Monte Carlo Retirement Simulation: What It Is and What It Gets Wrong

A Monte Carlo retirement simulation runs your plan through thousands of possible futures instead of one straight-line average. Here's what that actually means, and the three different engines behind it that most calculators never disclose.

A Monte Carlo retirement simulation runs your plan through many different possible futures instead of one straight-line average, then reports how often it worked. That’s the whole idea. Instead of growing your portfolio at a flat 7% forever and handing you a single ending balance, it plays your plan out hundreds or thousands of times, each time with a different sequence of ups and downs, and counts the share of those runs where you didn’t run out of money.

The name comes from the casino. Any single run tells you almost nothing - one lucky sequence, one unlucky one. But run the same plan thousands of times and the distribution of outcomes starts to mean something: a success rate, a range of ending balances, a sense of how much the result depends on luck versus your inputs. That distribution is the honest answer a single average line can’t give you.

Here’s the part most calculators never tell you: “Monte Carlo” is not one method. There are at least three genuinely different engines that all get labeled the same way, and they make very different assumptions about the future. A tool that says “we use Monte Carlo” without saying which kind is hiding the most important decision it made. This article walks through all three, plainly, including where each one is weakest.

Why one average line lies to you

Compound one fixed return and you get a smooth curve that no real portfolio has ever traced. Real markets arrive as a jagged sequence, and the order matters enormously - a rough stretch early in retirement drains a portfolio you’re also withdrawing from, leaving less to recover with. Two plans with the identical long-run average return can end decades apart depending purely on when the bad years landed. (That’s sequence-of-returns risk, and it has its own guide here - this article is about the machinery that measures it, not the risk itself.)

A Monte Carlo simulation exists to surface exactly that. By running many sequences, it lets the unlucky orderings actually fail, so the success rate reflects them. The question is where those sequences come from. That’s the choice that splits the three methods.

The three engines

The three approaches most tools lump together as "Monte Carlo." Each is a legitimate way to model uncertainty; each is wrong in a different, specific way. Knowing which one produced a number is half of reading it honestly.

1. Parametric: draws from a distribution you define

The parametric approach doesn’t touch historical data at all. You give it a mean return and a measure of spread (and, if you have multiple assets, how they move together), and it generates each year by drawing a random number from that distribution. Do that for every year of every run and you get thousands of synthetic futures shaped by the statistics you chose.

The strength is flexibility. Parametric is the only one of the three that can model a future the past never contained. Worried the next thirty years deliver lower returns than history? Dial the mean down and see what it does to your plan. You’re not trapped inside what happened to already happen.

The weakness is that it’s only as honest as the assumptions you feed it, and clean distributions flatter reality. A plain bell curve treats a once-in-a-generation crash as far more improbable than markets have actually delivered - real returns have fatter tails and cluster their disasters, and a tidy normal curve smooths both away. (This site’s engine draws from a lognormal shape rather than a raw normal one, so a single year can never lose more than 100%, but the broader caution stands: a distribution that’s too clean will quietly understate how bad a bad stretch can get.) Parametric answers “what if the future looks like my assumptions” - which is powerful only if you’re honest that the assumptions are doing all the work.

2. Historical block bootstrap: real history, reshuffled in chunks

The bootstrap builds new sequences out of real past returns. But the crucial detail - the one that separates a good implementation from a broken one - is that it resamples multi-year blocks, not single years.

Here’s why that matters. If you shuffle history one year at a time and draw them independently, you destroy the market’s memory. In reality a bad year is more likely to sit inside a bad stretch, and bull runs cluster too. Draw years independently and you scatter those clusters into statistical noise, which quietly averages away the exact sequence-of-returns risk the simulation was built to measure. You’d get a cheerful, wrong answer.

A block bootstrap instead grabs a run of consecutive real years - say five in a row - and keeps them intact, then grabs another block, and stitches them into a new sequence. A crash keeps the recovery, or the lack of one, that actually followed it. The clustering survives. You still get thousands of distinct trials, so the success rate is statistically meaningful, and each trial still respects how markets really behave year to year. This is the method this site’s simulator runs by default, with blocks kept in a modest multi-year range and every asset drawn from the same calendar year at once, so a stock crash and its matching bond move stay aligned.

Its honest limit: a bootstrap can only recombine what already happened. It can’t produce a decade the historical record never contained, and it inherits the record’s own biases - the long US series in particular describes an unusually lucky century.

3. Cycle mode: the literal historical windows

Cycle mode is the most conservative and the most transparent: no random draws at all. It takes your plan and replays it once starting from every real year on record - your plan beginning in 1929, then 1930, then 1931, and so on - and reports how each of those actual histories would have treated you. No distribution, no resampling, no model assumptions. Just: here is what would have happened, run after run, if you’d started in each real year.

The appeal is obvious. There’s nothing to argue with about the inputs, because there are no inputs beyond the historical record itself. Some people trust it precisely because it invents nothing.

The catch is subtler than it looks, and it’s the honest thing to say out loud: a long historical record produces far fewer truly independent trials than its raw count suggests. A run that starts in 1965 and a run that starts in 1966 share almost their entire path - they’re the same decades offset by twelve months, not two independent draws. Once you need multi-decade windows, a century of monthly data yields only a handful of genuinely non-overlapping periods. Cycle mode gives you the real windows with zero modeling assumptions, but it can’t give you many independent ones, so a comforting-looking spread can rest on very little that’s actually distinct.

How to read a Monte Carlo result honestly

None of the three is the “correct” one, and a tool that hides which it used is asking you to trust a number without its footnotes. A useful way to hold them together:

• Parametric answers “what if the future looks like my assumptions.” Powerful for stress-testing a future unlike the past, honest only if you admit the assumptions carry the whole result.
• Historical block bootstrap answers “what if the future rhymes with the past, crashes and clustering included.” The default here, because it keeps sequence risk real while still giving thousands of trials - but it can’t imagine a genuinely new disaster.
• Cycle mode answers “what literally happened, start year by start year.” Zero modeling assumptions, but only a handful of independent windows underneath the count.

A success rate is only as meaningful as the engine behind it and the return assumption fed in. That’s why this site shows you the deterministic straight-line answer first, then the historical probability band around it, and lets you dial the expected return below the historical mean rather than treating an unusually lucky century as a promise. The point of running the simulation instead of a spreadsheet isn’t a prettier number - it’s seeing the shape of what you don’t control before you lean on it. If you want the real historical survival record next, here’s what withdrawal rates actually did across every 30-year window.

Frequently asked