Funds & ETFsFlows, fees & vehicles
Preprint finds venture capital portfolios perform close to matched random benchmarks
A Crunchbase-based preprint finds that real venture capital portfolios perform about as well as random portfolios matched on timing, sector, region and size. It finds no edge among big winners, though it measures follow-on funding rather than exit returns.
From the Klement on Investing article of November 17, 2026 and cited sources: Knicker, Bouchaud and Benzaquen compared about 8,200 investors' Series A–C portfolios with 1,000 random draws each. [2] [3]
SOURCES[2] arXiv HTML full text, v1 arxiv.org[3] Do Venture Capitalists Beat Random Allocation?… arxiv.org
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Knicker, Bouchaud and Benzaquen compared about 8,200 investors' Series A–C portfolios with 1,000 random draws each. [2] [3]
- Original source
- Klement on Investing article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- Studies
- arXiv 2605.03980 (May 2026) [3] date from arXiv ID
- When
- November 17, 2026 article published
- Covered
- Nov 17 edition
- Evidence
- 6 sources across 5 sites
- 1Klement on Investing: Do venture capitalists beat a random…klementoninvesting.substack.com
- 2arXiv HTML full text, v1arxiv.org
- 3Do Venture Capitalists Beat Random Allocation?…arxiv.org
- 4Skill and Luck in Private Equity Performance (Stanford GSB)gsb.stanford.edu
- 5The Persistent Effect of Initial Success: Evidence from…corpgov.law.harvard.edu
- 6Institut Louis Bachelier news iteminstitutlouisbachelier.org
StrongHow it’s measured
Substack · 22 likes · 10 comments
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Knicker, Bouchaud and Benzaquen compared about 8,200 investors' Series A–C portfolios with 1,000 random draws each. The draws matched each investor's year, industry, region and number of deals. Actual outcomes came out very close to the random ones. Tests of the right tail, at multiples of 5 or more, found no outperformance. The only downside protection was weak: roughly 2 percentage points fewer zero-outcome portfolios, a result that depends on how zeros are read. [2] [3]
The study tests stock-picking within a market segment, not overall fund performance. It defines a company's multiple as its next funding round divided by its current one, with no valuations or exits. It also holds year, sector and region fixed, which removes the timing and deal-access channels that earlier research links to persistent VC performance. [2] [4] [5]
A widely read commentary says the paper shows VCs lose most by overpaying at entry. The paper cannot support that claim, because the authors state they have funding amounts but no valuation data. The commentary also describes exit multiples and a US-only sample, while the paper uses funding-round ratios and conditions on region. [1] [2]
The evidence covers deal selection within a given year, sector and region. If it holds, limited partners may weight market entry, vintage and deal access more than company-picking claims.
Read the full assessment
That is an inference, since fund-level cash returns were not tested.
Executive brief
The commentary's sharpest claim, that VCs lose most when they overpay at entry, does not appear in the paper it cites. The authors say they have no valuation data, only funding amounts (arXiv HTML). The core finding does hold up. Knicker, Bouchaud and Benzaquen built random portfolios that keep each investor's timing, sector, region and portfolio size. Real VC portfolio outcomes came out "remarkably close" to those random ones, with no edge in the high-multiple tail (arXiv abstract). This fits older work showing that VC returns are the hardest asset-class returns in which to tell skill from luck.
What changed and event timeline
Journal paper finds VC skill hardest to identify
Korteweg and Sorensen separate skill from luck. They find skilled private-equity firms beat others by 7–8% a year, but VC performance is the noisiest and its skilled managers the hardest to identify ().
Persistence linked to deal access
Nanda, Samila and Sorenson (Journal of Financial Economics) find that early IPO success predicts later success. About half of that persistence comes from investing in the right places and times, and early success seems to bring better deal flow ().
Random-allocation preprint posted
arXiv:2605.03980 v1 (8 pages, 10 figures) compares VC portfolios with constrained random benchmarks ().
Research institute promotes the result
Institut Louis Bachelier, where the authors' econophysics lab sits, summarises the paper as showing how hard it is to separate skill from luck when a few big wins drive returns ().
Commentary turns it into a polemic
Joachim Klement calls VC picks indistinguishable from "monkey" darts and says overpaying for trophy assets hurts returns ().
Capabilities and access
- Paper: arXiv:2605.03980v1, a preprint with no journal review noted on arXiv. Listed under econ.GN and cond-mat.stat-mech (arXiv).
- Data: Crunchbase funding rounds for startups founded 2010–2022, covering 26,522 startups and 8,249 investors.
- Authors: TU Munich and Institut Louis Bachelier. Bouchaud and Benzaquen are also affiliated with Capital Fund Management (arXiv HTML).
Read the full section
- Paper: arXiv:2605.03980v1, a preprint with no journal review noted on arXiv. Listed under econ.GN and cond-mat.stat-mech (arXiv).
