Aarogya Grid

Each state fits its own model and publishes statistics only: a monthly demand multiplier per catalogue item, its standard error, and a vacancy rate per cadre. Those are pooled into a national prior by random effects, with the between-state variance estimated from the nodes rather than chosen — so a state keeps its own estimate exactly to the extent its own data earns it.

At 30 days a newcomer has seen one month and cannot tell a seasonal month from an average one at all: it publishes no informative multiplier, takes the national prior outright, and forecasts 36.0% closer to observed demand than it manages alone — recovering 97% of the gap to a full-history fit of itself. By 60 days it starts keeping some of its own.

Limitation, stated plainly: All thirty-six states and union territories are generated by one seeded simulator, so genuine between-state heterogeneity is small by construction. The tau^2 recovered here is therefore largely an artefact of sampling, the pooling weights are a demonstration rather than a finding about Indian states, and the prior transfers better than it would between thirty-six real health systems.