Air Quality Forecasting for Dhaka: Data Integrity over Model Complexity

We show that data integrity and feature design dominate model complexity for Dhaka AQI forecasting. A two‑check audit reveals that a widely used 26‑year hourly dataset is synthetic before August 2022; training on it yields negative skill. On the verified record, LightGBM with perfect‑prognosis weather features significantly outperforms a Transformer fusion model at +24 h (Diebold–Mariano tested). An operational per‑horizon forecaster retains positive skill from +1 h to +72 h using only free public data.