food register. The Wayback Machine's spring captures of each council's food hygiene open-data file for 2013, 2014, 2016, 2018, 2019, 2022, 2025 and 2026. 2015, 2017 and 2020 have too few captures, and 2021, 2023 and 2024 failed systematically when the Food Standards Agency moved the files, so they're left out rather than estimated. A council is compared only with itself, between two springs it was captured in, and pubs are placed in today's councils by postcode. Closed for good, changed hands and new site are as the piece defines them. Councils whose registered pubs moved by more than half between the two springs are left out as recodings (42 for 2013 to 2019, 9 for 2019 to 2026); both versions are in the CSV. Nightclubs and bars share the register's type and ONS's code, so the two compare like with like. Council by council the register and ONS disagree (the correlation of their 2019 to 2026 changes is close to zero), so the register gives the national closure share and the regional split, and ONS gives every count by place.

Jobs. ONS Business Register and Employment Survey (Nomis dataset NM_189_1), Great Britain, employees and employment for SIC 56302 (public houses and bars), 11050 (beer), 11060 (malt) and 46342 (wholesale of wine, beer and spirits), 2015 to 2024, rounded by ONS, read on 15 October 2026. Employees are people on a payroll and employment adds working owners. It counts jobs, not hours or pay; Northern Ireland isn't in it, which is why jobs are given for Great Britain and sites for the UK, and the two are never divided. No industry estimate of jobs "supported" by pubs is used.

ONS. UK Business Counts, local units in SIC 56302, March, rounded to 5 (Nomis dataset NM_141_1): 2010 to 2026 for the UK, nations, regions and councils, with the split by legal status; for 2024 Westminster constituencies ONS publishes March 2025 and 2026 only. A local unit is a site registered for VAT or PAYE, so the smallest sites are missed.

Deprivation and rurality. The Index of Multiple Deprivation 2025 deciles and the 2021 rural-urban classification of the registered office's postcode, England only. Under 2% of pub companies' registered offices couldn't be placed, and a registered office is often an accountant's or the owner's home.

The policy dates are each checked against gov.uk or the House of Commons Library. The pretend-event test runs every month from January 2021 to April 2026 the same way as the real dates, and the table gives where each real date sits among the 64. The independent check is a text file you can download.

Limitations

Registered offices, SIC codes and the status of directors' other companies in the training data are today's values, not day-zero values, which slightly flatters the backtest; the daily collection fixes this for everything incorporated from June 2026. "Real" is defined from the register, so a micro-entity property company counts as real and a thriving sole trader who dissolved the company to go self-employed does not. Officer histories were capped at three directors and fifty appointments each. The most recent test cohort still has about 5% of companies whose first accounts aren't yet due, counted as not real. Four days in mid-June 2026 returned no incorporations from Companies House's advanced search, an upstream gap, and are marked on the charts. AUCs on samples of this size carry roughly ±0.02 of uncertainty. Council-level survival figures exclude the roughly 5% of companies whose registered-office postcode no longer resolves (nearly all of them companies the registrar moved to its default address before striking off), so they understate death rates by about 1.4 points nationally and by an unknown amount in the councils where those companies were originally registered.

Sources and licences

Companies House public register: bulk data, REST API and advanced search (Open Government Licence v3). GOV.UK official incorporation statistics. ONS Postcode Directory and ONS Business Demography, UK: 2024 (both OGL). Information Commissioner's Office register of fee payers (OGL). Food Standards Agency Food Hygiene Rating Scheme open data (OGL; ratings shown with their date). Companies' own public websites, fetched in accordance with their robots.txt with an identified user-agent. No data is bought from brokers and no personal data beyond the public register is collected. Questions about the method, or corrections: hello@salenovo.com.