Portfolio and risk · Long-run planning
Monte Carlo simulation: testing a plan against thousands of futures
How Monte Carlo simulation works in financial planning, with 10,000 simulated retirements, what a success rate means and how much it depends on the inputs.
On this page
A Monte Carlo simulation answers a question by trying it many times with random inputs. In investing, the question is usually "will this plan work?", and the random input is the sequence of future returns. The computer draws a possible return for each year from an assumed distribution, runs the plan through that sequence, and records the result. Then it does the same thing thousands more times. The spread of results shows the range of outcomes the assumptions imply, and the share of runs in which the money lasted is reported as a probability of success.
The method is the standard tool of retirement-planning software. It is useful because it replaces a single projected line with a distribution. It is also easy to misread, because the output looks like a measured probability and is really a restatement of the assumptions that went in. This lesson runs a simulation and then changes the assumptions to see how much of the answer they control.
The idea
A projection that assumes "5% a year" draws one line. Real returns do not arrive as an average. They arrive as a sequence, and Sequence of returns risk: why the order of returns matters shows that the sequence decides the outcome when money is being withdrawn.
History offers only one sequence. Back-testing reuses it in overlapping slices. A Monte Carlo simulation takes a different route: it treats each year's return as a random draw, manufactures as many sequences as wanted, and looks at how the plan fares across all of them.
The figure is the result: not a line but a fan. After 30 years the outcomes range from nothing to several times the starting balance, all from the same plan and the same assumptions.
How it works
A simulation has four parts.
- A model of returns. Typically an average return and a volatility for the portfolio, and a distribution to draw from. More elaborate models add several asset classes with correlations, changing inflation or fat-tailed distributions.
- A plan. The starting balance, the contributions or withdrawals, the time horizon.
- Many trials. For each trial, draw a return for every year, apply the plan and record the path.
- A summary. Percentiles of the balance at each date, and the share of trials that met the goal.
A worked example: 10,000 simulated retirements
The assumptions: a $1,000,000 portfolio; a fixed withdrawal at the start of each year, in today's dollars; 30 years; yearly real returns drawn independently from a lognormal distribution with a mean of 4% and a volatility of 11%, roughly what a balanced portfolio might be assumed to deliver after inflation; no taxes or fees. The random-number generator is seeded, so the figures below are reproduced exactly on every run.
With a withdrawal of $40,000 a year (4%):
| Year | 5th percentile | 25th | Median | 75th | 95th |
|---|---|---|---|---|---|
| 5 | $625,871 | $807,161 | $966,062 | $1,147,476 | $1,468,125 |
| 10 | $442,916 | $689,663 | $913,084 | $1,208,159 | $1,768,461 |
| 20 | $91,020 | $426,661 | $784,080 | $1,299,414 | $2,457,829 |
| 30 | $0 | $100,201 | $592,169 | $1,362,569 | $3,294,441 |
The plan succeeded, meaning money remained after 30 years, in 80.4% of the 10,000 runs. In the 1,958 runs that failed, the median year of running out was the 26th and the earliest was the 12th. At the other end, 34.3% of runs finished with more than the starting $1,000,000 and 14.4% with more than $2,000,000.
The simulation also reproduces sequence risk. Among the fifth of runs with the lowest balance after five years, 44.4% succeeded. Among the fifth with the highest, 98.8% did. The first five years largely settled the next twenty-five.
The answer depends on the inputs
| Withdrawal rate | 4% return, 11% volatility | 3% return, 11% volatility | 4% return, 15% volatility |
|---|---|---|---|
| 3.0% | 95.9% | 90.8% | 87.2% |
| 3.5% | 90.1% | 80.3% | 77.8% |
| 4.0% | 80.4% | 66.2% | 67.0% |
| 4.5% | 67.3% | 51.5% | 56.0% |
| 5.0% | 53.7% | 37.5% | 45.5% |
| 6.0% | 29.0% | 17.3% | 27.9% |
| 7.0% | 13.0% | 6.6% | 15.9% |
A change of one percentage point in the assumed return moved the success rate of a 4% withdrawal from 80.4% to 66.2%. Nobody knows the next 30 years' real return to within a point. The Equity risk premium (ERP): the extra return stocks earn for their risk lesson shows how wide the honest range is. So the difference between "80% likely to succeed" and "66% likely to succeed" is inside the uncertainty of the inputs. The second decimal of a success rate carries no information, and the first digit is debatable.
