Two families of models. Every figure on this page comes from one of two kinds of return model, and knowing which is which answers most "what does this slider actually change?" questions:

• Data-driven models replay or resample actual recorded returns. In portfolio mode, Historical windows and Block bootstrap use your holdings' own monthly real returns (SIM-extended history, deflated, at your live weights). The Broad-sample column uses somebody else's data entirely: the Jorda-Schularick-Taylor academic record of 16 developed countries, 1870-2020, bundled into pofo, held as a 60/40 domestic stock/bond portfolio (the broad-sample SWR literature's baseline allocation). The μ/σ/df sliders have no effect on any of these.
• Parameter-driven models (Student-t, Sequence stress, Lost decade) generate synthetic returns from three numbers: the mean μ, the volatility σ and the tail thickness df. Those three numbers are what the sliders control, and everything in the "Market model" group works by changing them.

Where the slider values come from. In parametric mode they are just your inputs. In portfolio mode they are seeded from your portfolio's history: μ is the mean realised annual real return, σ the monthly dispersion scaled to annual, df fitted from the monthly kurtosis (fat months → low df). Then, because a 20-year history cannot reveal 45-year risks, the central models quietly blend those fitted values toward a cautious world-equity prior (μ 4.5%, σ 13%, df 4) in proportion to how much the horizon exceeds the history, capped at 50/50. So the portfolio's history reaches the central case through its statistics, never through its actual sequence of years.

The model columns, one by one:

Which model do the detail sections use? The one you click. The hero table is the selector: click any column and every detail section of the page (spending, lifecycle, the decisive decade, income, the risk levers, the buffer arbitrage, the spending-rule frontier) re-runs under that lens (the amber underline marks the active one; the fans of section 01 always show all four side by side). By default everything runs on the calibrated central case (the Student-t column).

How much should you trust each column? None of them is "the truth"; they bracket it:

• Historical windows / Block bootstrap are the most faithful to your assets but the least faithful to the future: they can only recombine the one window your funds lived through, usually a favourable one, and a 20-year history contains not a single independent 45-year retirement. Read them as the optimistic bound.
• Student-t is the planning default: your portfolio's fitted statistics, blended toward a cautious world prior, with fat tails. Its known blind spot is independence: real markets trend and cluster, so for long horizons it is mildly optimistic about sequence risk.
• Sequence stress fixes exactly that blind spot (clustered bears at the same average). If you only stress one thing, stress this.
• Broad-sample is the empirical century: everything that actually happened to a 60/40 investor across 16 countries, 1870-2020, including the disasters. It is the most honest single estimate of long-horizon risk for a passive allocation, and it is deliberately not about your specific portfolio.
• Lost decade is a tail scenario, not a central estimate: plan so that it is survivable, not so that its number is low.

Practical reading: plan between Student-t and Broad-sample; treat Sequence stress as the "am I robust to ordering?" check and Lost decade as the stress test.

What each toggle really overrides:

• broad-sample prior just rewrites the three sliders with the cautious prior values. It changes the parameter-driven models and every detail section; it never touches the data-driven columns (they don't read sliders).
• anchor to CAPE replaces only the central mean with the valuation-implied return (1/CAPE, plus the volatility-drag correction). σ and df keep their fitted values; data-driven columns unchanged.
• rising-equity glidepath reshapes the central model's asset mix over time (30% equity gliding to 75%), trading return for early-years protection.
• monthly withdrawals refines the kernel's granularity and is visible only in sections 07-08; the model strip and analysis sections always compare annual kernels, for speed and comparability.

The withdrawal rules, one by one (every rule answers the same question (how much to take out this year) and they all sit on one trade-off: certainty of income versus certainty of never running out). Exactly one runs at a time: picking a rule in the rail clears the one you had, since the kernel could only ever apply a single policy.

