Historical city-level records
Economic realism is a testable goal—not a guarantee.
First generate an immutable-parameter forecast commitment without reading scoring labels, then evaluate Engine v0.5.0 against frozen v0.4.2 and a naive persistence baseline on 2023–2024 observations. The simulated starting point remains synthetic. Employment is measured against BLS unemployment rate, not a count of jobs. This is label-blinded execution, not an independent preregistered study.
Six economic measures
BLS city-level unemployment observations
Comparable records are absent for four requested indicators. The model calculates them, but the lab does not invent missing observations or report synthetic accounting checks as historical accuracy.
Holdout comparison
| Indicator | Policy branch | Legacy MAE | v0.5.0 MAE | Naive MAE | v0.5.0 vs legacy | 2024 sensitivity range | Holdout years |
|---|
MAE units are persons for population and percentage points for unemployment. Results are not a policy effectiveness ranking. A new engine can perform worse than its predecessor.
Uncertainty assessment — all six measures
The 2024 range below comes from the sealed engine plus 14 single-coefficient ±25% perturbations. This is a deterministic sensitivity envelope, not a confidence interval or an estimated probability of future outcomes. The missing four historical measures still receive internal model uncertainty tests, not historical accuracy scores.
| Economic measure | Low, 2024 | Base, 2024 | High, 2024 | Evidence |
|---|
Housing affordability is a model index, not a measured rent burden; private investment counts project commitments, not all regional capital formation. Government debt includes bonds; spending is annual modeled municipal outlays.
Methodology and limitations
- Historical records are provided in a fixed, bundled public catalog; no live updates or silent corrections occur.
Read Project Documentation for the model and data caveats. The historical backtesting lab continues to use frozen Engine 0.4.2.