Retirement Monte Carlo

For a given household — starting portfolio, contribution schedule, horizon, withdrawal rule, allocation, and Social Security — the engine runs 10,000 independent simulated paths of the portfolio over the years to and through retirement. Each path uses real (CPI-deflated) monthly returns sampled from history; contributions and withdrawals happen on a monthly grid. The headline output is the share of paths where the portfolio survives the full horizon — that is the success probability.

A path's monthly returns are constructed by block bootstrap: we sample contiguous 12-month windows from the historical series and glue them end-to-end. This preserves regime persistence — clustered drawdowns, momentum, mean reversion — without committing to any parametric distribution. The 1973–74 bear, the 2000–02 dot-com bust, and the 2008 GFC all live as samplable blocks rather than as independent monthly draws.

The classic Trinity-study Monte Carlo treats monthly returns as independent log-normal draws calibrated to a long-run mean and standard deviation. That misses two structural features of real markets that drive retirement-portfolio failure:

• Volatility clustering. Bad months come in clusters; volatility is autocorrelated. A pure log-normal sim spreads variance evenly through time, understating tail risk.
• Sequence-of-returns concentration. Retirement failure is dominated by what happens in the first ~5 years of decumulation. Block bootstrap keeps those bad-start sequences intact in their natural shape.

The production dataset is 1149 months of real (CPI-deflated) total returns for stocks, bonds, and cash, covering 1928-01 through 2023-09:

• Stocks:Robert Shiller's monthly S&P series (the maintained CSV mirror of the Irrational Exuberancedata). Real total return is computed from Shiller's real price plus real trailing dividend accrued monthly.
• Bonds: a constant-maturity 10-year Treasury incomeproxy built from Shiller's long-rate column (yield / 12). Duration return — capital gain or loss from yield moves — is deliberately excluded to keep the bond series a stable income stream; that design choice is documented in scripts/plan/refresh-historical-returns.ts.
• Cash:FRED's TB3MS 3-month T-bill rate where available (1934 onward); earlier months approximate cash as 0.4 × the long rate to capture the typical long–short spread.
• Inflation: the CPI column of the same Shiller series. All three return series are deflated by realized monthly CPI before they are baked into the dataset.

The dataset ships with the app and carries its asOf date (2026-05-10) and a sha256 of the source CSV body (b76e7613e7ec…). It is refreshed via scripts/plan/refresh-historical-returns.ts; the window above is read from the dataset header at render time, so this page always states the true coverage. If the production data has not been loaded — for example, in a fresh dev environment — the engine falls through to a small calibrated synthetic fixture and brands the result datasetKind="fixture", which the plan surfaces label as illustrative.

• Real returns, real cashflows.Every dollar in inputs and outputs is in today's purchasing power. In the default inflationMode="real", contributions and withdrawals stay flat in real terms and CPI is implicit in the already-deflated dataset.
• Stochastic inflation (optional). inflationMode="stochastic" samples realized CPI alongside returns, deflating any nominal-dollar cashflows (a fixed nominal pension, a nominal income target) by cumulative realized inflation month by month.
• Monthly rebalance to target allocation. Every simulated month, the portfolio is rebalanced back to the stock / bond / cash weights you specified (including along a glide path, if one is set). This is standard for retirement-projection MC and reduces tracking error vs the chosen allocation.
• Tax wrappers. The portfolio can be split across taxable, traditional, Roth, and HSA pools. Decumulation draws in sequence — taxable first, then traditional (grossed up by the effective retirement tax rate, default 20%), then Roth, then HSA. Omitting the split reproduces the simple single-pool behavior with no tax adjustment.
• Withdrawal strategies. Four are wired: fixedReal (constant real-dollar income per year), fixedPctInitial (percentage of the start-of-retirement balance, locked), guytonKlinger (guardrail bands at 1.2× / 0.8× the initial rate with 10% adjustments, per the 2006 paper defaults), and vpw (variable percentage withdrawal, annuitizing the remaining balance at a 4% default expected return).
• Social Security. A single real annual benefit starting N years after retirement, held flat in real dollars. Claim-age adjustments, spousal benefits, and WEP/GPO interactions are computed by the federal benefits engine and passed in as that single real benefit.

• It does not model healthcare cost trajectories or long-term care.
• It does not model taxes on investment income year-by-year (the tax-wrapper gross-up is a flat effective rate, not a bracket model).
• It does not model future legal / regulatory changes (tax brackets, RMD ages, SS solvency).
• It does not model real-estate equity or business equity.
• Success probability is not a guarantee. A 95% success rate on the historical bootstrap means 5% of past-shaped futures would have failed; nothing in the math forbids a worse future.

Every run is deterministic given (inputs, dataset version, engine version, seed). The default seed is 0xC0FFEE; two runs with the same inputs against the same dataset and engine version produce identical results. Saved scenarios store their inputs verbatim and can be re-rendered exactly at any time.