The latest filed accounts show an operating profit. Revenue is growing; nothing flags a problem. Three years later, the commercial court opens insolvency proceedings.
This is no isolated case. Among the 12,925 corporate insolvencies observed here, nearly one in four involved a company that was profitable at the operating level in its last published accounts. The income statement did not flag them. The balance sheet did.
More precisely, the gap between cash paid out and cash collected. The question here – the one every investment committee faces with a private target – is twofold. Can corporate default risk be detected in time? And can machine learning, applied to filed accounts and combined with explainability methods, bring out signals that conventional financial analysis misses?
The answers are quantified below, together with what they change in practice at each stage of a deal, from origination to negotiation.
Private companies leave a single trail
A listed company sends signals. The share price falls, analysts cut their estimates, the rating slips – all of it visible months before anything shows up in the accounts. In a comparable model built on US listed companies, this market verdict ranks among the three strongest signals.
A French private company sends none. No share price, no rating, no analyst. The smallest may even elect to keep their accounts confidential, and many do. Only one trail is common to every target in a pipeline: the annual accounts filed with the commercial court registry.
That is a constraint, and it is also what makes the exercise worthwhile. If the signal exists, it is in the accounts – or nowhere.
Data, scope and method
The data: all annual accounts filed with the French national business register, i.e. 4.85 million sets of accounts from nearly 1.15 million unlisted French companies, matched against 648,623 insolvency notices published between 2012 and 2025. Only openings of judicial reorganisation (redressement judiciaire) and judicial liquidation are retained.
The scope: we deliberately narrowed the population. In the raw data, micro-enterprises dominate the French sample of failed companies. We therefore retained, by convention, only companies with total assets of at least €1 million: below the €2 million ceiling used to define a micro-enterprise, but enough to move away from that population and closer to the universe that funds and trade buyers actually target. After this filter, the analysis covers 1,895,118 sets of accounts filed by 388,139 companies, including 12,925 insolvencies, or 3.3% of the retained population.
This three-year rate of 3.3% is measured on the entire retained population, with no weighting or matching. It is consistent in order of magnitude with official insolvency statistics, allowing for differences in coverage and period – the first sanity check any such figure should pass.
The question: not “will this company fail one day?”, which helps no one, but “will insolvency proceedings be opened against it within three years?”, a horizon consistent with a typical investment thesis. Throughout this article, default refers specifically to the opening of judicial reorganisation or liquidation proceedings, dated by the opening judgment; it does not cover every missed payment. Safeguard proceedings (sauvegarde) are excluded: they are opened before the company is unable to pay its debts as they fall due, while it is still meeting them, and counting them would classify as failures companies that had the foresight to act early. Judgments approving a reorganisation plan are excluded as well, for a different reason: they come several months after the opening, and using them would date the failure too late, distorting a measurement anchored on the three years following the last filed accounts.
The period: fourteen financial years, through 2025. The data therefore span the pandemic, the support measures that drove insolvency openings down, and the subsequent catch-up.
The test: the model is trained on financial years up to 2019 and evaluated only on 2022 accounts it has never seen. The test set comprises 283,788 sets of accounts, 5,170 of which precede an insolvency within three years. As an additional check, on companies for which it had never seen a single set of accounts, performance drops by about three points, with no breakdown.
One insolvency in four hits a company with an operating profit
Of the companies that failed, 23.6% reported an operating profit in their last published accounts: 3,047 companies, in a sample where losses remain the most common warning sign.
What sets them apart from the profitable companies that survive does not show in the income statement, but in the cash their business ties up: inventories and trade receivables to be financed, less the credit extended by suppliers. In other words, working capital.
Across 151,886 profitable companies and 93,991 loss-making ones, the pattern is clear. Among profitable companies, the insolvency rate almost doubles as operating working capital increases (×1.9 on the chart): 1.4% when it stays below 10% of total assets, 2.7% to 2.9% above 25%.
For loss-making companies, the chart shows a gap of only ×1.4, but that figure compares the two extreme bands: in between, the rate climbs to 10.3%. Working capital weighs on them too; the loss has simply already raised the flag, and at the same working capital intensity these companies fail two to four times as often as profitable ones. Put differently: working capital is not one indicator among many – it is the tie-breaker when the income statement looks reassuring. Which is precisely the situation of a target that has cleared an investment committee’s first screen.
