For over 80 years, America’s science fair system has been selecting unusually technical, self-directed young builders — teenagers who choose a hard problem, build evidence, and defend the work under scrutiny. From Westinghouse to Intel to Regeneron, the competitions form one of the longest-running talent-selection systems in the country.
Science Fair Fund is built around this community. We use a proprietary alumni dataset and stage-specific scoring to surface promising founders early, then combine that signal with conventional venture diligence to make pre-seed and seed investments. It’s a thesis about people — that the teenagers who chose to do something hard when no one was watching often become the founders who do it again when everything is on the line.
Cumulative, all years · verified funding threshold
45% of identified companies
Reach Series A+
131
Reach Series A+
Cumulative, all years · mature funded cohort rate: 70%
61% of all-years funded set
Source, score, diligence, and select
40
Target Initial Investments
Over the fund life · $50K–$100K initial checks
Portfolio target
The Opportunity
15.2%
Mature funded cohort · $1B+ peak / exit
70%
Mature funded cohort reaching Series A+
55,081+
Verified alumni tracked
29%
Mature funded cohort · $100M+ peak / exit
35
$1B+ alumni-founded companies · all years
-
Model AUC
-
Q1 Lift ($1B+) at Series A · backtest
-
Observed closed-company rate · mature cohort
How Scoring Drives Portfolio Construction
The model ranks alumni-founded companies at stage-specific decision points. At founding, it uses only signals knowable before institutional funding to prioritize sourcing and diligence.
At first funding, round and team information augment the early biographical signal. The score is decision support, not an automated investment decision or a calibrated company-level probability.
Initial checks remain within the $50K–$100K target range. Later scores, ownership, round terms, and company progress jointly inform where pro rata capital is most valuable.
The Structural Moat
Edge
What It Means
Community Flywheel
Participating in science fairs is a formative experience (age 14–18) — the kind that bonds people for life. That bond is the moat. 2,213 alumni said “willing to help” (70%+ response rate). The network compounds with every cycle: alumni open doors at their companies, write checks into each other’s rounds, and make the introductions that move companies forward. Active portfolio founders become the fund’s best source of referrals into the next generation of alumni startups.
Insider Information Asymmetry
ISEF 1st Place (’04) + multi-year Grand Awards Judge, paired with a proprietary dataset covering 25 years of ISEF alumni and STS records back to 1943 — 55,081+ verified profiles. A stage-specific model reaches walk-forward AUC 0.73 at first funding and 0.73 at First Funding, prioritizing the alumni most worth investigating as evidence accumulates. Recurring monitoring and community relationships can surface founding transitions, repeat founders, and warm referrals before those signals become broadly visible.
Observed Selection Edge
In the founding-stage walk-forward backtest (n=263), Q1 companies (n=66) reached $100M+ at a 33.3% rate, 2.0× the cohort baseline. Their $1B+ rate was 19.7% (2.3× baseline). This is evidence of ranking separation, not proof of realized fund returns.
The Manager: Anthony Atlas
I’m a Science Fair Winner. I won 1st at ISEF in 2004. That competition changed the trajectory of my life — and I’ve spent 20 years watching it do the same for others. I’ve been a Grand Awards Judge for multiple years and have deep relationships across the alumni community. This isn’t a thesis I researched — it’s a network I grew up in.
I Built the Database. Since 2006, I’ve built a proprietary dataset on this talent pool — 55,081+ verified profiles, company outcomes, and a 20-year reference cohort. It turns a community relationship into a repeatable sourcing and diligence system.
I Can Help Them Win. Raised >$75M in venture capital as an operator. Supported 3 deep-tech companies through ~10× valuation step-ups (seed → Series B). I know what early founders need because I’ve been in the room.
What the Data Shows
Metric
SFF Alumni
Denominator / Validation
Interpretation
$1B+ outcome rate
15.2%
23 / 151 funded companies
Peak, exit, or current value ≥$1B
Series A+ progression
70%
106 / 151 funded companies
Observed mature cohort
Closed-company rate
~7%
11 / 151 funded companies
Not equivalent to a realized write-off rate
Scoring AUC (First Funding)
0.73
Walk-forward CV · n=236
Ranking discrimination, not probability calibration
Reference cohort: alumni-founded VC-investable companies founded 2006–2020 that raised at least $500K. Outcome values use disclosed exits, current valuations, or documented peak private/public values; see Facts & Methodology for definitions and limitations.
Fund I focuses exclusively on ISEF and STS alumni. The scoring framework is designed to be extensible to analogous communities — Hertz Fellows, RSI, Math Olympiad winners, European science competitions — where formative competition selects for the same founder traits.
