Datasets:

LeadLagForecasting
/

llf_github

dataset_info:
  features:
    - name: id
      dtype: string
    - name: platforms
      sequence: string
    - name: created_date
      dtype: timestamp[ms]
    - name: first_create
      dtype: float64
    - name: first_push
      dtype: float64
    - name: first_star
      dtype: float64
    - name: first_fork
      dtype: float64
    - name: cumulative_pushes_w_creates
      sequence: int64
    - name: cumulative_pushes
      sequence: int64
    - name: cumulative_stars
      sequence: int64
    - name: cumulative_forks
      sequence: int64
  splits:
    - name: train
      num_bytes: 87650021696
      num_examples: 938567
    - name: labeled_test
      num_bytes: 21930381181
      num_examples: 234640
    - name: val
      num_bytes: 21906489783
      num_examples: 234615
    - name: unlabeled_test
      num_bytes: 50784636048
      num_examples: 1568622
    - name: train_extra
      num_bytes: 6294971369
      num_examples: 58077
  download_size: 1094625852
  dataset_size: 188566500077
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: labeled_test
        path: data/labeled_test-*
      - split: val
        path: data/val-*
      - split: unlabeled_test
        path: data/unlabeled_test-*
      - split: train_extra
        path: data/train_extra-*
    viewer_feature_names_to_ignore_statistics:
      - cumulative_pushes_w_creates
      - cumulative_pushes
      - cumulative_stars
      - cumulative_forks
license: cc-by-4.0
task_categories:
  - time-series-forecasting

Dataset Card for Lead-Lag Forecasting Benchmark: GitHub Dataset

This dataset is part of the benchmark introduced by the paper Benchmark Datasets for Lead-Lag Forecasting on Social Platforms. It provides time series data for over 3 million GitHub repositories, for modeling and predicting long-term impact (quantified by cumulative forks) from short-term signals (pushes and stars). Please visit our project page https://lead-lag-forecasting.github.io for more information.

Paper Abstract

Social and collaborative platforms emit multivariate time-series traces in which early interactions-such as views, likes, or downloads-are followed, sometimes months or years later, by higher impact like citations, sales, or reviews. We formalize this setting as Lead-Lag Forecasting (LLF): given an early usage channel (the lead), predict a correlated but temporally shifted outcome channel (the lag). Despite the ubiquity of such patterns, LLF has not been treated as a unified forecasting problem within the time-series community, largely due to the absence of standardized datasets. To anchor research in LLF, here we present two high-volume benchmark datasets-arXiv (accesses → citations of 2.3M papers) and GitHub (pushes/stars → forks of 3M repositories)-and outline additional domains with analogous lead-lag dynamics, including Wikipedia (page views → edits), Spotify (streams → concert attendance), e-commerce (click-throughs → purchases), and LinkedIn profile (views → messages). Our datasets provide ideal testbeds for lead-lag forecasting, by capturing long-horizon dynamics across years, spanning the full spectrum of outcomes, and avoiding survivorship bias in sampling. We documented all technical details of data curation and cleaning, verified the presence of lead-lag dynamics through statistical and classification tests, and benchmarked parametric and non-parametric baselines for regression. Our study establishes LLF as a novel forecasting paradigm and lays an empirical foundation for its systematic exploration in social and usage data.

Dataset Features

Name	Type	Description	Extraction / Notes
id	`string`	Repository identifier: `owner_name/repo_name`	From Ecosyste.ms metadata
platforms	`list[string]`	List of platforms the repo was released on (e.g., PyPI, npm)	From Ecosyste.ms metadata
created_date	`timestamp`	Date of GitHub repository creation	From Ecosyste.ms metadata
first_create	`float`	Days between first `CreateEvent` and creation date. `!=0` implies anomaly.	Comparison between GH Archive and Ecosyste.ms
first_push	`float`	Days between first `PushEvent` and creation date. `<0` implies anomaly.	Comparison between GH Archive and Ecosyste.ms
first_star	`float`	Days between first `WatchEvent` (star) and creation date. `<0` implies anomaly.	Comparison between GH Archive and Ecosyste.ms
first_fork	`float`	Days between first `ForkEvent` and creation date. `<0` implies anomaly.	Comparison between GH Archive and Ecosyste.ms
cumulative_pushes_w_creates	`list[int]`	Daily cumulative pushes + creates (of repo, branches, and tags) since creation date.	Summed per day from GH Archive events
cumulative_pushes	`list[int]`	Daily cumulative pushes without `CreateEvents` since creation date.	Summed per day from GH Archive events
cumulative_stars	`list[int]`	Daily cumulative `WatchEvents` (stars) since creation date.	Summed per day from GH Archive events
cumulative_forks	`list[int]`	Daily cumulative forks since creation date.	Summed per day from GH Archive events

Dataset Splits and Statistics

Split Name	Description	Number of Examples
train	Random sample from labeled packages with >= years of activity and no anomalies.	938,567
val	Random validation subset.	234,615
labeled_test	Random test subset.	234,640
unlabeled_test	Repositories with < 5 years of activity.	1,568,622
train_extra	Repositories with anomalies.	58,077

Citation

If you use this dataset in your research, please cite the following paper:

@article{kazemian2025benchmark,
  title={Benchmark Datasets for Lead-Lag Forecasting on Social Platforms},
  author={Kazemian, Kimia and Liu, Zhenzhen and Yang, Yangfanyu and Luo, Katie Z and Gu, Shuhan and Du, Audrey and Yang, Xinyu and Jansons, Jack and Weinberger, Kilian Q and Thickstun, John and others},
  journal={arXiv preprint arXiv:2511.03877},
  year={2025}
}