One caveat: negative working capital is no guarantee of strength. Among profitable companies it goes with a 1.9% insolvency rate, higher than in the 0–10% band. It is the hallmark of retail and restaurants, funded by their suppliers and vulnerable for other reasons. The safest level sits just above zero, not as low as possible.
What the corporate default risk model actually weighs
The finding above emerges from the raw data, without a model. The question is whether a machine learning model, trained with no prior guidance, reaches the same conclusion. It comes close, but the answer is more nuanced than one might expect.
Each model estimate can be broken down into contributions, line item by line item. Aggregated by family across the test set, these contributions produce the ranking above, and the first lesson is that no family dominates: the cash cycle (22.6%) and a company’s position relative to its sector (21.4%) are level, ahead of trajectory (17.7%). Next come financial structure (11.1%), profitability (10.9%), raw sector and size markers (10.8%), and internal cash generation relative to debt service (5.4%).
The white lines tell the story: the two leading families overlap, and either may come first depending on the measurement. Looking for the dominant factor would be a misreading – there are two, of equal weight: the cash tied up in the operating cycle and the gap to sector norms. Trajectory follows immediately, and it is the most stable weight on the chart: 17.7% in every measurement. The lower half of the ranking warrants more caution – only a few points separate financial structure, profitability and the size markers, and their order is not firmly established.
This narrowing is no accident; it follows from the scope set out above. For micro-enterprises, cash explains almost all of the risk, because there is often little else to observe: a thin financial structure, and sector comparisons that carry little information at that scale. Small and mid-sized companies already have a capital structure, a track record and a position in their sector, and these three dimensions come to matter almost as much as the cash cycle. Assessing a target of this size means reading them together.
Sector positioning deserves a closer look: it shares the lead with the cash cycle, whereas its contribution was marginal on the unfiltered population. Its message is easy to state and easy to overlook in practice: what matters is not the level of a ratio, but its distance from the sector norm. Working capital at 20% of total assets would be a red flag in food retail, where it is typically negative; in construction, it is business as usual.
Profitability sits mid-table, at 10.9% on average – the least robust figure on the chart, since it ranges from 6.2% to 16.0% depending on where tax, depreciation and financial items are classified. It does not support the conclusion that profitability is redundant. What does hold, and it comes not from this ranking but from the opening finding, is that profitability fails to discriminate where discrimination is needed: any rough screen will catch a loss-making company; it is the profitable company that has to be assessed – the one for which, by construction, operating profit says nothing.
Construction: a textbook case
The mechanism is visible to the naked eye in a single sector. Specialised construction activities form the largest sector in the sample, and the one with the most distinctive financial profile.
| Construction | All other sectors | |
|---|---|---|
| Working capital / total assets (median) | 13.9% | 6.7% |
| Trade receivables / total assets (median) | 24.5% | 15.1% |
| Three-year insolvency rate | 7.5% | 3.0% |
Working capital relative to total assets twice that of the rest of the economy; trade receivables more than 50% higher. The business model explains it entirely: work is carried out before it is paid for, progress billings are settled late, and retentions hold back up to 5% of the contract value for a year after handover. The contractor finances its client.
The result is among the most telling in the analysis. Among profitable construction companies, 5.2% become insolvent within three years, against 1.8% for profitable companies in all other sectors. Nearly three times as many, although all of them are profitable at the operating level. A positive margin does not make up for a stalled collection cycle.
Where insolvencies concentrate
We set out to rank sectors by their exposure to insolvency. The data provide that ranking, and something more useful: a dividing line. Five divisions of the French activity classification (NAF) reach or exceed 9.8%, three times the 3.3% average. Five others remain below 1.2%, at least three times lower. Between the two groups, the ratio is roughly ten to one. Within the most exposed group, however, the precise order is debatable: a few hundred companies per division are enough to establish that a division is exposed, but not, in this case, to separate the second from the two that follow. What matters, then, is not an activity’s rank but whether it belongs to the group around 10% or the group around 1%.
| Most exposed | Insolvency rate | Most resilient | Insolvency rate |
|---|---|---|---|
| Remediation and decontamination | 14.0% | Veterinary activities | 0.2% |
| Apparel manufacturing | 11.9% | Legal and accounting activities | 0.4% |
| Furniture manufacturing | 11.8% | Energy | 0.4% |
| Investigation and security | 11.3% | Real estate | 1.1% |
| Printing | 9.8% | Other mining and quarrying | 1.1% |
The line does not follow the expected boundary between manufacturing and services. It separates activities whose margins are renegotiated constantly – with every order, every tender – from those whose revenue is protected, by regulation or by an asset. On one side, apparel, furniture, printing and security; on the other, veterinary, legal and accounting services, energy and real estate. Companies in the first group fail ten times as often as those in the second.
