The Number Everybody Reports and Nobody Defines
Somewhere in nearly every modern retirement plan there is a percentage. It shows up under a fan chart, in a bold color, usually with a word like confidence or probability of success next to it. Ninety-one percent. Eighty-four percent. Ninety-seven percent.
Clients react to it immediately and almost always incorrectly. A 90% score gets heard as "there is a 90% chance my retirement works out." An 82% score gets heard as "there is nearly a one-in-five chance I end up broke." Neither reading is right, and the gap between what the number says and what people hear it say is one of the most consequential misunderstandings in retirement planning.
The number is not a forecast. It is the output of a specific counting exercise, run against a specific set of assumptions, measured against a specific and rather crude definition of the word "success." Understanding exactly what gets counted — and what gets thrown away in the counting — changes how you should use it.
A Success Rate Is a Count, Not a Forecast
"90% success" means: out of 10,000 simulated market histories, 9,000 of them left the portfolio with a positive balance on the last day of the planning horizon. It says nothing about how much was left, how close the other 1,000 came, when they failed, or whether the retiree would have noticed in time to react. Those are all different questions, and most of them matter more.
How Monte Carlo Simulation Actually Works
The traditional retirement projection is a straight line. Assume 7% returns, assume 3% inflation, assume you spend $65,000 a year, and see whether the money lasts to 95. It produces a single, clean, and deeply misleading answer, because no retiree will ever experience a 7% return. They will experience a sequence — twenty-nine, fifteen, minus eighteen, six, minus four, thirty-one — that happens to average something near 7%, and the order those numbers arrive in will matter more than the average does. That's sequence-of-returns risk, and a straight-line projection is structurally blind to it.
Monte Carlo simulation exists to fix exactly that blindness. Instead of one path, it runs thousands. The mechanics are simpler than the name suggests:
- Describe the market statistically. Rather than picking one return, you specify a distribution: stocks average 10% with an 18% standard deviation, bonds average 5% with a 6% standard deviation. That describes a range of plausible years, not a single one.
- Draw one random year at a time. For each year of retirement, the engine draws a random return for stocks and one for bonds from those distributions, and blends them by the portfolio's allocation. Most engines draw from a lognormal distribution rather than a plain normal one, because a normal distribution would happily generate a return below −100%, which is not a thing that can happen to a portfolio.
- Run the whole retirement. Apply the year's return, subtract the year's spending net of guaranteed income like Social Security and pensions, inflate next year's spending, and repeat until the ending age. Record whether the balance ever hit zero.
- Do it ten thousand times. Each of the 10,000 runs is a complete alternative history — a different order of good and bad years. Some open with three straight bear markets. Some open with a boom.
- Count the survivors. The success rate is that count, divided by the number of trials. That's the entire calculation.
Everything interesting about Monte Carlo lives in the details of step five and in the honesty of step one.
The Working Example
For the rest of this article we'll use one concrete plan, run through 10,000 simulated paths, so that every claim can be checked against a number rather than asserted:
| Input | Value |
|---|---|
| Age at retirement | 65 |
| Planning horizon | To age 95 |
| Starting portfolio | $1,000,000 |
| First-year spending | $64,000, inflated 3% annually |
| Social Security | $2,400/month beginning at 67 |
| Allocation | 60% stocks / 40% bonds |
| Stock assumption | 10% mean, 18% standard deviation |
| Bond assumption | 5% mean, 6% standard deviation |
| Simulations | 10,000 |
Result: a 90.4% success rate, with a median ending balance of about $2.55 million in nominal dollars — roughly $1.05 million in today's purchasing power after thirty years of 3% inflation.
That is the number a client would see. Now let's take it apart.
What "Success" Counts — and What It Throws Away
Success in a Monte Carlo engine is almost always binary: did the portfolio end the horizon above zero? A path that finishes with $47 is scored identically to a path that finishes with $8 million. A path that runs dry one month before the ending age is a failure; a path that runs dry one month after it is a success. This is a very blunt instrument, and four things get discarded in the process.
1. When the failures happen
This is the most under-reported statistic in retirement planning, and the one that most changes how the number should be heard.
In our 90.4% plan, 964 of the 10,000 paths failed. Here is where they failed:
| Depletion age | Failing paths | Share of failures |
|---|---|---|
| Before 80 | 0 | 0% |
| 80–84 | 39 | 4% |
| 85–95 | 925 | 96% |
Not one path in ten thousand ran out of money before age 80. The median failing path depleted at age 92. The earliest failure anywhere in the simulation was at 80.
