Blind-Window Forecasting: Real-Time Benchmarking and Multimodal Reconstruction for Tropical Cyclones

Tropical cyclone forecasting models are often trained and evaluated with reanalysis inputs. Reanalysis provides physically rich environmental context, but its most recent fields may be unavailable at forecast issuance. This can introduce a hindsight-style advantage and lead to optimistic estimates of real-time forecast skill. Satellite observations, in contrast, are available closer to real time but provide incomplete and noisy views of the storm system. We formulate this data-availability mismatch as forecasting through a reanalysis blind window. To study this setting, we introduce a real-time-faithful benchmark spanning 1980--2023 that combines lagged reanalysis with visible, infrared, water-vapor, and passive microwave satellite imagery. We then develop RECAST-TC, a multimodal reconstruction framework that fuses time-lagged environmental histories with real-time satellite observations to estimate a forecast-sufficient latent storm state from delayed, partial, and noisy inputs. We use idealized analysis-available forecasting and delayed-reanalysis-only forecasting as reference regimes. RECAST-TC consistently improves track and intensity forecasts over delayed-reanalysis baselines and moves noticeably closer to the idealized reference. These results highlight data availability and observation timing as central design variables for realistic evaluation and deployment-oriented AI forecasting of tropical cyclones.