SpikeProphecy benchmark reveals brain region predictability rankings

Neural population models—algorithms that predict the joint firing of many neurons forward in time—have long been evaluated by a single number: Pearson correlation between predicted and actual spike counts. A new pre-print from researchers at UC Santa Cruz, UC San Francisco, and other institutions argues that this scalar masks critical structure. They introduce SpikeProphecy, the first large-scale benchmark for causal, autoregressive spike-count forecasting on real electrophysiology recordings.

The benchmark uses 105 Neuropixels sessions from the Steinmetz 2019 and IBL Repeated Site datasets, totaling ~89,800 neurons. The core contribution is a population metric decomposition that separates aggregate performance into temporal fidelity, spatial pattern accuracy, and magnitude-invariant alignment. Testing seven architecture baselines—four state-space models (three diagonal plus one non-diagonal), a Transformer, an LSTM, and a spiking network—the authors uncover a brain-region predictability ranking that reproduces across all models and survives ANCOVA correction. They also identify a sub-Poisson evaluation floor and a negative result for KL-on-output-rates distillation in ANN-to-SNN transfer. The work is submitted to NeurIPS 2026 Datasets and Benchmarks Track.