How Bad Is Running Out of Money in Retirement?

Key takeaways

• Across the 69 rolling 30-year retirements in the 1928 to 2025 US return record, a 4% inflation-adjusted withdrawal from a 60/40 portfolio ran dry 4 times.
• The worst of those four, starting in 1966, funded 25.6 years of real income before the pot emptied. Failure meant a gap of 4.4 years, not 30.
• At 4% the worst case funded 25.6 years whether the plan was 25, 30, 35 or 40 years long, while the failure rate rose from 0% to 15.3%.
• Raising the rate from 4% to 6% moved the failure rate from 5.8% to 47.8% and the worst case from 4.4 years short to 16.4 years short.
• A UK pot that empties leaves the full new State Pension, worth £241.30 a week in 2026/27, which is more than 4% of a £250,000 pot pays.

Running out of money in retirement meant 25 years funded, not zero

Here's the number you came for. The plan is 4% of the pot in the first year of retirement, that cash amount raised with inflation every year after, held in 60% shares and 40% government bonds and rebalanced once a year. Run that plan through every 30-year stretch in the 1928 to 2025 record of US returns and inflation, and it emptied the portfolio in 4 of 69 retirements. The worst of the four began in 1966 and paid 25.6 years of income in full. The other three paid 27.4, 27.7 and 27.8 years.

So the headline, that a 4% withdrawal survived 65 of those 69 histories, is true and almost useless on its own. It gives the odds. It says nothing about the size. On this record, running out of money in retirement at 4% meant falling about 4 years short of a 30-year plan with everything before that paid in full.

The picture changes as the rate rises. At 6% the same portfolio emptied in 33 of 69 windows, and the worst one funded 13.6 years. That leaves 16.4 years unpaid. Same word, same metric, a different event entirely.

The failure rate counts failures and refuses to measure them

The metric everyone quotes is the failure rate: the share of historical or simulated retirements in which the money ran out before the end. Javier Estrada, a finance professor at IESE Business School, put its problem plainly in the Journal of Financial Planning in July 2023. The failure rate, he wrote, "neither distinguishes between a failure early or late in a retirement period nor does it account for any bequest left".

His example is the one that sticks. The failure rate "is indifferent between two strategies that fail, say, five years or 25 years into a retirement period; obviously no retiree would be indifferent between those two failures". One of those is a catastrophe. The other is a late trim to spending. The percentage treats them as the same event.

The success side is just as blunt. Morningstar's The State of Retirement Income: 2025, published with data as of 30 September 2025, defines its 90% target as 900 of 1,000 simulated paths ending with a positive balance after 30 years. Its own description of what counts: "If at least a penny remains in the portfolio at the end of year 30, then the trial is considered successful." A penny and a doubled portfolio score the same.

That is a defensible way to build a planning rule. It isn't a measurement of anything a retiree experiences. What a person experiences is a number of years and a size of cut, which is why the arithmetic below is in years.

Years of real income funded, by withdrawal rate and horizon

The table below runs the same 60/40 portfolio through every window the data allows, at six withdrawal rates and five planning horizons. Each cell holds two things: the fewest years of inflation-adjusted income any single window funded, and in brackets the share of windows that fell short of the horizon. Years funded past the end of the plan count the leftover balance at the same real withdrawal, which is how Estrada's work measures it.

Read the 4% row across and the point of the piece is in it. The worst window funded 25.6 years against a 25-year plan, a 30-year plan, a 35-year plan and a 40-year plan. The same 1966 retiree, the same empty pot, the same date. What moved was the failure rate, from 0% to 15.3%, because the yardstick got longer.

The failure rate is partly a claim about markets and partly a claim about how long you expect to live. The magnitude of the shortfall is only about markets. The chart plots the first number from the 30-year column, which is the worst historical outcome at each rate rather than the average one.

The rate itself does move the magnitude, and hard. The step from 3.5% to 4% took 10.3 years off the worst case. Each half-point after that took 3 to 5 years more, until 6% left a pot that funded 13.6 years. The probability and the magnitude both deteriorate; they just deteriorate for different reasons, and only one of them is reported.

Every 4% failure since 1928 began within five years of the others

The four failing 30-year windows at 4% started in 1965, 1966, 1968 and 1969. That is one episode: a cohort who began drawing in the late 1960s and then met the inflation of the 1970s with a portfolio already run down. Our piece on sequence risk works through why the order of returns does this.

At 5% the failures spread out, which tells you something the headline rate doesn't. 19 of the 69 windows fell short, starting in 1929, 1930 and 1937 as well as almost every year from 1956 to 1973. The median one funded 22.5 years, so the typical 5% failure was 7.5 years short. The worst, from 1969, funded 17.7 years.

At 6%, 33 windows failed and the median failure funded 19.6 years. A retiree in that median case had paid themselves in full for two decades and then had nothing from the portfolio for a decade more.

What the coverage ratio measures that a percentage can't

Estrada and Mark Kritzman proposed a replacement in 2019, and it is simple enough to compute at a kitchen table. The coverage ratio is the number of years of inflation-adjusted withdrawals a plan supported, during and after the retirement period, divided by the length of the plan. Below 1 the portfolio emptied early. Above 1 it funded the whole plan and left something over.

Estrada's own illustration uses a $1,000 pot and $40 withdrawals over 30 years. A plan that emptied in 24 years scores 0.8. One that ended exactly on time scores 1.0. One that left $240, enough for six more years of withdrawals, scores 1.2.

