Methods note

We let the data try to pick the weights. It couldn't beat judgment.

Fitting dimension weights to COVID-era outcomes across 3,080 counties produced weights no more predictive out of sample than the ones we chose — and far less stable.

Choosing the Weights: an Empirical CheckSep 10, 2026United StatesPublished

Read the full report

Chosen vs. fitted

The data wanted housing and diversification. We weight program reliance.

Dimension weight, share of composite (six structural dimensions)

Chosen (renormalized)

Economic diversification
19%
Employment
19%
Income & poverty
19%
Program reliance
25%
Human capital
9%
Housing
9%

Empirical fit, full sample

Economic diversification
28%
Employment
18%
Income & poverty
5%
Program reliance
16%
Human capital
11%
Housing
22%

Recovery dynamics is excluded from the fit because it is measured on the same shock (leakage).

Stability

The fitted weights move a lot between folds.

Standard deviation of each fitted weight across cross-validation folds, in percentage points. A weight that swings by 10 points is not a weight, it is noise.

1Economic diversification10
2Income & poverty9
3Program reliance6
4Housing3
5Human capital3
6Employment2

Cross-validated; seed fixed and recorded.

Out of sample

Neither set predicts the COVID unemployment spike.

Spearman correlation between composite and 2020 unemployment spike, held-out folds. All near zero.

0.036

chosen weights

0.014

empirically fitted weights

0.015

equal weights

How this is calculated, and what it does not say›

Six leakage-free structural dimensions; recovery_dynamics excluded. One shock only — evidence, not yet grounds to change the index.

The honest conclusion: one shock cannot set the weights. A second, conventional recession in the data would; until then the weights stay a disclosed judgment with a published sensitivity test.

Dig deeper

Methodology & research brief

National report

Data in this edition

See the report's methodology section.

Sources carry different vintages and lags by design and are never combined inside one figure. The write-up's caveats section says what the index does and does not claim.

Cite as

Alt-30 (2026). Choosing the Weights: an Empirical Check, Methods note (v1.0.0). Alt-30. https://portal.alt-30.com/reports/eri-weights-note