- Data: Crunchbase funding rounds for startups founded 2010–2022, covering 26,522 startups and 8,249 investors. That is 14,636 Series A, 7,593 Series B and 4,293 Series C deals across six sectors (arXiv HTML).
- Authors: TU Munich and Institut Louis Bachelier. Bouchaud and Benzaquen are also affiliated with Capital Fund Management (arXiv HTML).
Technical analysis for researchers and developers
- Outcome measure: a company's multiple is the next round's funding amount divided by the current round's, not an exit return.
- Benchmark: for each investor, the same number of deals is drawn with replacement from the same year, industry cluster and region, repeated over 1,000 Monte Carlo draws.
- Tests: two-sample Kolmogorov–Smirnov tests restricted to each tail, plus z-scores comparing each portfolio's realised multiple with what its rank would predict under random sampling.
Read the full section
- Outcome measure: a company's multiple is the next round's funding amount divided by the current round's, not an exit return. Companies with no new round within three years count as 0. A portfolio's multiple is the equal-weighted mean of its companies (arXiv HTML).
- Benchmark: for each investor, the same number of deals is drawn with replacement from the same year, industry cluster and region, repeated over 1,000 Monte Carlo draws.
- Tests: two-sample Kolmogorov–Smirnov tests restricted to each tail, plus z-scores comparing each portfolio's realised multiple with what its rank would predict under random sampling.
- Reproducibility: the method is simple to replicate, but Crunchbase is a licensed dataset.
Claims and evidence
- No edge in the upside tail: right-tail tests for multiples of 5 or more keep high p-values (arXiv HTML).
- At most weak downside protection: about 2 percentage points fewer zero-outcome portfolios than random.
- Top portfolios are not exceptional: z-scores mostly fall within ±2, and the best-ranked portfolios score negative.
Read the full section
- No edge in the upside tail: right-tail tests for multiples of 5 or more keep high p-values (arXiv HTML).
- At most weak downside protection: about 2 percentage points fewer zero-outcome portfolios than random. Left-tail p-values fall below 0.05 only once the lower bulk of the distribution is included, and the result depends on how zeros are read (arXiv HTML).
- Top portfolios are not exceptional: z-scores mostly fall within ±2, and the best-ranked portfolios score negative.
- The "overpaying" claim is not in the paper: Klement describes exit money multiples, US funds and entry-multiple z-scores (Klement). The paper instead ranks realised portfolio multiples, uses funding-round ratios, and conditions on region rather than describing a US-only sample (arXiv HTML).
Context and prior work
Persistence studies find that VC performance repeats across funds, but the Korteweg–Sorensen decomposition shows VC is the noisiest segment of private equity (Stanford GSB). Nanda et al. link persistence partly to where and when firms invest, and to better access after early wins (Harvard Law Forum). The new paper deliberately holds year, sector and region fixed.
Read the full section
Persistence studies find that VC performance repeats across funds, but the Korteweg–Sorensen decomposition shows VC is the noisiest segment of private equity (Stanford GSB). Nanda et al. link persistence partly to where and when firms invest, and to better access after early wins (Harvard Law Forum). The new paper deliberately holds year, sector and region fixed. That removes the "right place, right time" channel, so it tests stock-picking within a given market segment, not overall fund returns.
Limitations, safety and contested findings
The authors list these limits (arXiv HTML): The Institut Louis Bachelier summary calls the paper peer-reviewed (ILB), but arXiv lists only a v1 preprint (arXiv). No independent replication has been published.
Read the full section
The authors list these limits (arXiv HTML):
- No valuations, ownership stakes or investor-level cash returns.
- Every deal is weighted equally, whatever its size.
- Failures are under-observed in Crunchbase, so bad outcomes may be undercounted.
- A zero can mean failure or a company that grew without needing more funding.
- The benchmark ignores access, reputation and network advantages.
- The results "should not be interpreted" as proof that individual managers lack skill.
The Institut Louis Bachelier summary calls the paper peer-reviewed (ILB), but arXiv lists only a v1 preprint (arXiv). No independent replication has been published.
Business and practitioner implications
- Limited partners: the evidence supports judging managers on which markets and vintage years they enter and on their deal access, not on company-picking claims within a segment.
- AI founders and funding: the study's outcome is whether a startup raises a next round, so it measures fundraising momentum more than value created.
- Researchers: the conditional-random benchmark can be reused. Re-running it with valuation and exit data would test the commentary's overpayment claim directly.
Read the full section
- Limited partners: the evidence supports judging managers on which markets and vintage years they enter and on their deal access, not on company-picking claims within a segment. Equal-weighted tests cannot capture how much a manager puts behind each deal.
- AI founders and funding: the study's outcome is whether a startup raises a next round, so it measures fundraising momentum more than value created.
- Researchers: the conditional-random benchmark can be reused. Re-running it with valuation and exit data would test the commentary's overpayment claim directly.