Back-test
The numbers behind the methodology. Aggregates only: no company is named or numbered here.
The question
Does each registry signal separate UK companies that later entered liquidation or administration from comparable healthy ones, using only what the register showed 12 months before the event?
Samples
- First sample: 100 failed companies and 100 healthy ones. Used to explore and to set the first rules.
- Validation sample: 299 failed and 299 healthy, drawn independently with a different seed, with no company from the first sample.
- Failed companies: UK limited companies and PLCs whose liquidation (compulsory or creditors’ voluntary) or administration began in 2023 to 2025, drawn at random by incorporation month from 2006 to 2024. Members’ voluntary liquidations are excluded, because they are solvent. Events by year: first sample 2023: 21, 2024: 23, 2025: 56; validation 2023: 63, 2024: 76, 2025: 161.
- Healthy companies: for each failed company one company that is active today, of the same type, sharing its industry section, incorporated within a year either side of it, and evaluated at the same date.
- Median company age at the evaluation date: 8 years (first), 7 years (validation).
Method
- The evaluation date is 12 months before the failed company’s event; its healthy match is evaluated at the same date.
- The production signal code runs on a profile rebuilt for that date from the filing history: only filings delivered by then count, the status is forced to active, and nothing flagged only today is used (outcome and current-only fields cannot leak in).
- Check of the rebuild: run for today on the healthy companies and compared with the register: the confirmation statement due date matched for 100 of 100 (first) and 298 of 299 (validation); the accounts overdue flag for 95 of 100 and 291 of 299. Differences come mostly from changed accounting reference dates, which the filing list does not show.
- The ratio is the share of failed companies with the signal divided by the share of healthy ones. The interval is a 95% interval for that ratio; with a zero count 0.5 is added to each cell. The p value is a one-sided Fisher exact test.
Results: first sample
| Signal (level today) | Failed | Healthy | Ratio | 95% interval | One-sided p |
|---|
| Accounts overdue (Medium) | 20 (20.0%) | 10 (10.0%) | 2.0 | 0.99 to 4.05 | 0.0367 |
| Gazette notice for strike-off (Low, Medium or High) | 10 (10.0%) | 2 (2.0%) | 5.0 | 1.12 to 22.25 | 0.0165 |
| Registered office changed repeatedly (Low) | 2 (2.0%) | 2 (2.0%) | 1.0 | 0.14 to 6.96 | 0.6894 |
| Confirmation statement overdue (Info) | 9 (9.0%) | 4 (4.0%) | 2.25 | 0.72 to 7.07 | 0.1253 |
| Incorporated less than 12 months before (Info) | 8 (8.0%) | 3 (3.0%) | 2.67 | 0.73 to 9.76 | 0.1067 |
| Name changed in the last 12 months (Info) | 1 (1.0%) | 1 (1.0%) | 1.0 | 0.06 to 15.77 | 0.7513 |
| Overall level | Failed | Healthy |
|---|
| Low | 80 | 90 |
| Medium | 19 | 10 |
| High | 1 | 0 |
Results: validation sample
| Signal (level today) | Failed | Healthy | Ratio | 95% interval | One-sided p |
|---|
| Accounts overdue (Medium) | 42 (14.0%) | 16 (5.4%) | 2.62 | 1.51 to 4.56 | 0.0002 |
| Gazette notice for strike-off (Low, Medium or High) | 33 (11.0%) | 12 (4.0%) | 2.75 | 1.45 to 5.22 | 0.0008 |
| Registered office changed repeatedly (Low) | 11 (3.7%) | 0 (0.0%) | no healthy company | 1.36 to 388.56 | 0.0004 |
| Confirmation statement overdue (Info) | 26 (8.7%) | 20 (6.7%) | 1.3 | 0.74 to 2.28 | 0.2216 |
| Incorporated less than 12 months before (Info) | 28 (9.4%) | 26 (8.7%) | 1.08 | 0.65 to 1.79 | 0.4433 |
| Name changed in the last 12 months (Info) | 7 (2.3%) | 2 (0.7%) | 3.5 | 0.73 to 16.71 | 0.0882 |
| Overall level | Failed | Healthy |
|---|
| Low | 254 | 283 |
| Medium | 38 | 14 |
| High | 7 | 2 |
A single Medium signal sets the overall level to Medium.
What was decided from this
- Accounts overdue stays Medium. It was about twice as common before failure in both samples. Splitting it by how late the accounts were did not hold up: accounts overdue by at least 90 days looked strong in the first sample (ratio 4.67) but in the validation sample it was 1.75 (0.88 to 3.49), so no lateness threshold was adopted.
- Gazette strike-off notices were clearly more common before failure in both samples. Almost all were notices that were later withdrawn.
- Repeated registered-office changes were 2 against 2 in the first sample and 11 against 0 in the validation sample. The level is Low and is to be re-checked on a third independent sample.
- Confirmation statement overdue and company under 12 months old showed little or no difference in the validation sample, so they are context only (Info). So is a recent name change, which was too rare to judge.
Limitations
- Most failed companies showed nothing. In the validation sample 254 of 299 failed companies were Low overall a year earlier.
- Samples are small, so intervals are wide, and the validation sample was also used for the office-change decision.
- Failed companies come from those still listed as in liquidation or administration; ones already dissolved are missed, and they tend to be the faster, smaller ones. Healthy companies are those active today, so a later failure among them biases slightly towards no difference.
- Sampling is by incorporation month, not in proportion to how many companies were formed. Events lean to 2025.
- One lead time (12 months), UK limited companies and PLCs only.
- The case-control design has a 50% failed share. Real failure rates are far lower, so even a signal twice as common before failure mostly appears on healthy companies. This is not a failure model.
- Not back-tested: the website footprint, certificates, lookalike domains, sanctions name matches, SEC EDGAR signals and LEI consistency, because those sources only show today’s state.
Back to the methodology