The panel
Every pattern rests on one asset: a longitudinal panel of observed professional records resolved to company entities and expanded into annual observations. The panel is the unit of evidence, not a collection of point in time snapshots.
48M+
Observed company years
4.2M
Companies in the panel
1950 to 2020
Temporal coverage
5 years
Minimum history per scored firm
Construction
From raw record to scored company
Six deterministic stages. Each stage either transforms the series or removes a company from eligibility. Nothing is imputed to fill a gap.
- 01
Record assembly
Professional records are resolved to a single company entity, then to a single role interval with a start year and an end year. Overlapping intervals at the same employer are merged. Records without a resolvable employer or without at least a start year are dropped before any aggregation.
- 02
Company year construction
Each company is expanded into one row per calendar year. A person counts toward a company year when their role interval covers that year. This produces the observed headcount series, plus function, seniority and tenure distributions for the same year.
- 03
Normalisation
Absolute observed headcount is not treated as true headcount. Series are converted to year over year growth rates and internal share ratios, so a company is always compared against its own prior shape rather than against panel coverage levels.
- 04
Eligibility filter
A company is eligible for scoring when it has five or more consecutive observed years and a minimum observed base in the first year of the window. Companies below the base threshold are reported as insufficient evidence rather than scored.
- 05
Pattern assignment
Rule based classification runs over the normalised series. Every pattern has an explicit numeric definition, so assignment is deterministic and reproducible from the same input. Where two patterns qualify, the distress or integrity pattern takes precedence over the neutral one.
- 06
Outcome linkage
Outcomes are attached only from dated public events after the observation window closes: insolvency filing, delisting, acquisition under distress, mass restructuring, or documented misstatement. Undated or press only claims are not accepted as outcomes.
Definitions used throughout
These terms carry exact meanings in every Vivameda study and in every scoring run.
- Signal
- A numeric condition on a normalised workforce series that must hold over a stated rolling window. A signal is not a score and carries no probability. It states that a shape occurred.
- Rolling window
- All conditions are evaluated over three or five consecutive observed years. A single year movement never triggers a pattern, which removes most one off reporting and coverage artefacts.
- Threshold
- Growth thresholds are fixed in advance per pattern, for example minus 3 percent or worse for contraction and 8 to 15 percent for sustained moderate expansion. Thresholds are not fitted per company or per industry.
- Lead time
- The interval between the last observed year that satisfies the pattern and the dated public outcome event. Lead time is reported per case, never averaged into a single headline number.
- Insufficient evidence
- An explicit output state. A company with thin coverage or a short history returns insufficient evidence instead of a pattern, which keeps the false positive surface low.
Validation controls
What protects the result, and what the result does not claim.
Definitions fixed before testing
Pattern thresholds were derived on data through 2020 and written down before the contemporary cases were classified. No threshold was adjusted after seeing a case outcome.
Forward looking by construction
Every validation outcome resolved after the observation window closed. The test is not a holdout split inside one dataset, it is a later real world event set.
Deterministic classification
The same input series always produces the same pattern. There is no per case human judgement inside assignment, so results can be reproduced from the panel and the rule set.
Survivorship handled explicitly
Companies that ceased to exist stay in the panel with their final observed years. Removing them would inflate the apparent stability of the historical base.
Coverage bias stated, not hidden
Observed records skew toward white collar functions and toward more recent decades. Because scoring uses internal ratios and self relative growth, coverage level differences affect eligibility rather than pattern direction.
No financial inputs
Patterns use workforce structure only. No revenue, valuation or filing data enters the classification, which is what allows the signal to move before disclosure does.
Diligence
Reproducing the result
A data diligence team can rerun the recognition test with three inputs: the normalised company year series for the named companies, the written pattern definitions, and the dated outcome events. Because classification is rule based, the output is verifiable rather than reported. Requests for the definition sheet and validation workbook go through the data room.