Beyond Picking Winners: Correlation-Driven Tail Risk in Quant VC Portfolio Construction

The central claim: picking is necessary but not sufficient

Every venture capital firm obsesses over picking winners. Almost no one models how their picks correlate. This paper formalizes what happens when that gap is filled in. Hold individual deal success rates constant, then turn up the correlation between deals across the portfolio. Three things happen at the same time:

• Modest outcomes get rarer. The middle of the return distribution thins out.
• Extreme upside gets sharply more likely. The right tail thickens substantially.
• Expected return stays the same. The mean does not move; only the shape of the distribution does.

Translation, in plain language: picking is necessary but not sufficient. The shape of a venture portfolio's returns depends on how the deals move together, not only on which ones were picked. Two general partners with identical pick quality can produce wildly different return distributions if their portfolios have different correlation structures.

What the paper contributes to quant VC portfolio theory

The paper uses a Gaussian-copula construction to decouple deal-level success probability from portfolio-level dependence, then runs Monte Carlo simulations to trace what correlation alone does to the distribution of fund outcomes. Because the marginals (per-deal success rates) are held fixed, every change in the output distribution is attributable to correlation, not to screening quality.

The result reframes a long-running debate in venture capital. The dominant narrative says that legendary outcomes come from legendary picks. The data here suggests that some portion of legendary outcomes comes from correlation-induced tail amplification: concentrated portfolios in correlated themes (think AI infrastructure in 2023, crypto in 2021) generate fatter right tails than independence-based models would predict, even when the underlying pick quality is unchanged.

How the model works

Each deal is represented as a Bernoulli outcome with a deal-specific success probability. The dependence between deals is modeled through a Gaussian copula: the marginals stay where they are, while the joint distribution is parameterized by a correlation matrix that captures how shocks to common factors (sector, theme, vintage, geography) move outcomes together. The authors then sweep correlation from 0 (full independence) up to empirically realistic levels and observe how the full distribution of fund outcomes changes.

Why a copula. Gaussian copulas are the standard machinery for separating marginals from dependence in quantitative finance. They let the model preserve heterogeneous per-deal success probabilities (founders are not interchangeable) while making the dependence structure an explicit, tunable object. The same machinery underpins much of credit risk modeling, which is why co-author Hasan Ugur Koyluoglu (one of the original authors of the 1998 Koyluoglu-Hickman framework) appears on this paper too.

Why hold marginals constant. The methodological point of the paper is to isolate the contribution of correlation. Most existing venture portfolio analyses change pick quality and correlation at the same time and then attribute the result to one or the other based on intuition. The Gaussian-copula construction forces the math to be honest: any change in tail probability with marginals held fixed is correlation, by construction.

What the simulations show

Three findings stand out for quant VC portfolio construction.

Correlation thickens both tails, not just one. Increasing correlation while holding deal-level success probabilities constant produces heavier left tails (more zero-unicorn outcomes) and heavier right tails (more multi-hit clusters). The middle of the distribution (the modest, two- or three-unicorn outcomes that feel like a fine fund) gets rarer. This is consistent with the existing Portfolio Outliers finding on the left tail, and extends it explicitly to the right tail.

Concentrated thematic portfolios are correlation-amplified by construction. A 40-deal portfolio that is 100% AI infrastructure in 2023 is not just exposed to the AI thesis; it is exposed to a single common factor that affects every deal in the portfolio simultaneously. The same is true for crypto in 2021, climate-tech in earlier cohorts, or any other thematic concentration. The model shows that thematic concentration mechanically raises both the probability of zero-unicorn outcomes and the probability of multi-unicorn clusters, relative to a diversified portfolio with the same per-deal quality. Standard fund-level risk models that assume independence understate both effects.

Two GPs with the same pick quality can have very different return distributions. The paper's most consequential implication for LPs: fund-level performance comparisons that look only at average pick quality miss most of the variation. Two funds with identical per-deal probabilities and different correlation structures will produce visibly different empirical distributions of unicorn counts, and the correlation-induced difference can be large enough to dominate the picking-induced difference. LP-facing models that compare GPs only on a pick-quality metric are reading half the picture.