Half a point of withdrawal rate had a similar effect: from 4.0% to 4.5%, success fell from 80.4% to 67.3%.
How many trials?
Randomness adds its own noise. Eight separate runs of only 500 trials each, on the base assumptions, gave success rates between 79.8% and 81.8%. With 10,000 trials the figure is stable to a fraction of a point. Sampling noise is easy to remove by running more trials. The uncertainty in the assumptions cannot be removed that way: a million trials of the wrong model give a very precise wrong answer.
Using it
- Read the whole distribution. The 5th and 25th percentile paths say more about risk than the success rate. A plan with a median ending balance of $592,169 and a 25th percentile of $100,201 has less margin than "80% success" suggests.
- Change the inputs and watch the answer. A plan that only works with a 5% real return is fragile. One that still works at 3% is robust.
- Treat failure as a prompt to adjust, not a prediction of ruin. The simulated retirees never change their spending. A real person who was down 40% after five years would cut back. A 20% "failure" rate is closer to a 20% chance of needing to adjust.
- Do not aim for 100%. A plan with a 99% success rate under conservative assumptions is one in which the typical outcome is dying with most of the money unspent.
- Ask what the model assumes. The return and volatility used, whether inflation varies, whether returns have fat tails, whether fees and taxes are included.
- Compare with history. Safe withdrawal rates: how much can a portfolio pay out? runs a real sequence. A simulation and a back-test that disagree widely are a reason to examine both.
Where it fails or misleads
- Garbage in, garbage out. The output is the input assumptions, rearranged. The method adds no knowledge about future returns.
- Independent draws miss how markets behave. Simple models assume each year is unrelated to the last. Real markets show volatility that clusters, long stretches of high or low valuations, and inflation that persists. Whether this makes long-run risk higher or lower than the simple model implies is disputed.
- Normal-style distributions understate extremes. A lognormal draw almost never produces a year like 2008 or a day like October 1987. See Fat tails: why extreme market moves are more common than models say.
- Success rates hide magnitude. Running out in year 29 and running out in year 12 both count as failures. Ending with $1 and ending with $3 million both count as successes.
- Precision is mistaken for accuracy. "83.7% probability of success" sounds measured. It is a count of simulated paths under one set of guesses.
- Costs are often left out. A 1% annual fee lowers the return assumption by a point, which this example shows is enough to move success by fourteen points.
Compared with its neighbours
| Method | Source of return sequences | Strength | Weakness |
|---|---|---|---|
| Deterministic projection | One assumed average return | Simple | Ignores sequence and variation entirely |
| Historical back-test | The actual past, in rolling periods | Real crises, real correlations | One history; periods overlap |
| Monte Carlo simulation | Random draws from an assumed model | Unlimited scenarios; full distribution | Only as good as the model |
| Bootstrap | Random draws from actual past returns | Keeps real fat tails | Loses the order and persistence of history |
| Stress test | A named scenario | Concrete and specific | No probabilities |
How Finabase does it
Finabase does not run Monte Carlo simulations, estimate a probability of success or project a portfolio's future value.
What it offers is the other approach, real history. The Backtest page applies a trading rule to a symbol's actual past prices and reports the gain, the win rate and the trades, one historical path with no random draws. A symbol's chart at the maximum range and its History tab give the record from which return and volatility assumptions are usually taken, and the paper-trading account lets a plan be followed forward with simulated money as real prices arrive. Anyone using a simulator elsewhere can check its return and volatility assumptions against those histories.
Frequently asked questions
What does a Monte Carlo success rate of 80% mean?
That in 80% of the simulated futures, generated from the model's assumptions, the plan did not run out of money. It is not a measured 80% chance for the real world, because the real world may not match the assumptions.
