Reference

Monte Carlo and historical replay

How the chance of success is computed, with sampling methods, seeds, what counts as success, failure analysis, and a comparison with FIRECalc.

The chance of success replays your plan many times, each time with a different sequence of market returns, dividends and inflation, and counts how often the plan succeeds. Every trial is a full run of the same engine used for the deterministic projection.

What varies and what does not

Only the market varies: stock and bond returns, dividends, inflation, and optionally exchange rates. Your income, spending, tax rules and dates are the plan’s, in every trial. Crypto has no long history, so historical runs draw it from a distribution.

Sampling methods

In the app Method How a trial is built
Historical, random start years historical-random-restart (the default) Starts in a random historical year and restarts at a random year when the data runs out.
Historical, every start year historical-sequential One trial per historical start year, with no looping. The number of trials is set by the data.
Block bootstrap block-bootstrap Stitches random blocks of consecutive historical years (block length 1 to 30 years, default 5).
Random (normal distributions) normal Independent draws from correlated normal distributions for stocks, bonds and inflation. Optional log-normal shape (returns cannot fall below minus 100 percent) and fat tails (a Student-t with 2.5 to 1,000 degrees of freedom, keeping the standard deviation).

Trials run from 1 to 10,000, default 1,000. A series switched off for variation uses the deterministic projection’s rates. Datasets are described on Historical data.

Seeds and reproducibility

The random number generator is seeded, and each trial derives its own seed from the run’s seed and the trial number. Two consequences:

  • The same seed gives the same answer. Splitting a run across worker threads and merging the slices gives exactly the result of a single-threaded run.
  • Through the API and MCP, a run without a seed uses a fixed default seed, so “same plan, change one number” compares like with like and repeat requests come from cache. Ask for a fresh seed with randomSeed: true on the API.

In the plan editor’s Chance of success section, “Random seed” is either “New each run” or “Fixed” with a value. A fresh seed is recorded in the result so a run can be repeated.

What counts as success

The default is every year funded: a trial succeeds if every year’s spending was fully funded. The engine supports four definitions:

Definition Success means
all-funded No unfunded shortfall in any year.
liquid-floor Liquid net worth never drops below a floor you set (today’s money).
never-depleted Liquid net worth stays above zero (spending may still have been cut).
end-floor The final value of liquid net worth or net worth is at least a floor (shortfalls along the way allowed).

The app’s “Success means” control offers “Every year funded” and “Savings above a floor”. Results also sort trials into outcome categories: large surplus, comfortable, barely made it, almost made it, failed in the middle and failed early.

What the results include

Percentile bands for every series (in today’s money and nominal), the success rate, the outcome categories, a histogram of the years in which failing trials first run short, milestone timing across trials, and sequence-of-returns buckets (success by the returns in the first years, default 10, adjustable from 1 to 40). It also keeps representative best, worst, median, 10th and 90th percentile trials so you can inspect one simulated future in full.

Exchange-rate risk

Off by default. When on, each foreign currency gets a yearly volatility and a correlation. Zero volatility reproduces the deterministic run.

How it was checked

The engine’s Monte Carlo was compared with FIRECalc on one case: a 60/40 portfolio of 1,000,000 dollars, 4 percent inflation-adjusted withdrawals, 30 years and 0.18 percent fees. FIRECalc, on US data from 1871, gave 96.8 percent: 4 of 126 cycles failed. Glideworth’s bundled US data starts in 1928, so it has 69 cycles. With cash-flow timing at start (FIRECalc’s convention) 4 of them fail, all mid-1960s starts: 94.2 percent. Counting FIRECalc’s 57 earlier cycles, which all succeed, as successes gives 96.8 percent. At mid (the default) the result is 95.7 percent and at end 98.6 percent; the spread shows how much the timing convention alone matters. It is one comparison, not a published validation report.

Last checked against the code on .