Why this matters for LPs and GPs

For LPs, the practical implication is that the standard mental model of fund-level risk needs an explicit correlation parameter. A fund pitch that says “our average per-deal success rate is 4%” is reporting a marginal. A fund pitch that says “our average per-deal success rate is 4% and our portfolio correlation is X” is reporting something actionable. The first lets the LP estimate expected outcomes; only the second lets the LP estimate the probability of zero unicorns or the probability of a multi-unicorn cluster.

For GPs running concentrated thematic strategies, the paper says something uncomfortable: your variance is higher than your independence-based dashboards suggest. Both the downside and the upside. A thematic fund that hits in a hot vintage will produce a multi-unicorn cluster that looks like genius; the same strategy in a cold vintage will produce a zero-cluster that looks like incompetence. Correlation explains a lot of the gap between those two outcomes, and the underlying pick quality may be the same in both cases.

For diversified GPs, the implication is the inverse: your tails are thinner than thematic peers, but so are your chances of producing a marquee multi-unicorn cluster. Correlation-aware portfolio construction is not about picking one strategy over the other. It is about pricing both strategies honestly.

Where this fits in Vela's quant VC research program

Vela treats venture capital as a quantitative discipline. Correlation-Driven Tail Risk sits in the Probabilistic Portfolio Construction thread, which now contains two complementary papers:

• Portfolio Outliers (Sakamoto, Koyluoglu, Alican, Ihlamur, arXiv:2602.07761). Latent-factor model that computes the full distribution of unicorn counts P(U = 0), P(U ≤ 1), P(U ≤ 2) for any portfolio composition, accounting for sector, geography, and founder-type correlation. Foundational paper of the thread.
• Beyond Picking Winners: Correlation-Driven Tail Risk (this paper, Liang, Koyluoglu, Alican, Ihlamur, arXiv:2604.23087). Gaussian-copula construction that isolates the contribution of inter-deal correlation to the full shape of the fund-level return distribution, holding pick quality constant.

The two papers operate at different levels of the same problem. Portfolio Outliers builds the latent-factor estimator that turns deal-level probabilities into a fund-level distribution. Correlation-Driven Tail Risk uses the copula machinery to explain, in clean form, what correlation is actually doing to that distribution. Read together, they describe both the machinery and the mechanism.

Both papers consume outputs from Vela's deal-screening research: Think-Reason-Learn (GPTree, Random Rule Forest, Reasoned Rule Mining, Policy Induction), and the LLM-Augmented ML thread (LLM-AR, GPT-HTree, Rare-event prediction, Verifiable Reasoning, Learning What to Ask). The deal-screening pipelines produce per-deal success probabilities; the portfolio construction papers consume them. Empirical benchmarking of the screening side is in VCBench.

Limitations

The paper is explicit about several limitations. The Gaussian-copula assumption underestimates joint extreme clustering relative to heavier-tailed alternatives such as Student-t copulas; the venture market arguably has heavier tails than the Gaussian benchmark, and the right tail estimates here are if anything conservative. The correlation parameter itself is treated as static across the simulation, whereas venture cycles alternate between systemic regimes (where correlation rises) and idiosyncratic regimes (where it falls). The deal universe abstracts away from sector, geography, and founder-type interactions; a future extension can layer the three-factor decomposition from Portfolio Outliers onto the copula construction here. Finally, the paper studies the mechanism of correlation in isolation; combining it with deal-level screening improvements requires the joint analysis that the two papers together start to enable.

Read the paper

Beyond Picking Winners: Correlation-Driven Tail Risk in Venture Capital Portfolio Construction.
Yunqi Liang, Hasan Ugur Koyluoglu, Fuat Alican, Yigit Ihlamur.
arXiv preprint arXiv:2604.23087, April 2026.
Read on arXiv · Download PDF.

For the foundational paper in the same research thread, see Portfolio Outliers. For the deal-screening pipelines that produce the per-deal probabilities both portfolio-construction papers consume, see Think-Reason-Learn and the LLM-Augmented ML thread. For benchmarking, see VCBench.

Authored by members of the Vela team and our collaborators at the University of Oxford and Oliver Wyman. See the full roster of contributors.

For research collaboration on quant VC portfolio construction, correlation modeling, copula methods for venture, or LP-facing fund risk analysis, email engage@vela.partners.