Get the Fund Deck
Full investment thesis, team background, and portfolio construction in a 15-page deck.
Current Pipeline
Early-stage alumni-founded companies with founding scores ≥45, founded 2021+. Anonymized for LP review.
Pipeline shows a point-in-time set of scoring candidates, not a statement of current offering terms or final portfolio construction.
—
Pipeline Companies
—
Target Initial Investments
—
Avg Tier $100M+ Rate
—
Industries
ISEFSTSBoth
Portfolio Outcome Projections
Using observed founding-stage tier rates from the 2006–2020 walk-forward backtest, applied to the current pipeline as a hypothetical portfolio. These are tier-level frequencies, not company-specific forecasts.
—
Expected $100M+
—
Expected Unicorns
Outcome Range (P25 – P75)
Methodology: Each company’s outcome probability is treated as an independent Bernoulli trial. Aggregate distributions computed via Poisson approximation. Range shows 25th–75th percentile; marker indicates median.
Current-strategy model. The default view models the $5M target case with 40 initial investments at $50K–$100K. Additional capital can be modeled as pro rata reserve for the highest-scoring companies. The return distribution remains scenario analysis, not a forecast.
Fund Model Controls
$—M
— first checks + — follow-on + 20% fees
Portfolio Construction (current target: 40 initial investments)
Q1 Companies ⓘ
10
Q2 Companies
30
Q1 Check Size
$100K
Q2 Check Size
$100K
Follow-on Reserve ⓘ
$0.0M
Signal Retention ⓘ
100%
Entry & Dilution Assumptions
Seed Post-Money ($M) ⓘ
$10M
1.00% initial ownership at a $100K check
Dilution per Round ⓘ
20%
Unrealized Positions
Valuation Mode
Liquidity Discount
40%
Scenario Options
Exit Value Cap
Largest-Outcome Scenario
OpenAI: $852B post-money as of 2026-03-31.
The default interpolates the upper tail with a 20%–60% private-market haircut.
Primary source.
— companies · $—M first checks · $—M follow-on · $—M fund
Gross MOIC Distribution
Log scale · final bin includes P99+ simulations
Scenario
Gross MOIC
Net TVPI
Unicorns
$100M+ Exits
Gross Proceeds
P25 (Conservative)
-
-
-
-
-
P50 (Median)
-
-
-
-
-
P75 (Upside)
-
-
-
-
-
Largest-Outcome Sensitivity
Model
P25 Gross
P50 Gross
P75 Gross
P50 Net
Evidence-weighted
-
-
-
-
OpenAI excluded
-
-
-
-
Observed $852B mark
-
-
-
-
Return Probabilities
≥5x
-
≥10x
-
≥25x
-
≥50x
-
≥100x
-
Link copied to clipboard!
Methodology: Smoothed Empirical Outcomes
Each simulated portfolio uses 127 observed alumni-founded company outcomes (funded, VC-investable, founded ≤ 2017). Outcome severity is sampled continuously by log-interpolating adjacent empirical quantiles rather than repeatedly drawing exact company values. The $100M+ rate for each simulated fund is drawn from a Jeffreys beta posterior around the observed founding-stage tier rate, so finite-sample uncertainty is propagated rather than treated as known.
Outcome Pool (127 companies, founded ≤ 2017)
Outcome Tier
Count
% of Pool
Example Values
Source
$1B+ (Unicorn)
21
16.5%
$1B – $852B
Database
$100M – $999M
16
12.6%
$100M – $700M
Database
$1M – $99M
10
7.9%
$5M – $75M
Database
No known valuation
80
63.0%
45 active (no public valuation) · 19 acquired (value unknown) · 8 closed · 7 unknown · 1 IPO micro-cap.
Tier Selection — SFF Scored Mode
Quartile
Scoring Range
$100M+ Rate
$1B+ Rate
Source
Tier 1 (Gold)
Founding Q1 (n=66)
33.3%
19.7%
Database
Tier 2 (Teal)
Founding Q2+Q3 (n=132)
11.4%
6.1%
Database
Founding-cohort baseline
All (n=263)
16.8%
8.8%
Database
Signal Retention — Out-of-Sample Sensitivity
Historical tier separation may not persist out of sample. Signal Retention controls how much of that observed separation the scenario preserves. It is a user-selected sensitivity, not an estimate derived from AUC.
At 100% (default), the scenario directly applies the observed walk-forward founding-stage tier rates.
At 25%, it preserves one quarter of the historical separation and draws the remainder at the founding-cohort base rate.
At 0%, scored tiers receive the same founding-cohort base rate.
