RELIABILITY-ADAPTIVE FUSION FOR STREAMING MULTIMODAL TIME SERIES IN DECISION SUPPORT SYSTEMS

Streaming forecasting models based on multimodal data must remain robust when individual modalities undergo temporary degradation [1, 2]. This paper investigates Reliability-Adaptive Dynamic Fusion (RADF), a method that adjusts fusion weights in real-time based on calibrated reliability signals [4]. We present a stress-testing protocol involving alternating missingness and scale shifts and report practical findings from controlled benchmarks and real-world domains (Energy, Crypto) in a prequential setting [6]. Results show that RADF reduces Mean Absolute Error (MAE) by up to 93% in severe stress scenarios compared to naive fusion, while maintaining optimal performance in clean regimes via a “safety gate” mechanism.