Why is it called Monte Carlo?
After the casino district in Monaco. The method was developed in the 1940s by scientists including Stanislaw Ulam, John von Neumann and Nicholas Metropolis for physics problems that could be solved by repeated random sampling.
What success rate should I aim for?
There is no agreed figure. Planners often work with 75% to 90% for plans that allow spending to be adjusted. Very high targets force very low spending.
Is Monte Carlo better than using historical returns?
Neither is better outright. History contains real crises but only one sequence. Simulation gives any number of sequences, all from a simplified model. They are best used together.
How many simulations are needed?
A few thousand are enough to make the sampling noise small. In the example, runs of 500 trials varied by about two percentage points and 10,000 trials by a fraction of one.
Summary
Monte Carlo simulation runs a plan through thousands of randomly generated return sequences and reports the range of results. In a simulation of 10,000 thirty-year retirements, a 4% withdrawal succeeded 80.4% of the time under a 4% real return and 11% volatility, and about 66% of the time when either input was made slightly less favourable. The fan of outcomes is the valuable part: it shows how wide the possibilities are. The success rate is a summary of the assumptions, to be tested by changing them and never read to the decimal.
Check your understanding
5 questions on this lesson. Each answer comes with the reason, and nothing is scored.
What does a Monte Carlo simulation of a retirement plan actually do?
- Predicts the single most likely future
- Runs the plan through many randomly generated return sequences drawn from assumed inputs
- Replays the last 30 years of market history
Show the answer
B. Runs the plan through many randomly generated return sequences drawn from assumed inputs Each trial draws a random return for every year from the assumed distribution and applies the plan. The collection of results forms a distribution.
In the example, lowering the assumed real return from 4% to 3% changed the success rate of a 4% withdrawal from 80.4% to 66.2%. What is the main lesson?
- The simulation contains a programming error
- The success rate is highly sensitive to assumptions that cannot be known precisely
- A 3% return is impossible
Show the answer
B. The success rate is highly sensitive to assumptions that cannot be known precisely Future returns are uncertain by more than a point, so the success rate is uncertain by many points. It restates the inputs and cannot be read precisely.
Eight runs of 500 trials gave success rates from 79.8% to 81.8%. What causes that variation, and how is it reduced?
- Sampling noise from the random draws; reduced by running more trials
- Changes in the assumptions; reduced by using a lower return
- Inflation; reduced by using real returns
Show the answer
A. Sampling noise from the random draws; reduced by running more trials Each run uses different random numbers. More trials shrink this noise. They do nothing about uncertainty in the assumptions themselves.
Two simulated retirees are both counted as failures: one ran out in year 12 and one in year 29. What weakness of the success rate does this show?
- It ignores how severe or how late the failure is
- It counts too few trials
- It ignores inflation
Show the answer
A. It ignores how severe or how late the failure is A success rate is a count of pass or fail. It hides the size and timing of shortfalls, which is why the percentile paths should be read as well.
Why does a simple simulation with independent, lognormal yearly returns tend to understate the chance of extreme years?
- Because it uses too many trials
- Because real returns have fatter tails and clustered volatility than that distribution allows
- Because it ignores dividends
Show the answer
B. Because real returns have fatter tails and clustered volatility than that distribution allows Actual markets produce very large moves more often than a normal-style distribution predicts, and bad periods cluster together.
Sources and further reading
- Nicholas Metropolis and Stanislaw Ulam, “The Monte Carlo Method,” Journal of the American Statistical Association, vol. 44, no. 247 (1949)
- Paul Glasserman, Monte Carlo Methods in Financial Engineering (Springer, 2004)
- Moshe A. Milevsky and Chris Robinson, “A Sustainable Spending Rate without Simulation,” Financial Analysts Journal, vol. 61, no. 6 (2005)
- David M. Blanchett and Wade D. Pfau, “The Power and Limitations of Monte Carlo Simulations” (Advisor Perspectives, 2014)
- Simulations computed by Finabase: 10,000 paths of 30 years from a seeded random-number generator; the assumptions are stated in the text