Largest-Outcome Treatment
OpenAI is anchored to its company-announced $852B post-money financing value dated 2026-03-31. The default evidence-weighted scenario applies a randomly drawn 20%–60% private-market haircut when the top empirical interval is sampled. The robustness case removes that observation; the observed-upside case uses the announced mark. This preserves the observed right tail without turning one company into a repeatable point mass.
Fidelity
Interpretation
Effective T1 $100M+ Rate
100% (default)
Observed walk-forward tier rates applied directly
33.3%
25%
One quarter of historical tier separation retained
Flat sensitivity for each known round after entry (adjustable, 15%–30%); 25% is a downside case
Model
Avg rounds: unicorns
4.7
From actual company data
Database n=20
Follow-on (T1)
Super pro rata
Offsets 2 dilution rounds
Industry
Follow-on (T2)
Pro rata
Offsets 1 dilution round
Industry
Fund Economics (Fixed)
Parameter
Value
Notes
Source
Fee / expense load
20%
Conservative simplifying reserve; actual fees and expenses are governed by offering documents
Model
Carry
20%
Of profits above 1× return of capital
Industry
Scoring & Backtest
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 proprietary data and stage-aware models to prioritize alumni, while publishing each validation view only when its versioned snapshot supports that decision point.
The scoring system’s edge is early identification: at founding — before any institutional capital — the top quartile produces $100M+ outcomes at 2.00× the baseline rate. Later-stage decisions incorporate observed financing, ownership, lead quality, and milestone evidence rather than assuming the founding-stage result automatically transfers.
Primary metric: $500M+ outcomes (stable sample, directly relevant to fund returns). $100M+ and $1B+ shown for context.
Two evaluation frameworks: Selection Advantage (tier separation within each stage) and Model Validation (walk-forward CV + time-gated backtests).
Current published selection view: At founding, Q1 concentrates 24.2% $500M+ outcomes (2.10× lift).
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.
Outcome rows use the promoted reference-cohort counts and the funded benchmark denominator. Outcome-by-Series-A intersections are not inferred from aggregate counts.
192 of 266 companies (72.2%) are US/Canada-headquartered, consistent with the fund’s North American deployment focus.
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.
Signal Evolution by Stage
How signal categories shift in importance as companies mature — the model adapts its weighting at each decision point.
Signal Category
At Founding
At Funding
At Series A
Competition Record
●●●
●○○
●○○
Sector Fit
●●●
—
—
Capital Progression
●○○
●●●
●●●
Technical Depth
●●○
●○○
●○○
●●● = high ●●○ = medium ●○○ = low — = not available at this stage
Model 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.
Decision Point
Walk-Forward AUC
Fund Application
At Founding
0.65
Sourcing priority within alumni universe
At First Funding
0.73
Initial check sizing within the $50K–$100K target range
At Series A
0.70
Follow-on concentration decisions
An AUC of 0.70 at Series A means the model correctly ranks a random $100M+ outcome above a random non-outcome 70% of the time.
Precision/Recall at Q1 Threshold
Stage
Selection %
Precision ($500M+)
Recall ($500M+)
At Founding
25%
24.2%
53.3%
Selection Advantage
The ranking model’s cross-sectional test: within each stage cohort, does the score concentrate future winners into a smaller, actionable subset?
The current published score concentrates $500M+ outcomes toward the top of the ranked founding-stage list.
This snapshot publishes founding-stage validation only. Later-stage model charts will return when an explicitly versioned later-stage snapshot is supplied.
Published-stage scope matters. The current snapshot contains a founding-stage selection diagnostic; it does not publish later-stage quartile tables.
Early signal is observable before institutional capital. The founding score shows Q1 $100M+ rates at 2.00× the baseline rate.
The score is a rank, not P($500M+). The V4 score is a percentile rank. $500M+ probabilities are separately calibrated.
Definitions
$100M+
Exit value or last known valuation ≥$100M.
$500M+
Exit value or last known valuation ≥$500M.
$1B+ (Unicorn)
Reached ≥$1B via a realized liquidity event (IPO/M&A), a current valuation, or a priced institutional round within the fund’s ~10-year hold window. Excludes peaks that were never realizable (e.g. micro-float listings) or that occurred outside the hold window; companies that reached ≥$1B in a real round but later declined are counted and noted as peak.
Funded
Institutional capital raised ≥$500K.
Series A+
Venture stage at Series A or later, funded ≥$500K.
Walk-Forward CV
Trains on earlier cohorts, tests on later ones. No future data leaks into training.
AUC
Area Under ROC Curve. Probability a positive is ranked above a negative. 0.5 = random; 1.0 = perfect.
