Black-Mamba Improves Forecasting Under Distribution Drift with Biologically-Inspired Adaptation

Real-world forecasting often deals with non-stationary data where underlying distributions change over time, a phenomenon known as distribution drift. While some test-time adaptive models update continuously based on prediction errors, this can lead to inefficient adaptation by conflating transient noise with genuine, persistent shifts. Researchers have introduced Black-Mamba, a novel forecasting architecture inspired by biological processes. It formulates online adaptation as evidence-gated state tracking, augmenting a base predictor with a dynamic memory. This memory is updated only when a sufficient amount of temporally accumulated surprisal provides clear evidence of a regime change, making adaptation an event-driven rather than continuous process. Across various non-stationary forecasting benchmarks, Black-Mamba achieves competitive or superior predictive performance compared to existing test-time adaptation methods. Crucially, it significantly reduces the number of memory updates during inference, demonstrating a more efficient and robust way to handle distribution drift by distinguishing persistent shifts from transient noise.