Beta

Compdex is new: mistakes will happen. Tell us.

Report a problemSuggest a feature or give feedback
Skip to main content
How we measure

Methodology

How every model works: what it reads, what it states, and how we know how good it is.

The principles

Every figure on a company page comes from a stored source record with a link and a date, or from a documented model over such records. Anything modelled is labelled as an estimate, with a range, a confidence and the method. Where there is nothing to show, the page says so rather than filling the gap: a figure not stated is kept apart from a figure of zero.

Written summaries are checked number by number against the facts they were written from, and every page is complete with them switched off. Anyone can report an error on any page, and a person checks every report against the sources.

Profit estimate

Estimate
What it reads
Two consecutive balance sheets: retained earnings, dividends where the accounts disclose them, the corporation tax owed at the year end, cash, net current assets and net assets. It runs only for a company that files no profit and loss account; a filed profit always replaces it.
What it states
Profit before tax for the year, as a likely range and a wider one. It starts from the floor: the growth in retained earnings plus any dividends is the profit kept or paid out after tax, grossed up for corporation tax at the rates in force. The tax owed at the year end sets the figure where it implies more, unless it implies more than 3 times net assets and is more than 1.5 times the tax owed a year earlier. The top of the range allows for dividends nobody disclosed (50% of what the company could have paid out, and 5% of its net assets), is widened where similar companies filed profits outside it, and is kept below the turnover estimate where there is one. Try the floor rule.
How it is tested
The filed profit of 92,477 held-out companies fell inside the wider range 82.7% of the time, and inside the likely range 53% of the time (model profit-1.10.0). The confidence each page states is that measured share, not a feeling. See the accuracy record.

Turnover estimate

Estimate
What it reads
What a small company does file: staff numbers, debtors, cash, creditors, the breakdown lines in the notes, and how each moved from the year before. A missing line is treated as missing, never as zero.
What it states
Turnover as a likely band, the middle half of what similar companies filed, and a wide band holding nearly all of it. Five learned models, one for each point of the spread, were fitted on the Companies House accounts archive, and more staff, debtors, cash or creditors never lower the figure. Each estimate lists what moved it. Where the learned model cannot run, sector ratios stand in. Filed turnover always replaces it.
How it is tested
On 7,107 held-out company-years that did file turnover, the figure fell inside the wide band 80% of the time and inside the likely band 52% (model turnover-2.7.0). Confidence is measured per size band, so a small company is not credited with a large one’s accuracy. See the accuracy record.

Estimated profit and loss account

Estimate
What it reads
The turnover and profit estimates above, and the median margins of the company’s trade from filed accounts.
What it states
The estimates laid out as a profit and loss account, each line labelled as estimated, taken from the sector, or derived by arithmetic from the lines above it. Gross profit is the trade’s median margin applied to the estimated turnover, with the trade’s spread beside it. A line that cannot be connected honestly to the others is left out with a note rather than bridged by a number standing for nothing. A filed profit and loss account replaces the whole thing.
How it is tested
Not tested as a whole. Its estimated lines carry the turnover and profit estimates’ own measured confidence; its sector lines claim none.

Tax, dividends and what the owners took out

Estimate
What it reads
The tax charge in a filed profit and loss account, dividends in the accounts, the corporation tax owed at the year end, and directors’ pay and loans from the notes.
What it states
For each year: corporation tax, dividends, the payout ratio, the effective tax rate, the share of profit kept, directors’ pay and the movement in loans to directors, each marked filed or estimated. Tax not filed is implied by the year-end creditor. Dividends are estimated only where independent evidence supports it; otherwise the page shows a dash. A company is said to have paid no corporation tax only from filed figures. Whether a year-end creditor can be read as one year’s tax is decided by one rule, shared with the profit estimate.
How it is tested
The filed corporation tax fell inside the estimate’s range 75% of the time on 21,355 companies, and the filed dividend inside its range 70% of the time on 8,207 (model distributions-1.2.0). See the accuracy record.

Compdex Health Score

Rules
What it reads
The filed accounts, the Companies House register, the filing history against its statutory deadlines, officer appointments, insolvency notices in The Gazette, and, where we hold them, payment practices reports, employment tribunal decisions, the disqualified directors register, the FCA and CQC registers, the company’s proved website and its customer reviews. A source not read for a company is listed as not checked.
What it states
A score from 0 to 100 and a band, from three parts each scored 0 to 100. Financial strength, 50% of the score, reads the accounts: net assets and their movement over one and two years, liquidity, cash against what falls due within a year, the debt ratio, a filed profit or loss and retained earnings, each only where the accounts state the lines it needs; a line not stated is never read as nil. Stability reads the register: status, insolvency and strike-off, years on the register, the board and office moves. Conduct reads how late the latest accounts were and whether earlier ones were late too (a few days once is close to nothing; months late again and again is serious), the confirmation statements, and how the company pays its suppliers where it reports. A part with nothing read drops out and the others carry the score. Each factor is listed with the points it moved the score by and its evidence. A company the register shows closed or in insolvency, with a recent insolvency notice in The Gazette, or with strike-off action in progress is capped however healthy the rest looks. Beside the score is the list of what it has not checked, such as county court judgments; an unchecked source is never shown as clear. Filing deadlines are worked out exactly or not at all, by the same rules as the deadlines tool.
How it is tested
Fitted and tested on the Companies House accounts archive: the accounts measures, the years on the register and filing punctuality are weighted by how well they told apart the companies that later entered insolvency or were proposed for striking off before their next accounts. On 38,367 later filings held out of the fit it ranks them with an AUC of 0.68, against 0.65 for the earlier version on the same filings; 0.6% of band A and 3.2% of band D failed, against 1.4% overall. The other rows are judgement, recorded as such.