Precision
Fraction of selected companies that are actual $500M+ outcomes.
Recall
Fraction of all $500M+ outcomes captured by the selected set.
Lift
Outcome rate in a tier divided by the baseline rate for the full stage cohort.
SHAP
SHapley Additive exPlanations. A method that assigns each feature a contribution score for a given prediction, based on cooperative game theory.
Everything an investor (or AI) needs to evaluate the thesis.
1. The Thesis
Science fair alumni are an underfollowed founder talent pool. Every year, teenagers compete at ISEF (International Science and Engineering Fair) and STS (Science Talent Search), building original work and defending it under scrutiny. Those competitions create an unusually early record of technical depth, independent drive, and communication under pressure.
In SFF’s mature 2006–2020 cohort, 23 of 151 funded alumni-founded companies reached $1B+ (15.2%). Science Fair Fund is designed to turn that observed founder signal into a repeatable pre-seed and seed sourcing advantage.
The fund’s structural edge is threefold: (1) a proprietary database of 55,081+ verified alumni, (2) a stage-specific ranking model validated on historical outcomes, and (3) community access — the GP is an ISEF 1st Place Winner (’04) and multi-year judge.
2. America’s Farm System for Innovation
In 1942, Westinghouse Electric launched the Science Talent Search to find the nation’s most promising young scientists. It wasn’t philanthropy — it was talent infrastructure. The STS became known simply as “The Westinghouse,” a 57-year institution that functioned as an intellectual audit of America’s high schools. Seniors submitted original research papers judged on depth and rigor, often conducted in university labs. Alumni include Ray Kurzweil (1965) and Regeneron co-founder George Yancopoulos (1976). SFF’s database includes STS records back to 1943.
Meanwhile, the broader ecosystem of local and regional science fairs — organized by the same nonprofit, Science Service (now Society for Science) — coalesced into a national competition in 1950 and went international in 1958 as ISEF. Where the Westinghouse was the elite selection, ISEF was the farm system: a pyramid of ~400 affiliated regional fairs worldwide, with roughly 175,000 students competing each year for ~1,700 finalist spots. Projects are presented on display boards and defended in live oral interviews — selecting for the ability to build, communicate, and hold up under scrutiny.
After Sputnik, Congress passed the National Defense Education Act, and overnight these fairs shifted from hobby to national priority. The two programs served complementary functions: STS was the intellectual audit, ISEF was the public spectacle. Students often competed in regional fairs as underclassmen before submitting to STS as seniors. Westinghouse subsidized Science Service — the organization that administered both — effectively funding the pipeline that fed its own prestige competition. Together they created a talent system that has been quietly compounding for over 80 years.
The sponsorship passed from Westinghouse (1942–1998) to Intel (1998–2016/2019) to Regeneron (2016/2020–present) — and that last transition closed a remarkable loop. George Yancopoulos was himself a top Westinghouse STS winner in 1976. He has said the competition was a pivotal moment that gave him the confidence to pursue science. Forty years later, he brought the sponsorship home — the first time a single entity has held title sponsorship of the entire ecosystem.
SFF tracks alumni from both competitions. The fund’s thesis rests on the observation that these programs have consistently selected teenagers with the specific combination of technical depth, independent drive, and competitive resilience that compounds in venture-backed company building. Today this ecosystem produces ~2,000 finalists per year — teenagers who chose to spend a year solving a hard problem when no one was making them. SFF’s database tracks what happens to them next.
3. Database Methodology
The SFF Alumni Intelligence Engine contains 55,081+ verified alumni profiles spanning ISEF and STS competitions from 1950 to present. Data sources include official competition records, Society for Science archives, professional profile enrichment, and public company databases.
Each alumni record includes: competition history (year, placement, project), educational trajectory (undergraduate and graduate institutions), career path (current role, company, location), and founder status (linked to specific companies with founding dates and roles).
Professional profile enrichment covers approximately 12% of the full database; coverage is not uniform and is higher for identified founders. The database currently identifies 215 companies with at least $500K in verified funding. Portfolio value displays use documented exits, current valuations, or peak priced-round/public values and should not be read as current aggregate NAV.
Data quality is maintained through automated health checks, deduplication pipelines, and promoted metric snapshots. The public site and machine-readable fund export are built from the same promoted snapshot and canonical fund profile.
4. Scoring Model
The SFF scoring model is a stage-specific ensemble trained on the eligible alumni-founded company universe. It ranks companies for $100M+ outcomes using signals across competition results, education, company fundamentals, founder biography, and sector.