Most exposed
Remediation and decontamination. These companies rehabilitate sites: asbestos removal, soil and groundwater decontamination, restoration of former industrial land. A project-based business, sold at fixed prices to public or industrial clients, with regulatory costs to be pre-financed and a final payment received only on acceptance. Fourteen per cent of them faced insolvency proceedings within three years of their last filed accounts: four times the overall average, and the highest rate among the seventy-two divisions retained. The explanation probably lies in the business model rather than the technical nature of the work: few fixed assets, working capital that absorbs cash between contracts, and margins renegotiated at every tender. One clarification: waste collection, sorting and treatment belong to a neighbouring division, made up of much larger operators, and are not included.
Apparel, furniture, printing: 9.8% to 11.9%, three to four times the average. These three divisions share a profile: labour-intensive manufacturing exposed to price competition. Production equipment to depreciate, short runs, highly price-sensitive customers and, against them, imports that set the market price without bearing the same costs. Margins are renegotiated with every order, with neither liquid assets nor pricing power to absorb a shock. These are not badly run sectors; they are sectors whose business model leaves no room for error.
Most resilient
The lowest rates range from 0.2% to 0.4%: veterinary activities, legal and accounting activities, and energy.
Each is measured on several hundred to several thousand companies, with very few insolvencies – not because the samples are too small, but because failure is genuinely rare there: regulated professions, inelastic demand, or activities backed by assets and long-term contracts.
These three rates are too close to be ranked against one another; what is robust is that all three sit well below the 3.3% average.
Just above them, one sector deserves attention, less for its rank than for what it teaches.
Real estate: 1.1%. One company in ninety, against one in eighteen in building construction. The asset ties up substantial capital, but it also serves as collateral: a building can be refinanced, sold or let, whereas a specialised machine tool finds a buyer only at a steep discount. Rental income cushions shocks that an industrial business would absorb in full.
Yet the most instructive contrast is not between the top and the bottom of the ranking, but within a single value chain. Real estate loses one company in ninety; building construction, which produces the assets the former go on to hold, loses five times as many. Same building, same economic cycle, opposite positions: one owns a refinanceable asset and collects rent; the other pre-finances the project and waits to be paid. What protects is not the sector, but the company’s place in the cash collection cycle.
The lesson extends well beyond construction. High working capital, revenue renegotiated at short intervals, services delivered before they are paid for: this combination describes many of the industrial and business-services targets that cross a deal team’s desk. It is not a sector effect but a balance-sheet signature, and it calls for the same caution on a target whose sector appears lightly exposed.
Negative equity: the strongest warning is also the rarest
Taken individually, some threshold crossings separate failing companies from the rest better than the level reached.
Negative equity is found in 27.4% of failing companies, against 8.1% of healthy ones – that is, companies that did not fail within three years: a situation 3.4 times more frequent among the former. Swinging into negative equity from one year to the next is 7.1 times more frequent, the highest ratio in the sample.
This figure needs to be read for what it is. The swing affects only 11.3% of failing companies – one in nine. And when it occurs, it does not seal the company’s fate: roughly one in five goes on to fail within three years, six times the average. It is a strong but rare warning, which explains why it carries little weight in the average ranking while warranting immediate action when it is triggered.
The contrast with the bottom of the chart tells the rest. “Seven or more deteriorating signals” applies to 44.8% of failing companies, almost one in two – but also to 30.4% of healthy ones, a ratio of only 1.5. On one side, a frequent and weakly discriminating signal; on the other, a rare and far more selective one: what truly separates is crossing a threshold, not gradual erosion.
The practical takeaway fits in one sentence: a persistently mediocre company is a business model, not a warning sign. It is the one that was comfortable three years ago and is merely adequate today that justifies the extra week of work.
How well does it work in practice?
There are two ways to answer: one technical, one operational.
