Calibration and validation¶
-
27 choice models: sample, fit and hold-out of each.
-
The full results of all 27 models: coefficients, robust standard errors, p-values.
-
The demographic levers of the land-use model, the only calibrated parameters.
-
Every stage against data it never saw.
-
Interactive map of every count station against the simulated traffic.
-
CVAE vs IPF and IPU, against the original land-use and activity models.
-
What the model does not do, stated plainly.
Headline results¶
| Stage | Test | Result |
|---|---|---|
| Synthetic population | ACS 2014–18, Maryland | every attribute class within 3.9 pp |
| Land use | out of sample, 2021–24 | commute flows correlate 0.995 with the ACS; household size and cars within 2.1 pp in 2024 |
| Activities | survey hold-out half | +1.2 % activities, start times within 2 pp per hour |
| Assignment | 2,308 count stations, 2024 | 0.77 of all counted vehicles carried by residents alone; R² = 0.89 |
The land-use, activity and traffic results are updated after the final runs of the chain with the current models.