Walk-forward AUC is 0.73 at first funding (n=236) and reaches its best published value of 0.73 at First Funding (n=236). At founding (before any funding data is available), the model still achieves meaningful separation between tiers. Walk-forward cross-validation is used to prevent data leakage — the model is always trained on companies founded before the test period.
At founding, Q1 companies show a 33.3% $100M+ outcome rate versus the cohort baseline, a 2.0× lift in the current walk-forward backtest (Q1 n=66; total n=263). This is historical ranking separation, not a company-level probability or a guarantee of future performance.
Feature importance shifts across stages: competition and biographical signals matter earlier, while capital progression becomes more informative after funding. Investment decisions add diligence, ownership, round terms, and portfolio fit to the model’s ranking.
5. Backtest Methodology
The Monte Carlo simulation uses 127 observed company outcomes: 47 companies with known positive outcomes (ranging from $5M to $852B) and 80 companies assigned zero realized exit value. The zero group includes active companies without disclosed liquidity, acquisitions with undisclosed consideration, closed companies, and unknown outcomes; it is a conservative realized-value treatment, not a claim that all 80 failed.
Simulation methodology: continuous log interpolation between adjacent empirical outcome quantiles, with observed round counts retained for dilution. Each simulated fund also draws its tier-level $100M+ rates from Jeffreys beta posteriors around the walk-forward observations. The default scenario uses the current target construction of 40 initial investments at $50K–$100K, applies position-level dilution modeling, and calculates net fund returns including management fees and carry.
The default evidence-weighted case applies a 20%–60% private-market haircut to the upper interval anchored by OpenAI’s company-announced $852B post-money valuation dated 2026-03-31. The simulator shows that case beside a robustness case excluding the observation and an observed-upside case using the announced mark. It assumes a $10M entry post-money — 1.0% initial ownership for a $100K check — and a flat 20% dilution sensitivity per subsequent known round. High modeled returns are reported rather than compressed to a market-looking range; they may reflect genuine right-tail potential, model misspecification, selection effects, or valuation limitations. Follow-on sensitivity uses realized outcome rank as a proxy for later-stage rescoring, so reserve cases should be read as upper-bound allocation scenarios rather than forecasts.
The interactive simulator allows investors to adjust portfolio construction and downside assumptions and see their effect on return distributions in real time. Results are scenario analysis, not an underwritten forecast.
6. Definitions and Denominators
Metric
SFF Alumni
Denominator / Meaning
$1B+ outcome rate
15.2%
23 / 151 mature funded companies; peak, exit, or current value
Series A+ progression
70%
106 / 151 mature funded companies
Closed-company rate
~7%
11 / 151; not a realized write-off measure
$100M+ outcome rate
29%
44 / 151 mature funded companies
ML scoring AUC (first funding)
0.73
Walk-forward discrimination for $100M+ target; not probability calibration
Best published ML scoring AUC (First Funding)
0.73
Walk-forward discrimination for $100M+ target; not probability calibration
Reference cohort: alumni-founded VC-investable companies founded 2006–2020 that raised at least $500K. All-years inventory counts are shown separately and should not be mixed with mature-cohort rates.
7. Important Disclosures
For Accredited Investors Only. This material is provided for informational purposes only and does not constitute an offer to sell or a solicitation of an offer to buy any securities. Any such offer would be made only pursuant to a definitive offering memorandum and subscription agreement.
Past performance is not indicative of future results. The backtest results presented are based on historical data and Monte Carlo simulation. Actual fund returns may differ materially from simulated results. All investments involve risk, including the possible loss of principal.
The scoring model’s best published predictive accuracy (AUC 0.73 at First Funding) is measured on historical data and may not persist in future periods. Model performance may degrade as market conditions change.
Science Fair Fund is not registered as an investment adviser under the Investment Advisers Act of 1940. This communication is not investment advice. Prospective investors should consult their own legal, tax, and financial advisors before making any investment decision.
Notes
Scoring & Backtest metrics.
Reference cohort: 2006–2020 founding years, VC-investable alumni companies.
AUC = cross-validated area under ROC curve.
Lift = Q1 outcome rate ÷ baseline rate.
Source: Science Fair Fund scoring model, cross-validated on historical outcomes.
Portfolio statistics.
Funded = raised ≥$500K. Unicorns = peak valuation ≥$1B.
Peak / exit value = sum of documented peak priced-round or public values and disclosed exit values; it is not current aggregate NAV.
The inventory summary is regenerated from the canonical all-years company universe. The visible browser shows the top 50 companies by documented peak / exit value rather than every company in that universe.
Source: Science Fair Fund alumni database, all founding years.
Welcome!
Two quick questions to personalize your experience.