A Forecasting Benchmark of Large-Scale Neural Populations during Task-Driven Behavior

While substantial advances have been made in decoding behavioral or stimulus variables from 2-photon calcium imaging data—the forward problem of predicting future neural activity from past population dynamics remains underexplored. We leverage a novel, large-scale calcium imaging dataset from the International Brain Laboratory, where mice perform a complex decision-making task, to systematically evaluate neural forecasting. Using 11,393 neurons recorded simultaneously for over an hour, we applied standard forecasting techniques to assess the predictability of this system. Through experiments examining population scaling, prediction horizon, and cortical state transitions, we find that forecasting performance exhibits diminishing returns with population size and degrades substantially when models encounter unseen brain states. Simple univariate models remain competitive with sophisticated multivariate architectures, suggesting current approaches fail to exploit higher-order population codes. This work establishes a single-session benchmark for neural population forecasting on a large-scale mesoscope recording from behaviorally-engaged mice.