Vivameda

Models & Research

The Models Behind Our Intelligence

Vivameda develops numerical forecasting models around longitudinal company and workforce evidence. This page documents what each model estimates, the evidence it uses, how it was tested, and where its results currently stand.

Isometric illustration of layered historical data feeding a numerical forecasting engine

Featured prediction model

Vivameda Workforce Forecast

Research Preview

A numerical forecasting model that estimates the direction and scale of change in a company's observed workforce one year ahead. It is built entirely from longitudinal company records, so a company is always compared against its own prior trajectory rather than against a market average. Two variants are maintained: a core growth model and a tenure-enhanced model that adds employment structure where that evidence exists.

Core Growth Model

The baseline estimate. It uses the two quantities present for nearly every company with usable history in the panel, which makes it the reference behaviour every later variant is measured against.

Inputs
Observed company size at the forecast origin and prior-year growth
Returns
An estimate of next-year change in observed workforce profiles
Coverage
Calculated for the widest eligible population in the panel
Role
Establishes the reference result for comparison

Tenure-Enhanced Model

The same estimate recalculated with employment structure included. It investigates whether the shape of employment inside a company carries information that size and recent growth do not already provide.

Inputs
Observed company size, prior-year growth and workforce tenure distribution
Returns
An estimate of next-year change with tenure composition in the feature set
Coverage
Reported for companies where tenure evidence exists, so the population is narrower
Role
Tests the contribution of employment structure to the estimate

How a forecast is produced

Every run follows the same sequence, from the observation window through to the comparison that makes the result auditable.

  1. 01

    Observation window

    The model reads a company's continuous observed history up to one stated forecast origin. Only records dated no later than that origin are available to it, which keeps later information out of the estimate.

  2. 02

    Feature construction

    Two quantities carry the core estimate: observed company size at the forecast origin and the change recorded over the preceding year. Observed counts are treated as relative measures rather than absolute headcount, because panel coverage differs between companies and between decades.

  3. 03

    Estimate

    The model returns an expected change in observed workforce for the year following the origin. The output is a workforce trajectory estimate, not a probability of success, failure or acquisition.

  4. 04

    Outcome comparison

    Each estimate is scored against the change actually observed in the following year and against simple forecasting benchmarks calculated over the same company years.

Target
Next-year change in observed workforce
Horizon
One year
Output
Expected workforce trajectory for the following year
Evidence
Longitudinal company and workforce records only
Variants
Core growth and tenure-enhanced
Deployment
Vivameda-controlled infrastructure

Evaluated retrospectively on historical company data. The model is published here as a research preview while validation for present-day forecasting continues.

Research evidence

Measured Against Historical Outcomes

347,942

Historical company-year predictions evaluated

50,084

Company IDs represented in the historical evaluation

10

Annual forecast origins, 2010–2019

The historical evaluation compares the numerical models with simple forecasting benchmarks over the same company years. The structure of that evaluation is fixed and documented below.

Held-out companies

Company IDs used in the evaluation are excluded from model fitting, so no company is scored against a model that already saw it.

Label cut-off per origin

Training labels are restricted to years no later than each forecast origin, which keeps later outcomes out of the fit.

Repeated origins

The evaluation runs across ten annual forecast origins, so the result is not tied to a single year or a single economic moment.

Benchmark comparison

Numerical estimates are compared with simple forecasting benchmarks over the same company years rather than presented alone.

The evaluation population is smaller than the broader data foundation, because only companies with sufficient continuous history can be scored. These figures describe retrospective development evaluation, rather than prospective forecasting accuracy.

Scope and interpretation

What the Models Support and What They Do Not

What the models support

  • Comparing a company's expected workforce trajectory against its own recorded history
  • Screening a defined population of companies before deeper analysis begins
  • Framing diligence, allocation and market questions with a longitudinal baseline
  • Reproducing an estimate from the same records and the same fitted parameters

What they do not do

  • No probability of failure, success or acquisition is produced by these models
  • No revenue, financial statement or valuation data enters the described variants
  • No blending of historical panel evidence with present-day observations into one series
  • No prospective accuracy claim is made while present-day validation is ongoing

Research workflow

Vivameda Research Agent

Operational — Internal Research

Our research agent retrieves project evidence, runs installed historical models and preserves conversation and research context. It brings numerical analysis and supporting documentation into a connected research workflow.

Evidence Retrieval

Searches the project knowledge archive and surfaces supporting references.

Model Execution

Runs installed numerical models for supported historical research cases.

Research Memory

Preserves conversations and research records for continuity.

The agent uses locally hosted language model support as underlying technology. Research memory stores records and conversations; it does not retrain model weights.

Data foundation and infrastructure

Built on Longitudinal Evidence

Vivameda’s broader data foundation spans 4M+ companies and 48M observed company-year records covering 1950–2020. Our research infrastructure brings this historical context together with numerical models, project evidence and persistent research memory.

Model code, fitted parameters and research records are maintained on Vivameda-controlled infrastructure, supporting reproducible development and continuity.

The model evaluation above uses a smaller eligible population drawn from this broader foundation.

Bring a Company Question to the Research

Explore how longitudinal workforce evidence can support company analysis, target assessment and investment research. We start with your question, assess the available evidence and define an appropriate research scope.

Discuss Your Research Case