• Fixed real (the 4%-rule shape): take the same inflation-adjusted amount every year, whatever happens. Maximum lifestyle certainty, all the risk lands on ruin. The benchmark everything else is measured against.
• Flex cut: fixed real, but cut a bounded slice (say 10-20%) while the portfolio is in a deep drawdown, restore it after. The mildest adaptive rule: a known, capped sacrifice in bad years roughly halves ruin.
• Guyton-Klinger guardrails: ±10% spending moves whenever the current withdrawal rate leaves a ±20% band around the initial one. Powerful but unbounded downward: in a persistent bear the cuts compound (60k → 54k → 49k → …), which is exactly where its spectacular ruin reduction comes from. The floor slider bounds that descent at your incompressible standard.
• Ratchet: only-up moves, raise the standard when wealth is well above start, never cut. Prices the lifestyle option; slightly raises ruin.
• VPW (percent of portfolio): spend a fixed share of whatever remains. Mathematically cannot run out, but income swings one-for-one with the market.
• Bounded % (Vanguard dynamic spending): VPW smoothed, the yearly change in real spending is capped at +5%/-2.5%. Livable income path; some ruin risk returns because the descent is capped.
• ABW / TPAW (amortization): each year, the actuarial payment that exhausts current wealth exactly over the remaining horizon at the expected return, a reverse mortgage re-quoted yearly. Never ruins early, never strands wealth, adjusts in small continuous steps; income tracks the market. The rule much of the recent literature converges on; pairs naturally with an annuity floor for the incompressible part.

Mechanics worth knowing:

• The cash buffer is carved out of the starting capital (buffer = years × spending, the growth sleeve gets the rest), not added on top. Displayed wealth is always growth + buffer.
• Tail df in plain words: at the same volatility, it sets how much more often extreme years happen than a bell curve allows. At df 5, a catastrophic year (-30% real, a 3-sigma event) is roughly ten times more likely than at df 30 (≈ normal); ordinary years barely change.
• The red lines on the wealth fans are example failed futures. The eight thin sample paths are picked at evenly spaced ranks of final wealth (worst, ~14th percentile, ~29th, … best), so the count of red ones tracks the failure rate: one red line means "the worst future fails", two mean roughly "one future in seven fails", three "one in four", and so on. The bold red baseline is simply zero.
• The fans' vertical axis is capped at 10× the starting capital ("upside clipped ↑"): the upper cone can compound far beyond, but showing it would crush the zero line, which is the part that matters.
• Safe spend is always solved on the plain fixed rule at your acceptable ruin (the conventional definition of a safe withdrawal rate), even when your plan uses flex or guardrails; Ruin and Median wealth are evaluated under your actual policy.
• Taxes gross up every sale: withdrawing 60 k€ net sells more than 60 k€ of assets, and an annuity premium is a sale like any other, so raising it costs its capital-gains tax before a euro of income is bought. Only the gain share is taxable, and the embedded-gain slider says how large that share is on day one (half the capital by default); the basis is then tracked as you spend, so the effective burden climbs toward the full rate. Naming a PEA or an assurance-vie under the envelopes fold replaces the single blended rate with ordered pockets, drained taxable account first.
• The annuity is priced where mortality is drawn (section 05) and nowhere else: a joint-life, inflation-linked income at a 1% real rate, on an annuitant table, less the insurer margin you set. Under the fixed horizon of every other section, a lifelong income is paid for longer than it was priced for, which is a free lunch rather than a plan.

Le FIRE tranquille

Reference points: the classic US 4%/30y backtest ≈ 95% success; broad century-long developed-market samples give ≈ 75-85% for the same fixed 4% rule (Anarkulova, Cederburg & O'Doherty), with a ~2.3% safe rate at 5% ruin. These models are i.i.d. (no mean reversion), so long FIRE horizons read a little stricter than history.

These results are model-based and tend to be more optimistic than the empirical safe-withdrawal evidence. The main reasons, in rough order of impact:

• Dynamic spending. The "spending cut in downturns" (flex) reduces withdrawals in bad markets, which roughly halves ruin vs a true fixed rule. It defaults to 0 (the canonical fixed rule); section 04 shows what enabling it costs in lived spending.
• Optimistic return inputs. A μ/σ fitted from a fund's own history reflects a favourable window. The Broad-sample and Lost-decade columns supply the deep-past counterpoint; the central case blends toward the broad-sample prior when the fund history is shorter than the horizon.
• I.i.d. draws miss the persistence of real returns; the Sequence-stress column clusters bad years at the same long-run mean.
• Taxes. Every sale is grossed up at your rate on the gain share only, so the effective burden starts at the embedded gain share of your rate and drifts up as unrealised gains compound. Two figures set it: the rate (your blended rate across accounts, gain-weighted) and the embedded gain, half the capital by default. A book made entirely of cost basis, which this page used to assume because the control did not exist, flatters the sustainable withdrawal rate by about 0.3 point.
• Mortality. The ruin figures ignore death by default; section 05 shows the mortality-weighted picture.

For a life plan, read the pessimistic columns, keep the horizon past your life expectancy, and treat the ruin figures as ordinal (compare scenarios, don't trust decimals).