Systems and tools Illustrative · magnitudes withheld

Labor market digital twin

The twin builds a state’s workforce accounts directly, with five inflows, three outflows and demand estimated four ways, and solves for the gap between supply and demand.

Economics · policy · data science · engineering. First end-to-end run projects a state nursing workforce to 2033 and is informing state policy.
01

The problem

States plan their workforce with statistics that are years out of date and kept in separate places. Employment, graduations and migration each come from a different source, and they do not share an accounting framework. Projections come as black boxes, with a number at the end and no way to see how it was reached. The twin models the market as stocks and flows that have to balance, and every term is measured, sourced and reported with its uncertainty.

02

How it works

01Supply as conserved flows

Cohort accounting over 501 occupations and 51 jurisdictions: graduates in from the training pipeline, migration in and out from nineteen vintages of census microdata, exits and retirements from empirical age-specific rates. The accounting identity has to hold, so no term can be adjusted to make the numbers fit.

Economics · data science

02Demand built four ways, never averaged

Demand is estimated from requisition signals, structural projections, population-based forward need and facility blueprints, and the four estimates are kept separate. Where they agree you can act on them. Where they diverge, the difference points to the estimate that is missing something.

Economics · data science

03Sourced terms, flagged gaps

Transfers are removed from exits before anything is projected. Thin cells fall back to national rates and say so. Where two decades of occupational codes are harmonized, uncertain matches are flagged.

Data science · engineering

04From projection to decision

The engine also produces transition pathways limited by credential requirements and wage premiums, forward-looking market signals, and policy levers graded by the strength of their evidence, so a projected gap comes with options for closing it.

Policy · economics
Model

For one occupation in one region, five inflows add to the stock and three outflows reduce it, demand is built four separate ways, and the gap between them is the output.

The whole model Supply flows through the stock, demand is built four separate ways, and the gap between them is the output.
Adjacent population

Supply for an occupation includes the people already in it and everyone one retraining step away. For a target occupation the twin lists that pool, showing which occupations are skill-adjacent, how many people are in each, and what training it takes to move across.

Each occupation around the target is one retraining step away. Near steps are drawn solid, mid dashed, far dotted.
Adjacency from 798k occupation-pair similarities, tiered by retraining distance. The reachable pool is what the model feeds into the adjacent-population inflow, discounted by the credential and wage distance of each move.
Demand, from four angles

Demand is estimated four separate ways, each picking up a different signal over a different time horizon, and the four are not averaged. When they agree, that agreement is the reason to be confident, and when they diverge, the difference shows which horizon is changing.

Requisition signal
Structural growth
Forward need
Blueprint
Four estimates of the same demand, each measuring something different.
Short horizons are forecast from requisition dynamics with ARIMA and SARIMA, longer horizons come from structural projections and population-anchored need, and gradient boosting is in testing as a cross-check.
03

Tools used

Python, pandas, numpy, scipyCohort accounting over 501 occupations and 51 jurisdictions.
statsmodelsARIMA and SARIMA demand projection at short horizons.
scikit-learnGradient boosting, in testing as a demand cross-check.
ACS PUMS microdataBilateral migration flows across 19 census vintages.
Kaplan-Meier, cohort methodsExits, retirements, and retention.
Supabase, PostgresThe workforce panel and engine outputs.
Dash, PlotlyThe state dashboards over the engine.
04

Final state

This is the structure of the production system, which runs the same flows, identities and demand layers over real census microdata, training-pipeline records and employer panels. The first end-to-end run projects one state's nursing workforce through 2033 across all of its counties, and is informing state workforce policy and training investment. Where the engine can be checked against the federal system, it has been, and employer-survey forward demand for the pilot occupation comes within 7% of the federal projection. The curves and magnitudes shown here are illustrative, and the real runs carry their sources and uncertainty with them.

← Back to systems and tools