Most flight-delay prediction systems rely on real-time meteorological data, which are rarely available in large historical aviation archives. This paper shows that competitive performance is still achievable using only historical operational data. We train a cost-sensitive CatBoost classifier on a 240-million-record dataset of U.S. domestic flights, under strict chronological validation, and introduce a Temporal Undersampling strategy that addresses the severe class imbalance by removing the oldest On-Time records rather than random ones. The proposed approach outperforms four baseline imbalance-handling strategies and, on a balanced evaluation set, exceeds the best F1-Score reported by a recent IEEE DASC reference study that relied on weather data.