SpikeProphecy: A Large-Scale Benchmark for Autoregressive Neural Population Forecasting
SpikeProphecy: A Large-Scale Benchmark for Autoregressive Neural Population Forecasting
Authors: John R. Minnick, Jinghui Geng, Kamran Hussain, Jesus Gonzalez-Ferrer, Ash Robbins, Mohammed A. Mostajo-Radji, David Haussler, Jason K. Eshraghian, Mircea Teodorescu Date: 2026-05-13 Paper ID: openalex:2605.12992
Summary
SpikeProphecy is a large-scale benchmark designed to advance autoregressive neural population forecasting beyond aggregate correlation metrics. By introducing a novel population metric decomposition, the authors separate performance into temporal, spatial, and alignment-based dimensions, revealing critical model behavior previously hidden by scalar metrics. Evaluated on over 89,000 neurons across seven architectural families, the benchmark exposes biophysical constraints in spike forecasting and provides a rigorous evaluation protocol for neural population dynamics.
Key Contributions
- Introduces SpikeProphecy, the first large-scale benchmark for causal, autoregressive neural population spike-count forecasting.
- Proposes a population metric decomposition that isolates temporal fidelity, spatial pattern accuracy, and magnitude-invariant alignment to provide deeper model evaluation.
- Demonstrates consistent brain-region predictability rankings across seven diverse model architectures that persist after controlling for firing statistics.
Open Questions & Future Work
Key Concepts
- population-metric-decomposition: A framework for decomposing neural spike count forecasting metrics into temporal, spatial, and magnitude-invariant components.
Archivist Review
I approved the population metric decomposition as it represents a sophisticated, reusable methodological contribution for evaluating high-dimensional time series. I also approved the sub-Poisson forecasting bottleneck because it highlights a fundamental limitation in current neural forecasting metrics that requires structural research beyond just adding data or architecture depth. I rejected the benchmark datasets as they are established community standards and not standalone contributions that necessitate a new vault entry in this context.
Approved Concepts
- Population metric decomposition: Provides a rigorous framework for evaluating high-dimensional neural population dynamics by avoiding the information loss inherent in aggregate scalar metrics.
Approved Open Questions
- Sub-Poisson forecasting bottleneck: Identifying the root of the sub-Poisson noise floor is essential for accurately assessing predictive progress in neural forecasting and avoiding metric-induced bias.
Rejected Candidates
- [dataset] Steinmetz 2019 / IBL Repeated Site (
steinmetz-2019-ibl-repeated-site) - low_impact: These datasets are specific to a domain and are often used as standard reference sets; I am prioritizing more novel methodological concepts over specific benchmark datasets per policy.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.12992
- paper_id
- 2605.12992
- paper_source
- openalex
- domain
- nlp
- tags
- time-series-forecastingneurosciencebenchmarking
- architectures
-
- datasets
-
- concept_slugs
- population-metric-decomposition
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-16T06:01:39Z