The technical answer. Given two companies drawn at random, one that failed and one that did not, the best model assigns the higher risk to the failing one 87 times out of 100.
| Model | Discriminatory power | Insolvencies captured in the riskiest 20% |
|---|---|---|
| Gradient-boosted decision trees | 87.3% | 78% |
| Random forest | 86.7% | 76% |
| Logistic regression | 84.6% | 73% |
With 5,170 insolvencies in the test set alone, the margin of error on this measure is about three tenths of a point. The gap with logistic regression is therefore real; the six tenths separating the top two models remain close to that margin. Above all, the differences are modest: logistic regression, the classic statistical method, comes within three points of the best model. The choice of method does not drive the result; the quality of the data does. Machine learning is therefore not a methodological breakthrough here: it brings a systematic reading of corporate default risk to a universe too large to be reviewed company by company.
The operational answer. Review first the 20% of companies ranked riskiest, and you capture 78% of those that will fail. Within that fifth of the portfolio, 7.1% of companies go on to fail, against 1.8% in a random selection: almost four times as many.
The curve is even more telling at its far end. Among the riskiest 5% of companies, nearly one in six does fail within three years – more than eight times the base rate. For so rare an event, concentrating failures on a short list in this way is exactly what a first screen should deliver when an investment committee has days, not weeks, to decide.
Can these figures go into an investment memo?
With caution, and only if the model has been calibrated for the purpose – a step often skipped. A raw score ranks; it does not quantify. It says that one company is riskier than another, nothing more.
The banner on the chart above states the good news: after recalibration, the order of magnitude holds. The curve tells the rest of the story, which is better disclosed upfront than discovered later: it stays above the diagonal along its entire length. The model therefore underestimates risk in every band, and the gap widens towards the most exposed companies. In the riskiest decile of the test set, the model predicts 7.6% on average; 11.0% of companies actually fail.
The practical consequence is simple. A percentage from this model should be read as a floor for a group of companies, not as a point estimate or a guaranteed minimum for each one: observed outcomes are higher, particularly for the riskiest companies – precisely those that reach the investment committee. The model ranks better than it quantifies, and that is how it should be used: to prioritise a universe and focus attention, not to state a default probability to the decimal point in an investment memo.
One file, opened factor by factor
An average ranking shows what matters in general. An investment committee needs something else: why this company, and on what evidence. A model that answers “risky” without saying why is not wrong; it is unusable, because no committee approves a decision it cannot scrutinise.
The chart reads like a credit memo, not a model output. Each pink bar raises the estimated risk, each blue bar lowers it, and the length of the bar measures its contribution. The bars cover line items, but also ratios, their year-on-year changes and their gaps to sector norms – all derived from the filed accounts. These contributions explain the model’s estimate; they do not prove that a factor caused the failure.
This company illustrates the argument of this article precisely: the model put its risk at 13.8%, four times the overall average, and insolvency proceedings did follow within three years. The size of its balance sheet worked in its favour; outweighing it were cash too thin relative to its liabilities other than borrowings, heavy tax and social-security liabilities, a declining return on assets, and a solvency score and equity ratio well below sector norms. A dashboard built on profitability would have waved it through.
Note that none of the largest contributors to the estimated risk is a level of earnings: two cash-cycle items and a decline in profitability come first, followed by two gaps to sector norms. It is a single-case demonstration of what the family ranking shows on average.
It is also what makes the conclusion testable in the field. If the breakdown rests on an item the data room shows to be exceptional, it can be dismissed on the evidence rather than on instinct. Every bar traces back to the filed accounts, which anyone holding them can verify: nothing rests on the model’s authority.
What this changes at each stage of a deal
Origination and target search : the French private-company universe is vast and opaque – no ratings, no analyst coverage, no share price. Filed accounts are the only information available on every target at once. Using them for a first screen by corporate default risk replaces no judgement; it sets the order in which companies are reviewed. On a long list of eighty names, that alone determines how a deal team spends its weeks. The same screen applies beyond acquisitions: ranking by default risk the counterparties one is about to take exposure to – a key customer, a single-source supplier, a consortium partner – calls for the same reading.
Opportunity analysis : two targets priced at the same EBITDA multiple are not the same asset. The one whose operating cycle ties up a quarter of its balance sheet will have to fund its growth before collecting the proceeds; the other will not. The difference does not show in a multiple, it shows in the balance sheet – and it changes the investment thesis, not just the price.
