Problem
Forecast macro indicators with sparse, lagged public data and no access to private feeds.
Scarcity components exercised
Constraint ledger (data latency + availability), experiment rail (nowcasting with synthetic gaps), narrative layer (communicating uncertainty ranges).
What worked
Bootstrapping with proxy signals and robust uncertainty estimates kept forecasts honest.
What failed
Seasonality adjustments were brittle when public releases slipped; models overreacted to noise in small series.
Next steps
Test online-learning variants that adapt gracefully to release delays; expand proxy set and monitor drift aggressively.