paper-sub/neurips-3231

Impermanent: A Live Benchmark for Temporal Generalization in Time Series Forecasting

Impermanent is a live benchmark that evaluates time-series forecasting models under open-world temporal change. Unlike static benchmarks with fixed train-test splits, Impermanent scores forecasts sequentially over time on continuously updated data streams, enabling the study of temporal robustness, distributional shift, and performance stability rather than one-off accuracy on a frozen test set.

The benchmark is instantiated on GitHub open-source activity data sourced from GH Archive, a public record of GitHub event streams. It covers the top 400 repositories by star count and tracks four event types across three forecast frequencies, with daily updates and standardized leaderboards for reproducible, ongoing comparison.

Dataset Structure

Schema

Each row in the dataset represents a single observation for a repository–event-type pair at the selected frequency.

Column	Type	Description
repo_id	int64	Unique numeric GitHub repository ID
repo_name	string	Repository name in owner/repo format
metric	string	Event type: stars, issues_opened, prs_opened, or pushes
unique_id	string	Series identifier in {repo_id}:{metric} format
ds	timestamp[us]	Observation timestamp (UTC, start of period)
y	int64	Event count for the period

Configurations

The dataset has three configurations corresponding to the three forecast frequencies used in the benchmark. Each is a separate folder of Parquet files.

Configuration	Forecast horizon	Max context	First cutoff	Cutoff step
daily	h = 7 days	512 days	2026-01-04	7 days
weekly	h = 1 week	114 weeks	2026-01-04	1 week
monthly	h = 1 month	24 months	2025-10-01	1 month

Event Types

• stars: New stargazers added to the repository per period
• issues_opened: New issues opened per period
• prs_opened: New pull requests opened per period
• pushes: Push events per period

Data Source and Provenance

The data is derived entirely from GH Archive (Grigorik, 2011), a public dataset of GitHub event streams released under a Creative Commons license. GH Archive archives all public GitHub events as hourly JSON files.

Ingestion pipeline: Hourly JSON archives are downloaded from GH Archive, filtered for the four target event types, aggregated per-repository using DuckDB, and rolled up to daily, weekly, and monthly granularities. Completeness thresholds are applied: 90% of constituent hours for daily, 95% for weekly, and 99% for monthly. The pipeline runs on Modal serverless infrastructure with artifacts stored on Amazon S3 before publication here.

Repository selection: The 400 repositories were selected by star count at the time of dataset creation. Each repository–event-type pair is treated as an independent univariate series.

Time range: Data begins 2023-01-01 and is updated continuously. The evaluation window in the first release spans 20 cutoffs from 2025-10-01 to 2026-04-19.

Intended Use

This dataset is intended for:

• Benchmarking time-series forecasting models for temporal generalization — measuring whether model performance persists as data distributions evolve over time
• Studying distributional shift and concept drift in real-world event count data
• Evaluating probabilistic calibration of forecasting models on intermittent count series
• Reproducible, ongoing comparison of forecasting methods via the live leaderboard

Out-of-Scope Use

This dataset is not suitable for:

• Surveillance, tracking, or profiling of individual GitHub contributors or repository maintainers. All data is repository-level aggregate counts; no individual user activity is included.
• Drawing conclusions about the productivity or quality of specific repositories or organizations. Event counts are noisy proxies shaped by tooling, community norms, and external events.
• Use as a general-purpose time-series benchmark outside the temporal generalization setting. The benchmark is designed for sequential, prequential evaluation and may not reflect performance in standard static split settings.

Known Biases and Limitations

• Popularity bias: Only 400 repositories are included. This skews the benchmark toward popular, widely-used, and predominantly English-language software projects. Smaller, newer, or domain-specific communities are underrepresented.
• Single domain: All data comes from GitHub software development activity. Whether findings about model rankings and temporal generalization transfer to other domains (finance, energy, healthcare, etc.) requires separate evaluation.
• Univariate only: Each repository–event-type pair is treated as an independent series. The current protocol does not exploit cross-repository dependence, repository metadata, natural-language context (e.g. release notes), or other covariates that could matter in deployment.
• Short evaluation window: The first release spans 20 cutoffs over approximately six months. This is sufficient to expose short-run rank instability but is not long enough to draw strong conclusions about long-run temporal generalization.
• Platform and tooling changes: GitHub's own platform changes (API behavior, event definitions, UI changes that affect contributor behavior) can introduce structural breaks that are external to any model's forecasting target.

Personal and Sensitive Information

This dataset contains no individual-level user data. All measurements are repository-level aggregate event counts derived from public GH Archive data. No usernames, email addresses, commit messages, code content, or other personal information are included. The data is fully public and was collected from a public archive of GitHub's public event stream.

Evaluation Protocol

At each cutoff date, models receive a context window of historical observations and must produce point and probabilistic forecasts for the next h periods before any ground truth is available. Cutoffs are spaced exactly one horizon apart, and the most recent cutoff is always excluded because observations may still be incomplete.

Metrics:

• MASE (Mean Absolute Scaled Error) for point accuracy
• Scaled CRPS for probabilistic calibration, estimated from nine quantile levels Q = {0.1, 0.2, ..., 0.9}

Both metrics are scaled by the ZeroModel (constant-zero forecast) score for each evaluation group to normalize magnitudes across subdatasets and frequencies.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.

The underlying GH Archive data is also released under CC-BY-4.0. See gharchive.org for details.