Buy-side due diligence: a data room rarely holds more than three years of accounts. Filed accounts, by contrast, are public and take minutes to obtain: two extra years are often enough to change the picture, and to check that the trajectory presented matches the one on file. For long-cycle targets, two points merit a written question: the seasonality of working capital, since a year-end at the seasonal low flatters the balance sheet, and the timing and recoverability of retentions.
Valuation, structuring and closing: this is where the analysis turns into money. Structurally high working capital is a permanent funding need, and it belongs in the valuation itself – not in the completion-accounts adjustment, nor in the agreed definition of net debt. Three negotiating points follow: set the normalised working capital target on a multi-year average rather than the latest balance sheet, link any earn-out to cash actually collected rather than revenue invoiced, and tie the warranty and indemnity package to trade receivables. The history of filed accounts provides the evidence for each.
What can still be refined
Coverage. The 3.3% rate applies to companies whose accounts are filed, usable and within the selection criteria. Filings made under a confidentiality election are flagged by the registry and can therefore be counted: measuring their share sector by sector is the natural next step. Companies that file no accounts at all remain, by construction, beyond the reach of an accounts-based model.
Scope. The analysis focuses on companies with total assets of at least one million euros – the universe of a transaction. Micro-enterprises, which make up the bulk of the French business population, warrant a study of their own.
Definition and period. Failure is defined here as the opening of judicial reorganisation or liquidation proceedings; a safeguard procedure brought to a successful conclusion calls for a different reading. The years 2020 and 2021, when insolvencies collapsed (about 27,000 in 2021 according to the Banque de France, against more than 50,000 in 2019), fall within the observation window of part of the training years: weighting them differently is one avenue.
Sectors and horizon. Sector rates are frequencies observed over a given period. Seventy-two divisions have a sufficient number of companies; the remaining fifteen need more observations, as would a balance-sheet threshold raised to two million euros or a horizon other than three years.
What the data room adds. Management quality, governance, litigation, customer concentration, off-balance-sheet commitments: these often decisive factors emerge in due diligence. The model prioritises the files; the data room completes them.
Conclusion
One thread runs through the whole analysis: in the French private-company market, profitability is not enough to decide once companies are large enough for their financial structure and trajectory to be analysed. Nearly one insolvency in four hits a company with an operating profit, and among these companies operating working capital almost doubles the risk. But for small and mid-sized companies alike, the cash cycle does not decide on its own: it shares the lead with the company’s position relative to its sector, with trajectory close behind. These are readings to be combined, not ranked.
Sectors tell the same story from another angle – not which trades suffer insolvencies, but which share the same structure. It is not the most capital-intensive businesses that fail, but those whose margins are renegotiated with no asset to cushion the blow: apparel, furniture, printing, security. At the other end, regulated or asset-backed activities hold firm, with exposure roughly ten times lower – a gap no EBITDA multiple will ever reveal.
For a fund, a head of M&A or an adviser, the implication is direct. The income statement does a poor job of separating profitable companies – precisely those that reach the investment committee. The balance sheet, read against sector benchmarks, does.
Machine learning, combined here with explainability methods, does not replace credit analysis: it ranks corporate default risk and makes every conclusion traceable back to the accounts. That last point is not a detail: it is what separates a tool a committee uses from one it listens to politely before deciding otherwise.
Discussing a live deal
We contribute to a transaction in three ways: making a series of financial years genuinely comparable before reading a trend into it, positioning a target against its sector rather than in the abstract, and turning a balance-sheet weakness into a figure that holds up at the negotiating table.
Ranking a list of targets by corporate default risk, defending a normalised working capital figure challenged in negotiation, investigating a business whose reported profits look reassuring but whose cash profile raises questions: these are matters we work on.
Note on method. The data are split chronologically, calibration is fitted on a period separate from the test, and every step halts if a check fails. Other choices of size threshold, horizon or split would produce different figures, and use under current conditions would require recalibration. None of the above constitutes a production credit model, investment advice or a recommendation.
This study was built iteratively, over several versions, with the assistance of artificial intelligence. Versions were discarded more often because of data flaws than model weaknesses; what is published here has passed every check.
About LJ Advisory
M&A and capital reallocation advisory for mid-market investment funds, energy and industrial companies, and executives involved in cross-border transactions.