TestedThe Compdex Health Score weighs the accounts, the years on the register and filing punctuality by how well they told apart the companies that later failed, and on later filings it ranks them better than our earlier score did. Its other rows are judgement. Read it as a summary of the evidence, not a prediction.

Credit score and suggested limit

Rules
What it reads
The last two filed years, the current estimates, the filing history and its deadlines, registered charges, insolvency notices in The Gazette, payment practices reports, the other companies its directors served where we hold their records, and the measured failure rate of its trade.
What it states
A score from 0 to 100, starting at 50, with every factor scored from a weights table the page shows, banded A to D. A suggested limit on thirty-day terms: the smallest of 10% of turnover, 25% of net assets and 50% of the working capital headroom, scaled by the band. And what lenders’ usual rules make of the company, product by product. County court judgments are not read and are listed as unchecked. The band is public; the score, the limit and the lender views come with Pro. How the limit is worked out.
How it is tested
On 41,870 held-out company-years, the score’s factors with their measured weights ranked the companies that later closed or became insolvent above the rest with an AUC of 0.73, against 0.67 for the score’s own points (model credit-1.4.0). The bar for publishing a failure probability is 0.75, so none is shown yet. The limit and the lender views are not yet measured against later defaults. See the accuracy record.

ProvisionalThe score, the limit and the lender views follow the weights shown here. Tested against which companies later closed or became insolvent, the score ranks them better than chance but not yet well enough to publish a probability, and the limit and lender views are untested. Not a credit reference.

Twelve-month forecasts

Rules
What it reads
The register, the filing history and the latest balance sheets.
What it states
Whether the next accounts will be late and whether the company will borrow more, each as likely, possible or unlikely with the rules that put it there, and headcount as a range from the filed series. Never a percentage: without a fitted model there is no honest one. A forecast whose inputs are not read, such as county court judgments, says it is not available and why.
How it is tested
Not yet measured against what happened next.

ProvisionalThese bands follow stated rules over the register and the latest balance sheet. They are not probabilities, and they have not yet been tested against what happened next.

Valuation

Estimate
What it reads
The latest filed operating profit, or profit before tax with interest added back, or the likely band of the profit estimate; cash, bank loans and long-term creditors; and who runs and controls the company.
What it states
An enterprise value and an equity value, each a range. The earnings are multiplied by the trade’s listed multiple scaled to 30–50% for a private company; cash is added and debt taken off, and 20% comes off where one person is both the only director and the only person with significant control. Every step is shown, and so are the succession signals a buyer asks about first. Model valuation-1.5.0. Try it.
How it is tested
No confidence figure is given, because nothing measures a private company’s price until it sells.

ProvisionalAn earnings multiple over our own estimates, with every step shown. It has not been tested against sale prices, and it is not a valuation for any transaction.

Change of control

Rules
What it reads
The company’s register events over 90 days: people with significant control ceasing and arriving, directors leaving and joining, new charges or every charge satisfied, the registered office moving, shares allotted, a change of name.
What it states
A pattern in words, likely or possible, with fixed points for each rule and every event that raised it. Never a probability. The likely acquirer, inferred from register links, is shown only to signed-in readers on a paid plan.
How it is tested
25,279 companies are flagged, 11,509 of them as likely (model control-change-1.0.0); none has yet been checked against announced deals. See the accuracy record.

ProvisionalA pattern read from the register by fixed rules, not a finding that control has changed. It has not yet been tested against announced deals.

Peers

Arithmetic
What it reads
The Companies House bulk register and the latest accounts of every active company that files them.
What it states
Where the company sits on four measures a balance sheet supports: growth in net assets, liquidity, growth in headcount, and margin where profit and turnover are both filed. Against three sets of peers: the same trade in the same postcode area, the same trade across the UK, and every trade in the same town. The trade is the five-digit SIC code where that gives enough peers, otherwise its division. Fewer than 20 peers stating a measure and no percentile is given; the peer count is always shown.
How it is tested
Arithmetic over filed figures, not an estimate, so there is nothing to measure. The sector medians are medians of filed accounts and state their count.

True sector

Classification
What it reads
The company’s name, its SIC code descriptions, its proved website address and its brief.
What it states
Which leaf of our own two-level taxonomy the company actually belongs to, with the confidence the model gave and the evidence it quoted, copied from the inputs and checked. It is shown beside the SIC code, never instead of it, and not yet used for peers or filters.
How it is tested
On 10 hand-labelled companies, it chose the right leaf for 100% of those it classified and the right group for 100% (prompt v1). See the accuracy record.

Website matching

Proof
What it reads
A search for the company’s name and town, with directories, filing services and social networks struck out, and then each remaining page itself.
What it states
The company’s own website, only when the page proves it: the registered number in the page, which the Companies Act requires a company to show on its websites, or the registered name on a domain that also looks like the name, a weaker proof and labelled as such. Nothing unproved is shown, and no model guesses a domain. Reviews and online visibility are read only for a proved website.
How it is tested
A proof, not a statistical model, so there is no hit rate to measure. Each match is kept with the evidence that proved it.

Where the data comes from is on the sources page, and the free tools apply several of these rules to figures you type.