Bankruptcy probability (12 months)
A statistical model estimating, per company, how likely a bankruptcy opening is within the next 12 months, trained on our own corpus of Belgian bankruptcies, with a real backtest. Everything is below: the factors, the numbers and what the model can NOT do.
What it is
A logistic regression, deliberately not a black box, trained on Belgian legal entities with a bankruptcy opening in the Belgian State Gazette between 2018 and 2026, compared against a multiple of comparable active companies from the same period that did NOT fail. The heart of the approach is time discipline: for a company that failed at date T, the model only uses data that was public at least 12 months before T. So the model genuinely predicts forward; it never peeks at the future.
Training set: 4,852 bankruptcies (2018-2023) against 27,208 controls, nationwide, Flanders, Wallonia and Brussels. Validation was out-of-time on 11,853 bankruptcies from 2024-2026 the model never saw, and the absolute probabilities are calibrated against Statbel's official bankruptcy statistics.
The factors
All factors come from public sources (CBE + Belgian State Gazette). This is the full list, there are no hidden inputs:
- Age, age since incorporation; failure risk peaks between years 2 and 8, then declines.
- Legal form, BV, NV, VOF/CommV, CV or VZW; a non-profit, for instance, fails far less often than a BV.
- Sector, the historical failure rate of the NACE sector (hospitality and construction high, professional services low), computed on the training years.
- Judicial reorganisation, the strongest red flag: an earlier judicial-reorganisation procedure.
- Interim administrator, an appointed interim administrator or comparable measure.
- Ex-officio strike-offs, strike-off of address, UBO registration or for non-filing.
- Recent signals, the more of those warnings in the last 24 months, the heavier they weigh.
- Establishments & registration, number of establishment units and whether a NACE activity is registered.
- Region, the structural difference in failure rate between Flanders, Wallonia and Brussels (registered-office zip), corrected for differences in our publication coverage per region.
- Annual accounts, when filed annual accounts are available: solvency (equity / total assets), profitability, liquidity, cash position, size, staleness of the figures and the trend versus the previous year. NO accounts available expressly means NO adjustment either way: the absence of figures in our mirror says nothing about the company.
On the dossier we show, per company, the three factors that move the score most, with their direction (+ raises, − lowers).
The backtest
Trained on bankruptcies up to and including 2023, validated on 2024-2026, data the model never saw. The AUC on that validation is 0.79 (0.5 = coin flip, 1.0 = perfect). Concretely: give the model a random company that failed within 12 months and a random survivor, and it ranks the failure higher in 79% of cases.
Because our publication coverage differs per region, we also validate per region separately, AUC: Flanders 0.79, Wallonia 0.76, Brussels 0.81.
Calibration: we split the validation companies into ten equal groups by predicted probability and compare with what actually happened:
| Group | Predicted | Realized |
|---|---|---|
| 1 | 0.03% | 0.05% |
| 2 | 0.08% | 0.07% |
| 3 | 0.16% | 0.21% |
| 4 | 0.25% | 0.24% |
| 5 | 0.39% | 0.40% |
| 6 | 0.57% | 0.52% |
| 7 | 0.75% | 0.67% |
| 8 | 1.02% | 1.04% |
| 9 | 1.68% | 1.79% |
| 10 | 3.42% | 3.36% |
In the middle groups the model slightly overestimates risk (the cautious side); in the top and bottom groups, where the bands matter, it tracks reality closely.
The bands
- Very low, < 0,5%
- Low, 0,5% - 1,5%
- Moderate, 1,5% - 4%
- Elevated, 4% - 10%
- High, ≥ 10%
For reference: according to official Statbel figures, on average about 1.0% of active Belgian legal entities go bankrupt per year. We show one decimal and cap the display at "≥ 25%", false precision in the tail helps no one.
Honest limitations
- Annual-accounts figures only when available, our mirror of the NBB annual accounts is still growing: for part of the companies no figures are processed yet. For those, the model relies on the remaining factors, with no adjustment either way for the missing figures. The backtest is therefore validated separately on companies with and without available accounts.
- Publication coverage per region, our linking of State Gazette publications to enterprise numbers is historically less complete for French-language courts. The model is corrected using official Statbel statistics per region, and the backtest is validated per region separately. Still, for Walloon and Brussels companies recent warning signals may be missed slightly more often, a missed signal lowers the score, it never wrongly raises it.
- Young companies, for companies younger than 12 months we show no score: there simply is no history.
- The model can't see private information, a lost key customer, a bank refusal, a shareholder dispute, none of that is public before it ends up in a procedure. A low score is no guarantee and a high score is no verdict: it is a statistical estimate based on public sources, not credit advice nor a finding of fact.
Questions or a correction about your company? Email hallo@checked.be. See also the general methodology.