What is modelled. A fund puts equal stakes in N independent startups. Each one returns 0 with probability pf; otherwise its multiple X on the invested money follows a Pareto (power-law) distribution above a minimum xm, P(X ≥ x) = (xm/x)α, truncated at 1000x because markets are finite. The fund's gross multiple is the average of the N outcomes. Smaller α means a fatter tail; for α below 2 the untruncated variance is infinite, so the result is dominated by rare giants and the sample mean converges slowly.
Closed forms used. Truncated-Pareto mean E[X|survive] = α xmα (H1-α - xm1-α) / ((1-α)(1 - (xm/H)α)) with H = 1000; the fund mean is (1 - pf) times that. Draws use inverse-transform sampling, X = xm (1 - u (1 - (xm/H)α))-1/α. The median survivor is xm (1 - (1 - (xm/H)α)/2)-1/α.
Why it matters. The mean can be comfortably above 1x while the median fund returns less than the capital invested: the average is carried by outcomes that most funds never draw. Holding more startups narrows the spread of fund results (roughly as 1/√N when variance is finite) and makes the typical fund resemble the mean, but a fat tail slows that convergence a lot. The second chart shows concentration: often the single best startup returns most of the money.
Checks. (1) The simulated mean outcome against the closed-form mean, with a tolerance of five standard errors. (2) The closed-form mean against a numerical integral of x f(x) over a logarithmic grid. (3) The share of startups returning at least 10x against the formula. (4) The share of survivors below the analytic median is one half.
Not modelled. Time to exit and IRR, fees and carry, follow-on rounds with pro-rata, correlation between startups and with the market cycle, dilution, preferred-share terms, survivorship in published data, and any link to real companies or funds. Parameter values are illustrative: the claim that outcomes are heavy-tailed is widely discussed, but the numbers here are not calibrated to any dataset. This page is an educational probability model and not financial or investment advice.
The towers, flat discs and plinth were modelled and Cycles-baked in Blender; sampling, analytic checks and rendering run in Three.js r160.