Hackalysis: Detecting Hackathon-Origin Software Through Temporal and Structural Repository Traces

Abstract

Hackathons are among the most visible contemporary sites of time-bounded collective innovation: self-organizing teams converge for one to three days to produce a shared artifact. This paper addresses the software-centric subset of these events, in which the output is a working code prototype. Once an event ends, however, the repositories these teams leave behind are difficult to identify within the broader open-source ecosystem. Because hackathon repositories are rapid prototypes rather than sustained projects, conflating them with long-lived software distorts analyses of repository health, maintenance, and open-source sustainability. This paper presents Hackalysis, a metadata-driven framework for detecting and characterizing hackathon-origin GitHub repositories using only publicly available signals. From LauzHack (2023–2025) and an expert-curated control set of sustained academic and professional open-source projects, a labeled dataset of 458 repositories with complete GitHub metadata is constructed. Temporal, activity, social, and structural features are derived, and four classification strategies of increasing complexity are compared: rule-based scoring, correlation-weighted scoring, a Random Forest classifier, and an unweighted majority-voting consensus. A modest Random Forest is the strongest of the four (87.1% accuracy, 0.946 AUC-ROC under 5-fold cross-validation), and the majority-voting consensus does not improve on it. Across all methods, hackathon-origin repositories exhibit a consistent dual signature: concentrated development bursts together with structural minimalism, including markedly lower prevalence of tests, CI/CD workflows, deployment configuration, documentation, licenses, and contribution guidelines. The single most discriminative signal is the active-development window: hackathon repositories concentrate essentially all commits within a 1–3 day burst. These results show that short-term collective innovation leaves detectable behavioral and structural traces in software repositories and provide a reproducible basis for isolating innovation-contest software and studying how its outputs relate to the sustained open-source ecosystem around them.