Origin of the forecast horizon limit in neural activity forecasting

Determine whether the observed limit on informative probabilistic forecasting horizons for spontaneous mouse cortical widefield calcium imaging data—approximately up to 1–1.5 seconds ahead—arises from design constraints of current time series forecasting models (including autoregressive models and deep learning architectures such as PatchTST, TiDE, Chronos, and related methods) or from intrinsic sources of variability and characteristic time scales in mouse cortical neural activity.

Background

The paper benchmarks classical statistical models and modern deep learning and foundation models for probabilistic time series forecasting on spontaneous mouse cortical activity recorded via widefield calcium imaging at 35 Hz. Models such as PatchTST, TiDE, and fine-tuned Chronos outperform classical approaches, but their predictive accuracy deteriorates with increasing horizon.

Step-wise analyses show that beyond roughly 1–1.5 seconds, predictions converge toward the mean and variance of training data, and uncertainty estimates (ratio of predicted distribution standard deviation to training data standard deviation) approach one. This suggests a practical ceiling on forecast informativeness within the tested setting.

The authors explicitly state that it remains unknown whether this ceiling is imposed by current model designs or is intrinsic to the variability and time scales of neural activity, motivating future work to disentangle these factors through improved modeling and experimental approaches.