What makes an accelerator
actually work.
PANDA Research publishes open-access papers on accelerator effectiveness, founder readiness, and the math behind every score in the PANDA Lab. We publish what we learn so other accelerators — and other founders — can use it.
The Accelerator Effect — what fifteen years of data say about whether they work.
A meta-analysis of 4,237 cohort outcomes across Y Combinator, Techstars, 500 Global, Antler, Plug and Play, MassChallenge, and a long tail of regional and university accelerators between 2010 and 2025. Survivorship-adjusted. We rank the factors that actually predict 5-year survival, capital raised, and founder satisfaction — and we name the things accelerators say they do that don't show up in the data.
Three companion papers. Same methodology.
Each paper is short, peer-reviewed by PANDA's mentor board, and built around a single empirical question. All four together = the manual behind the PANDA program.
The Idea Score — scoring math + dimension calibration.
The math behind the PANDA Idea Score widget. Twenty questions, seven dimensions, calibration via logistic regression against 4,237 cohort outcomes. Includes the full calibration table, weighted dimension equations, and the cross-validation methodology used to lock the v3.1 weights.
"We present the calibration methodology for the PANDA Idea Score (v3.1), an open scoring system that predicts founder outcomes from a 20-question survey. Logistic regression against a hold-out set of 1,059 cohorts produces a 0.81 AUC for 5-year survival prediction…"
Read PDFThe Student-Stage Gap — why college founders need a different curriculum.
Standard accelerator curricula are built for full-time founders with prior experience. We show that the modal college founder fails the program at 3.2x the rate of their post-graduation peers — and that the failure mode is not capability but program-fit. We propose a 14-week structure designed around the unique constraints of student founders.
"We analyze 1,612 dropout events from college-stage cohorts (n=4,237) and find that 71% cluster in weeks 5–7 around three predictable failure modes: misaligned time budget, mentor-curriculum gap, and the 'cofounder mismatch' phenomenon…"
Read PDFFounder Readiness — the five dimensions that predict program completion.
The companion paper to the Founder Readiness diagnostic. We isolate the five founder-side variables that most reliably predict program completion (separate from outcome): grit, runway, time budget, network depth, and learning velocity. We publish the question bank, the scoring rubric, and the cross-validation results.
"We construct an 18-item Founder Readiness scale across 5 dimensions and validate it against program completion (n=4,237). Internal consistency: Cronbach's α = 0.79. Concurrent validity vs. mentor-rated readiness: r = 0.66…"
Read PDFThe 2026 Accelerator Landscape — trends, mergers, and the rise of vertical programs.
The first PANDA Research quarterly brief. Tracks the consolidation of horizontal accelerators (Techstars closing locations, 500 restructuring), the rise of vertical programs (biotech, climate, fintech, defense), and the explosion of university-affiliated programs (38 new launches in 2025 alone).
"2025 marked an inflection: for the first time, the number of vertical accelerator programs (n=412) exceeded the number of horizontal programs (n=389). University-affiliated accelerators grew 18.4% YoY…"
Read PDFBuilt on a public dataset. Reproducible by anyone.
Every paper in PANDA Research is built on the same underlying dataset. We started with the public outcome data from Y Combinator, Techstars, 500 Global, Antler, Plug and Play, and MassChallenge — supplemented by 38 university programs that share their cohort outcomes publicly. We layered our own PANDA cohort data on top. The combined dataset spans 4,237 cohorts and ~31,000 founders.
The full dataset, the scoring math, and the analysis code are published. If you want to replicate Vol. 04 Paper 02's logistic regression against your own program's data, every input we used is downloadable.
Use it. All PANDA Research is published under Creative Commons Attribution 4.0. Reprint it, adapt it, build on it. The only ask: attribute PANDA Research and link the source. If you're building a competing accelerator and our work helps you do it better, we consider that a win.