So the honest translation of "you have a 10% chance of failure" in this plan is not "there is a one-in-ten chance you go broke in retirement." It is "there is a one-in-ten chance that, somewhere in your late eighties or nineties, this portfolio stops covering the gap between your spending and your Social Security — with twenty-plus years of advance warning that it was heading there."
Those are wildly different statements, and only one of them is worth losing sleep over.
A 'Failure' Is Rarely a Cliff
A depleting portfolio does not surprise anybody. It declines visibly for a decade first. A retiree in one of those 964 failing paths has ten to fifteen years of falling balances to respond to — by trimming spending, tapping home equity, or leaning harder on guaranteed income. The simulation, by design, assumes they do nothing at all and keep spending on schedule until the account hits zero. That assumption is what generates most of the failures, and no real retiree behaves that way.
2. How badly they fail
The engine records whether a path hit zero, not how deep the hole was. A plan that comes up $9,000 short in the final year and one that runs out at 82 are the same data point. Since we now know almost every failure in this plan lands in the last decade, most of that 9.6% consists of near-misses — plans that would have been rescued by a 5% spending trim a decade earlier.
3. How enormous the upside is
Here is the distribution of ending balances at 95 in the same plan:
| Percentile | Ending balance at 95 (nominal) |
|---|---|
| 10th | $31,000 |
| 25th | $979,000 |
| 50th | $2,551,000 |
| 75th | $5,007,000 |
| 90th | $8,339,000 |
The median path ends with two and a half times the money it started with, in nominal terms, after thirty years of withdrawals. The 75th percentile ends with five times. A retiree fixated on the 9.6% failure probability is ignoring that the far more likely outcome — by an enormous margin — is dying with a fortune they could have spent, given away, or enjoyed.
The success rate has no vocabulary for that. Unspent money and enjoyed money score exactly the same.
4. Whether you'd have adapted
Covered above, but worth stating plainly as a rule: the standard success rate models a retiree who never changes their behavior. It is a measure of the plan's robustness to total passivity, not a measure of how your retirement will go.
The Precision Trap: How Solid Is That Number?
Software prints "90.4%" with a decimal point, which implies a precision it does not have. A Monte Carlo result is a sample estimate. Run the same plan with a different random draw and you get a slightly different answer.
We measured it. Running the identical plan twenty times with different random draws:
| Simulations per run | Range across 20 runs | Spread |
|---|---|---|
| 1,000 paths | 89.4% – 92.1% | 2.7 points |
| 10,000 paths | 89.6% – 90.8% | 1.2 points |
At 10,000 paths the answer is trustworthy to roughly a percentage point. At 1,000 paths — which some tools still use to keep the interface responsive — it wobbles by nearly three. Either way, the tenth of a percent is decoration. A plan that scores 89.4% and a plan that scores 90.1% are the same plan. Nobody should be adjusting an allocation, delaying a retirement date, or feeling relief over a difference that lives entirely inside the engine's sampling noise.
Ask Whether the Engine Is Seeded
An unseeded simulation draws fresh randomness on every run, so opening the same plan twice gives two different scores. For a client conversation that's merely awkward. For an advisor whose recommendations are recorded in a compliance file, a figure that cannot be regenerated from its own stored inputs is a genuine problem — you can no longer demonstrate what produced the number you presented. Good engines derive their random seed from the inputs, so the same plan always reproduces the same answer while still sampling the full distribution.
The Assumptions Do Almost All of the Work
If the output has a one-point margin of error, then any input change that moves the score by more than a point matters more than the output's precision. Almost all of them do. Same plan, one assumption changed at a time:
| Change | Success rate | Change from 90.4% |
|---|---|---|
| Stock return 10% → 9% | 85.2% | −5.2 |
| Stock return 10% → 8% | 77.5% | −12.9 |
| Inflation 3% → 3.5% | 83.0% | −7.4 |
| Inflation 3% → 4% | 74.8% | −15.6 |
| Plan to 100 instead of 95 | 83.4% | −7.0 |
| Plan to 90 instead of 95 | 95.8% | +5.4 |
| Spend 5% less ($60,800) | 94.6% | +4.2 |
| Claim Social Security at 70 | 90.1% | −0.3 |
Two lessons fall out of this table.
First, the assumptions dwarf the output. A single percentage point on the assumed stock return moves the score five points — five times the engine's own margin of error. Half a point of extra inflation moves it seven. If two advisors run the same client and one uses 10% equity returns while the other uses 8%, they will report a 90% plan and a 78% plan for identical circumstances, and neither will be lying. Always ask what return, inflation, and longevity assumptions produced a success rate before reacting to it.
