Probabilities are estimated from factors (base rate × capital density, tech maturity, comparable exit activity, founder/role fit) via a logistic model; payoffs are anchored on comparable exit valuations minus liquidation preferences and tax; the two paths are compared on both expected value and a risk-aversion-adjusted certainty equivalent (CRRA utility). All loaded numbers are a neutral example — overwrite them.
1 · Base rates (reference-class forecasting). Start from the empirical exit rate for the path's reference class, not a gut guess. Anchors: of US startups founded 2018 on Carta, ~49% shut down, ~5% were acquired, ~0.2% IPO'd; only ~6.9% of seed-stage startups exit for more than they raised; any-exit odds run ~18% (Series A) to ~28% (Series C); ~66% of software exits are M&A, and sub-$200M exits are M&A-dominated while IPOs need roughly $100M+ ARR, 40%+ growth and 120%+ net retention.
2 · Factor adjustment (Bayesian / logistic). logit(p) = logit(base rate) + w·Σ(factor scores), each factor scored −2…+2 vs the comparable baseline. Capital density (funding + acquirer appetite in the space), technology maturity (lower execution risk), comparable exit activity, and founder/role fit each shift the odds. Empirical anchor for founder fit: previously-successful founders succeed ~30% of the time vs ~18% for first-timers (Gompers, Kovner, Lerner & Scharfstein, JFE 2010).
3 · Heavy-tailed outcomes (power law) + liquidation waterfall. Success mass is split across a thin "big" tail and fatter "mid/small" outcomes, reflecting the power-law shape of venture results. Payoff = stake × comparable valuation − liquidation preferences, then tax — at low exits, preferences can take most of the proceeds before common stock, so the "small exit" branch is deliberately squeezed.
4 · Risk & wealth (expected-utility / CRRA). Instead of comparing raw averages, each path's gamble is scored with constant-relative-risk-aversion utility u(w)=w1−γ/(1−γ) over total wealth (buffer W₀ + floor + payoff), then converted to a certainty equivalent — the guaranteed sum you'd accept instead of the gamble. This is why a thin buffer plus a high chance of ≈$0 can make the high-EV path lose: the concave utility heavily penalizes near-ruin outcomes (von Neumann–Morgenstern; Arrow–Pratt).
Factor scores and base rates are your judgement calls — the models make the reasoning explicit and consistent, they do not remove the uncertainty. Empirical anchors are population averages, not your specific odds. This is illustrative modeling, not financial, tax, or legal advice; confirm equity %, dilution, liquidation preferences and QSBS eligibility against the real cap table and with a professional.