When Context Matters: State-Correction Forecasting for Tropical Cyclone Intensity

Intuitively, adding contextual information such as satellite images or reanalysis fields to a pressure-history model should improve tropical cyclone intensity forecasts. However, we empirically find the opposite is often true. On two datasets, naive fusion baselines are less accurate than pressure history alone at the earliest forecast steps, with MAE up to 47\% higher at the shortest lead time. The extra information can therefore hurt precisely where the pressure history is most predictive. This suggests that contextual information should not necessarily influence every forecast to the same extent. In this paper, we propose State-Correction Forecasting (StCF) to naturally integrate pressure-history and contextual information while preserving predictive performance. More specifically, StCF employs a pressure-history forecaster as its backbone and adds a state-derived correction gated by a learned, sample and horizon-dependent weight. The gate controls, for each sample and horizon, how strongly contextual information modifies the pressure-history forecast. Because the forecast decomposes natively into a pressure-history term and a correction term, every prediction can be audited: how much came from history, and how much from context. Across Digital Typhoon and ERA5, StCF remains close to the pressure-history model at short lead times, where naive fusion underperforms. At longer horizons, as contextual information becomes more beneficial, StCF matches or outperforms the fusion baselines. A shuffled-state control shows that the gain depends on informative contextual state: it disappears when context is mismatched across storms.