Second, the ending age is an assumption, not a fact. Changing the planning horizon from 90 to 100 swings this plan from 95.8% to 83.4% without touching a dollar of the portfolio. Planning to 90 is not a more optimistic plan — it is a plan that stops measuring before the risky years begin. A 65-year-old couple has better than even odds that one spouse reaches 90, which is precisely why cutting the horizon there flatters the result so effectively.
The Social Security row deserves its own note. Delaying to 70 barely moved the headline number here, because the plan already survives in over 90% of paths and the extra benefit mostly arrives in paths that were fine anyway. But it substantially strengthens the failing paths — the late-life shortfalls that make up 96% of this plan's failures are exactly the ones a larger inflation-adjusted lifetime check addresses. This is a case where the success rate is the wrong lens entirely; see when to claim Social Security and survivor benefits and the higher-earner delay for the decision that actually matters there.
Why a "Safer" Portfolio Can Score Better and Leave You Poorer
Here is the sharpest illustration of the binary-metric problem. Same plan, same spending, only the stock allocation changed:
| Stock allocation | Success rate | Median ending balance |
|---|---|---|
| 20% | 86.4% | $934,000 |
| 40% | 91.7% | $1,729,000 |
| 60% | 90.4% | $2,551,000 |
| 80% | 88.1% | $3,432,000 |
| 100% | 85.4% | $4,146,000 |
Read that carefully. The allocation with the best success rate — 40% stocks — produces a median ending balance of $1.7 million. The allocation with the worst success rate — 100% stocks — produces a median ending balance of $4.1 million, nearly two and a half times more.
A retiree optimizing on the success rate alone would choose 40/60, gain 1.3 percentage points of simulated survival, and surrender roughly $2.4 million of median wealth for it. That is not obviously a good trade, and the metric that recommended it cannot even see the cost.
The mechanism is straightforward: cutting volatility narrows the distribution of outcomes. Narrowing the distribution pulls the worst paths up above the zero line, which is all the success rate measures — while simultaneously pulling the best paths down, which the success rate cannot see at all. Beyond a point — note the 20% row — the drag on returns overwhelms the volatility benefit and the score falls again.
Never Optimize a Portfolio on Success Rate Alone
Success rate is a floor metric. It answers "how often does this end above zero" and nothing else. Used as an optimization target it systematically pushes plans toward low volatility and low growth, because it credits the reduction in bad outcomes and ignores the destruction of good ones. Read it alongside the median and 25th-percentile ending balances, always.
Dynamic Spending: Why 100% Is Not Free
The most powerful single change you can make to a Monte Carlo result is to stop assuming the retiree spends on autopilot.
Apply Guyton-Klinger guardrails to our plan — cut spending 10% when the current withdrawal rate drifts 20% above where it started, raise it 10% when it drifts 20% below — and run it on the identical 10,000 return paths:
| Withdrawal rule | Success rate | Median ending balance |
|---|---|---|
| Fixed, inflation-adjusted | 90.2% | $2,583,000 |
| Guyton-Klinger guardrails | 100.0% | $2,456,000 |
Zero failures out of ten thousand. Not one path ran out of money.
That result is real, and it is also a trap if you read it as "guardrails made the plan safe." The guardrails did not conjure returns. They eliminated failure by cutting spending in the paths that were heading toward it. The risk didn't vanish — it changed form, from a small chance of running out of money into a certainty of adjusting the budget when markets disappoint.
That is very likely a trade worth making. Most retirees would rather trim 10% of discretionary spending in a bad year than face a 10% chance of a shortfall at 92. But it should be chosen with eyes open, and it argues for a different question than "what's my success rate." The better question is: how much would I have to cut, how often, and for how long? A dynamic-spending plan should be presented as a distribution of spending outcomes, not a single reassuring number.
What Success Rate Should You Actually Target?
The instinct is "as high as possible." It's wrong, and understanding why is the last piece of reading these numbers correctly.
- Below ~70% — the plan needs structural change. Not a tweak: a materially lower spending level, a later retirement date, more guaranteed income, or some combination.
- 70–85% — workable with adaptability built in. If the retiree has real spending flexibility, guardrails, or meaningful discretionary spending they'd happily trim, this range is often perfectly sound. Remember that most failures cluster deep in the plan, with years of warning.
- 85–95% — the practical sweet spot for most plans. Enough margin to absorb a bad first decade, not so much that the retiree is obviously underspending.
- Above 95% — investigate whether this person is denying themselves. A 99% plan is a plan whose owner will almost certainly die with a large unspent balance. If that's a deliberate legacy goal, excellent. If it's anxiety, the simulation has just quietly recommended thirty years of unnecessary frugality and scored it as a perfect result.
