Data & Methodology

Scoring System, Evaluation Framework & Results

Executive Summary

Science-fair competitions (ISEF, STS) select for a rare combination of technical depth, independent research ability, and competitive drive — traits that compound in venture-backed founders. Science Fair Fund uses a proprietary ML scoring system to evaluate these alumni at each investment stage, enabling three portfolio construction decisions: prioritizing sourcing, sizing initial checks, and concentrating follow-on capital.

• Primary metric: $500M+ outcomes (stable sample, directly relevant to fund returns). $100M+ and $1B+ shown for context.
• Two evaluation frameworks: Selection Advantage (does scoring concentrate winners within each stage?) and Predictive Validation (does the model generalize forward in time?).

Data Universe

Reference cohort = VC-investable alumni-linked companies founded 2006–2020, with outcomes observed through early 2026. Every company has at least five years of maturity.

Over 90% of $1B+ outcomes are US-headquartered, consistent with the fund's North American deployment focus. Outcome rates are stable across founding vintages.

Scoring Methodology

The fund operates two complementary evaluation systems, each answering a different question about alumni founders.

Ranking Model (V4 Ensemble). Each company receives a stage-specific score based on its highest-scoring founder. The score is a within-cohort rank — not a calibrated probability — designed to sort companies relative to peers at each decision point. It answers: who should we prioritize?

Outcome Prediction (Walk-Forward CV). Separately, a walk-forward cross-validated model tests whether the ranking generalizes forward in time — trained on earlier cohorts, evaluated on later ones — using only information available at each stage. It answers: does the model actually predict who wins?

Three Decision Points

At Founding. Competition results, education trajectory, and sector signals available before institutional funding. Focuses sourcing on the highest-potential alumni.

At First Funding. Founding signals augmented by round size, investor quality, and co-founder composition. Primary signal for initial check sizing.

At Series A. Funding signals plus milestone progression — time to Series A, round scaling, traction markers. Drives follow-on concentration.

Feature Importance by Stage

Ranked by feature importance share within each stage. The top four named categories account for over 70% of model weight at every stage. At founding, competition results and sector lead; by Series A, capital progression and technical depth dominate — execution evidence rather than biographical priors.

Selection Advantage

The ranking model's cross-sectional test: within each stage cohort, does the score concentrate future winners into a smaller, actionable subset? Companies are grouped into quartiles by rank and measured at the $100M+, $500M+, and $1B+ thresholds.

At Founding

n=78 • Baseline $500M+: 9.0% • Top quartile: 30.0% (3.3×)

At First Funding

n=123 • Baseline $500M+: 18.7% • Top quartile: 32.3% (1.7×)

At Series A

n=79 • Baseline $500M+: 27.8% • Top quartile: 55.0% (2.0×)

Quartile Lift by Stage

$500M+ outcome rate by score quartile at each decision point.

AT FOUNDING

AT FIRST FUNDING

AT SERIES A

Precision/Recall Operating Curves by Stage

Cumulative top-k precision/recall curves for $500M+ outcomes.

AT FOUNDING

AT FIRST FUNDING

AT SERIES A

Predictive Validation

The predictive test: does the model generalize forward in time? Walk-forward cross-validation trains on earlier founding-year cohorts, evaluates on later ones, and uses only information available at each stage — no future data leaks into training.

An AUC of 0.84 at Series A means the model correctly ranks a random $100M+ outcome above a random non-outcome 84% of the time. The founding-stage AUC of 0.62 represents a modest early signal that strengthens as information accumulates through each subsequent stage.

Application to Fund I

The scoring system translates proprietary data on nearly 40,000 science fair alumni into stage-specific decision support. Each stage score provides a different lens on the ranked alumni pipeline.

Sourcing (Founding Score)

At founding — before any institutional capital — the ranked list shows $100M+ outcomes at 3.5× the baseline rate near the top. The founding score provides early signal across a pipeline of 50–100 active early-stage companies, helping prioritize outreach within a broad alumni universe.

Check Sizing (Funding Score)

At first funding, the model reaches a 0.57 walk-forward AUC, with top-quartile $500M+ rates of 32.3%. This is the fund's primary pre-commitment signal — the coverage tables above show how hit rate and recall trade off across different portfolio sizes at this stage.

Follow-On (Series A Score)

The strongest decision point: 0.84 AUC, with the coverage table showing hit rates from 55% in the top quartile down to 27.8% at baseline. Unicorn rates in the top quartile reach 50%. Signal is strongest here because the model incorporates milestone progression, round scaling, and traction markers alongside the founding and funding features.

Interpretation Notes

• Stage cohort sizes differ because each decision point uses its own eligible universe (78 at founding → 123 funded → 79 Series A+).
• Early signal is observable before institutional capital. The founding score shows top-quartile $100M+ rates at 3.5× the baseline rate, using only competition results, education trajectory, and sector signals.
• Signal strengthens with information. Top-quartile $100M+ lift ranges from 1.9–3.5× across stages, and CV AUC rises from 0.62 at founding to 0.84 at Series A. Each stage adds non-redundant predictive value.
• The score is a rank, not P($500M+). The V4 score is a percentile rank of stage-specific $100M+ output. $500M+ probabilities are separately calibrated and are not row-level monotonic with score.

Definitions