ARGOContributor Decision Intelligence

open-source contributor churn · GitHub Archive via BigQuery

Command Suite

Cohort health across 24 Python repositories: who is contributing, how much is shipping, and how much of last month's cohort came back.

Recent GitHub activity

Manual updates

Loading recent activity…

Explore the full community report →

The analysis below, model benchmarks, and retention queue use the fixed 2023 study.

Historical study · 547,434 events · 24 repos2023-07-01 → 2024-01-01

Loading cohort metrics…

0s
Streaming warehouse telemetry...

Contributor Activity Summary

What is ARGO?

ARGO predicts open-source contributor churnfrom 547,434 real GitHub Archive events across 24 production Python repositories. It answers one falsifiable question: given a contributor's behaviour across a four-month observation window, will they go completely inactive over the following two months?

Why Maintainer Attrition Matters

A project does not die when its issue count spikes. It dies when the three people who understand the build system quietly stop showing up. That signal sits in public event streams nobody aggregates, and by the time a maintainer announces they are stepping back the knowledge has already left.

Pipeline

01Extract

Partition-pruned BigQuery scan over githubarchive.day.2023*. The payload column is never read; bots are filtered at source.

02Land

Zero-copy Arrow handoff into DuckDB. No CSV round-trip, no dtype coercion, no intermediate serialisation.

03Transform

dbt marts with half-open observation and target windows, recency anchored to the window edge, seven schema tests gating the build.

04Model

XGBoost fitted on the chronologically earlier cohort, tuned with forward-chaining CV, explained per prediction with SHAP.

No synthetic data. A model trained on generated events can only rediscover the rules its author wrote — the earlier synthetic build scored 0.97 ROC-AUC and meant nothing. On real events the honest number is 0.7553, measured on a temporal holdout the model never saw.