The 100% Plan Problem
There is no such thing as a 100% safe plan, only a plan conservative enough that this particular model can't break it. Push spending low enough and any engine returns 100% — it has simply stopped being able to see the risks it doesn't model, which include long-term care, a market regime worse than the assumed distribution, and a tax code that will change several times over a thirty-year retirement. Treat a 100% score as information about the assumptions, not as a guarantee about the future.
What Monte Carlo Does Not Model
Every engine makes simplifications. Knowing which ones your tool makes is the difference between using the output and being used by it. The common ones:
- Correlation between asset classes. Many engines draw stock and bond returns independently, which implicitly assumes zero correlation. Real correlation is not zero, and it has been sharply positive in inflationary periods — exactly when a retiree most needs bonds to zag. An independent-draw model credits the portfolio with more diversification than it has, so its success rate is, if anything, optimistic.
- Fat tails. Normal and lognormal distributions understate the frequency of extreme years. Real markets produce more 2008s than the bell curve predicts.
- Long-term care. A single extended care event can dwarf every other variable in the plan and appears nowhere in a standard simulation. See long-term care planning.
- Taxes and account location. Many simulations treat the portfolio as one undifferentiated pot. In reality, which account a dollar comes from changes what it's worth. See withdrawal sequencing.
- Spending that isn't a flat inflated line. Real retirement spending is usually front-loaded — the "go-go, slow-go, no-go" pattern — then spikes again for healthcare late. A constant inflated withdrawal is a convenient fiction.
- The COLA gap. Social Security's cost-of-living adjustment has historically run below general inflation. If an engine inflates spending at 3% while inflating Social Security at 2%, that widening gap is doing real work in the result, and it should be disclosed rather than buried.
None of these are reasons to distrust Monte Carlo. They're reasons to treat the output as one input to a judgment rather than a verdict.
How to Read a Success Rate in Six Questions
When a plan hands you a probability of success, these six questions extract the actual information:
- What return and inflation assumptions produced it? A two-point difference in assumed equity returns moved our plan by thirteen points.
- What ending age? Stopping at 90 instead of 95 was worth five points on its own.
- How many paths, and is the engine seeded? Below 10,000, treat two- to three-point differences as noise. If it isn't seeded, the number changes every time you look at it.
- When do the failures happen? A 90% plan whose failures all land after 88 is a very different plan from one that fails at 74.
- What's the median and 25th-percentile ending balance? This is where the cost of a "safer" allocation shows up.
- Does it assume fixed spending? If yes, it is measuring the plan's tolerance for a retiree who never reacts. Ask what guardrails do to it — and what they cost in spending.
Why This Matters for Advisors
A success rate is the most persuasive number in a planning conversation and the easiest one to misuse in both directions. Present it bare and a client either relaxes into false comfort at 94% or panics at 78% over failures that would all have occurred past age 88 with a decade of warning.
The conversation that actually lands has three parts. Show the fan chart, so the client sees the range rather than a point. Show when the failing paths fail, because "you might run short at 92" prompts a completely different reaction than "one in five chance of failure." Then show the plan again with guardrails on, and name the price — this version never runs out, and here is the spending cut that buys it. That sequence turns a scary percentage into a decision the client can actually make.
It also protects you. An assumption set is a recommendation. If a 10% equity assumption and a 90-year horizon are what turned a marginal plan into a 90% plan, that needs to be visible, documented, and reproducible — not buried in a settings panel that renders a different number next quarter.
How RetirementForge Helps
The Monte Carlo Analyzer runs 10,000 randomized paths and reports the parts most tools omit: the full percentile fan, the distribution of failure ages, and the median and quartile ending balances alongside the headline rate. Its random draw is seeded from the inputs themselves, so a figure presented in a client session can always be regenerated from the record that stored it. Guyton-Klinger guardrails run against the identical return paths as the fixed-withdrawal baseline, so the dynamic-versus-static comparison is genuinely apples to apples, and a side-by-side scenario lets you model annuitizing a slice of the portfolio on those same paths.
Pair it with the Withdrawal Planner for sequencing, the Social Security Optimizer for the claiming decision that strengthens the late-life failure paths, and the Roth Conversion Analyzer for the tax layer the simulation doesn't model. Every assumption is documented and every client session is captured in an immutable audit trail. Get started free and run a plan through 10,000 futures in about a minute.
This article is for educational purposes only and does not constitute financial, tax, or legal advice. All figures shown are outputs of a simulation run against stated assumptions and are illustrative, not predictive. Investment outcomes depend on market conditions, individual circumstances, and tax law, all of which change over time. Consult a qualified financial advisor for guidance specific to your situation.