Diversify your angel portfolio

· 6 minute read

Venture outcomes don’t behave like normal investments. The distribution is so tail-dominated that every dataset you’ve ever seen underestimates the true expected return — it’s missing the 10,000x outcomes that haven’t happened yet (or weren’t captured). This breaks most portfolio advice you’ll read.

α ≈ 2where empirical VC tail indexes cluster, right at the point where the mean stops existing
37%average share of a 20-company portfolio's proceeds from its single best company
+2ppwhat going from 20 to 100 companies adds to your odds of a 5x portfolio
15-30companies: enough to survive, few enough to actually help
Headline numbers from the model and simulation below (power law, α = 2.0, 35% of deals returning at least 1x).

The math under the claim

Jerry Neumann’s Power Laws in Venture derives the whole return distribution from two numbers an investor actually picks: the growth rate g of the companies they back, and the average time to exit i. If value compounds at g and exits arrive exponentially with mean i, returns follow a power law with tail index

α = 1/(g·i) + 1

A typical VC with four-year holds and ~30% year-over-year growth lands at α ≈ 1.96. Neumann collects empirical estimates across a dozen datasets and they cluster near 2, which the model says is no accident: fund lifetimes cap the time to exit, and the supply of fast-growing companies caps the growth rate.

Your α is a strategy choice: it falls out of the growth rate you select for and how long you hold. The dot marks a typical VC (4-year holds, 30% YoY growth, α ≈ 1.95).

The α matters because everything about portfolio construction follows from it.

One company drives everything

The return of your portfolio will almost certainly be driven by a single investment. Not your “top 3” — one. And this never stops being true no matter how big your portfolio gets. It’s a property of the distribution: the expected largest outcome among n picks from a power law grows as n^(1/(α−1)). At α = 2 the biggest winner scales linearly with portfolio size. Invest in 10 companies and your likeliest best is ~10x; invest in 100 and it’s ~100x.

At α ≤ 2 (the empirical VC regime), the expected biggest winner grows at least linearly with portfolio size. Diversification raises the stakes of the winner instead of diluting it. Log scale.

As Neumann puts it: below α = 2, “one company is likely to return the entire amount invested in all of the successful companies.” The implication is that your job is to not miss that one company, not to optimize the average of everything else.

The cliff at α = 2

The mean multiple of a power law is (α−1)/(α−2). At α = 3 that’s a stable 2x. At α = 2.2 it’s 6x. At α = 2 it stops existing: the sample average never settles, however much data you collect. The empirical estimates put venture right at that edge, and Neumann argues supply and demand pins it there. Whenever fatter tails are available, more money shows up until α returns to 2.

The mean multiple explodes as α approaches 2 from above. Empirical VC alphas (shaded) sit at the edge of the singularity, where the average is barely defined.

One caveat. At portfolio sizes anyone actually holds, “the mean is infinite” is a modeling statement, not an observable. Real outcomes truncate somewhere, and with a few hundred exits you cannot statistically distinguish α = 1.9 from α = 2.2, or a power law from a lognormal (Clauset, Shalizi & Newman). The claim that survives is weaker but still decisive: the mean is so tail-dominated that your realized return is effectively one draw from the tail.

The dissent: batting averages

BIP Ventures argues the power law as practiced is a story Silicon Valley tells to excuse 90% loss rates. Top-quartile funds, they say, run deal batting averages of 30-35% (the best above 40%), and trading the 25x lottery ticket for a portfolio of 5-10x wins produces more consistent returns. (They’re a fund marketing their own strategy, and the batting-average figures come without a citation, but take the position seriously.)

Read through Neumann’s model, BIP isn’t denying the math. They’re choosing a different α. Later stage, capped magnitude, higher hit rate is what α ≈ 3 with a bounded tail looks like. It has a real mean, so diversification works there and results repeat. That’s a legitimate product. The question is which α you can sample from, and an early-stage angel doesn’t get to buy BIP’s distribution.

Run the argument yourself

The charts above are exact formulas. The probabilities below are simulated: 20,000 portfolios per size, per-company returns drawn from a Pareto distribution, portfolio return the equal-weight average. Three worlds: the orthodox α = 2.0 with 35% of deals returning at least 1x, a heavier-tailed α = 1.8, and a BIP-style α = 3.0 with a 60% hit rate and outcomes capped at 30x. Toggle between them and see which conclusions survive.

Downside protection saturates fast, but the odds of a great portfolio crawl. In the orthodox world, P(≥1x) hits ~87% at 20 companies and ~95% by 40, while P(≥3x) climbs from 21% at 20 companies to only 29% at 100. In the batting-average world the curves invert: safety is instant, and a 3x portfolio is out of reach at any size.
This is the chart the two camps disagree about. Under a power law the single best company's share of proceeds barely decays: ~37% of a 20-company portfolio, still ~30% at 100. In the batting-average world it collapses. Everything else follows from which line you believe.

Simple rules that follow

Numbers from the α = 2.0 simulation; the qualitative shape holds across the power-law scenarios.

More companies in your portfolio always reduces your chance of losing money, but it barely increases your chance of hitting a great return. Going from 20 to 100 companies buys you about 2 percentage points of probability on a 5x portfolio. That’s not worth the dilution of your attention.

You need ~40-50 companies to be 95% confident you won’t lose money overall. But ~20 already gets you to ~85%. The marginal safety of each additional company drops fast.

Your outcome is one company. At 20 companies the single best investment averages 37% of all proceeds, and that share barely moves as the portfolio grows.

Helping your companies likely shifts the distribution in your favor, and the model makes this precise: help that raises g lowers your personal α, which fattens your tail. If that’s true at all, then concentrated portfolios (15-30 companies) where you can actually be useful dominate over 100+ company spray-and-pray, because the tiny probability gain from diversification doesn’t offset the loss of involvement.

Certainty of great returns is essentially impossible. A 3x portfolio is plausible with 20 companies (~21% odds), but even 200 companies only gets you to ~34%. You have to be comfortable with the variance!

Diversify enough to not go to zero (~20-50 companies), concentrate enough to actually help (~15-30), and accept that your outcome is basically one company.

Sources and model notes

  • Power Laws in Venture (Jerry Neumann, 2015): the generative model, the empirical α estimates, and the x_max scaling. The estimates are a decade old; treat the exact values as dated and the structure as durable.
  • The Power Law of Venture Capital: Fact vs. Fiction (Mark Buffington, BIP Ventures, 2025): the batting-average position.
  • Power-law distributions in empirical data (Clauset, Shalizi and Newman, 2009): why claimed power laws often aren’t statistically distinguishable from alternatives.
  • Simulation: per-company multiples X = xmin · U^(−1/(α−1)) with xmin set so the hit rate P(X ≥ 1x) matches each scenario (35% for the power-law worlds, per Correlation Ventures’ ~65% money-losing deals; 60% for the batting world). Equal checks, no follow-on modeled, no fees, no time value.
angel investing