Run that on his data, which is Robert Shiller's US series from 1872 to 2021 and 121 rolling 30-year periods, and the 60/40 portfolio at 4% has a failure rate of 3.3%. The median coverage ratio in the same sample is 2.02, and the mean is 2.53. The typical retirement didn't scrape home. It funded more than twice the job. Even the first percentile, the worst outcome in a hundred, came in at 0.87, which is 26.1 years of the 30.

At a 6% withdrawal his numbers turn: a failure rate of 48.8%, a median coverage ratio of 1.06 and a first percentile of 0.43. That last figure is a portfolio that funded 12.9 years of a 30-year plan. One metric, two completely different retirements, and the failure rate on its own can't separate them.

The run here on the 1928 to 2025 data gives a 5.8% failure rate where Estrada's gives 3.3%. The gap is the denominator, not the diagnosis. A 3.3% rate on 121 periods is 4 failures, the same count this run finds in 69 windows. His sample starts in 1872, so the same handful of bad start dates is divided into a longer history.

A UK pot that empties falls to £12,547.60, not to zero

Every figure so far describes a portfolio, not a household. Morningstar is explicit that its safe rates exclude "Social Security or other nonportfolio income sources", and the historical tables work the same way. For a UK retiree that omission is most of the income.

The full rate of the new State Pension rose to £241.30 a week for 2026/27, up from £230.25. That is £12,547.60 a year, paid for life, and the step up from £230.25 is an increase of 4.8%. Set it beside the pots people actually have. In the FCA's retirement income market data for April 2025 to March 2026, 8.6% of pots accessed for the first time were worth £250,000 or more. A 4% withdrawal from £250,000 is £10,000 a year, so for more than nine pots in ten the State Pension is the larger of the two incomes.

So running out of money in retirement, for most UK households, is a step down to a floor rather than a step to zero. That reframes the worst case. The 1966 retiree's portfolio stopped paying 4.4 years early. The State Pension did not stop. What the retirement shortfall actually looked like was 4.4 years of £10,000 going missing from a household income of about £22,500 a year, which is a cut of roughly 44% rather than the loss of everything. For a smaller pot the cut is smaller still, because the floor is a bigger share of the total.

The arithmetic runs the other way for large pots and for anyone retiring before State Pension age, which is where the portfolio is the whole income and the gap is the whole gap. Our piece on the 25x rule covers how big the pot has to be before that stops being true.

The horizon you fail against is a guess about your own lifespan

If the failure rate depends on the length of the plan, and the length of the plan is a guess, then part of the failure rate is a guess too. The Office for National Statistics put cohort life expectancy at 65 in the UK in 2024 at a further 20.0 years for men and 22.7 for women, in figures released in May 2026. For someone reaching 65 in 2049 the projection is 22.0 years for men and 24.6 for women.

Those are averages, and half of each cohort outlives them. The 35 and 40-year columns are not eccentric. They are the rows a couple in their early sixties might reasonably read, and they are where the 4% failure rate climbs to 9.4% and 15.3% while the worst case stays at 25.6 years. Lengthening the horizon changes how often you'd call the plan a failure. It does not change how far the money fell short.

The strongest objection: nobody keeps drawing from an empty pot

Here's the serious case against everything above. The fixed real withdrawal is a mechanical assumption, and no real retiree follows it into the ground. Someone whose pot halves cuts their spending, and the warning was visible early: the 1966 portfolio held 9.3% of its real starting value with seven years of the plan still to run.

The numbers support the objection. Morningstar's base case for a fixed real income is a 3.9% starting rate, and allowing the withdrawal to move with the portfolio lifts that to as much as 5.7% of the starting balance. The gap of 1.8 percentage points is bought entirely by agreeing to spend less in bad years. Our piece on withdrawal guardrails works through the rules that formalise it.

Behaviour runs the other way too. In the FCA's 2025/26 data, 320,801 of the 702,412 plans making regular partial withdrawals took 8% or more of the pot in a single year, and 401,137 plans entered drawdown over the same twelve months. Most people are not running a 4% rule at all. They are running something the historical tables do not describe.

Estrada turns the same kind of criticism on averages of his own metric, and it is the fair one to put here. A high average coverage ratio can hide an unacceptable failure rate: he gives the case of a strategy that depletes in year three half the time and leaves 60 years of withdrawals the other half, producing a flattering mean of 1.6 and a 49% chance of ruin. Magnitude without probability is as misleading as probability without magnitude. The argument here is that the two belong side by side, not that one replaces the other.

What 69 overlapping windows cannot tell you

These are backtests, not forecasts, and the sample is thinner than the row counts suggest. The 69 thirty-year windows drawn from 98 years of data share most of their years with each other. The 1966 start date is not an independent draw; it appears, with its neighbours, in every window that straddles the 1970s. Four failures at 4% is one historical episode counted four times.

The data is also American. US shares and US Treasuries over this period were a good draw by world standards, and Estrada's cross-country work finds failure rates above 27% at 4% for the average country in his 21-market sample. Our piece on the safe withdrawal rate outside the US covers how far that moves the answer.

Neither the historical runs here nor Estrada's tables charge any fee or tax. Estrada states the omission directly: the figures "ignore the impact of taxes", and transaction costs are left out as well. Platform charges, fund costs and tax on a drawdown withdrawal all come out of the same return the income is taken from, so every years-funded figure in this piece is an overstatement by the size of those costs.

And the success-rate framing still has its uses, which is why the Trinity study grid in our piece on withdrawal rate success rates remains worth reading beside this one. A probability answers the question of whether to accept a plan. It just cannot tell you what accepting it costs when it goes wrong.

What would change these numbers

A worse episode entering the record would move the worst case directly. A 30-year window needs a start date no later than 1996, so retirements begun after that join the count one start date a year as their windows close. The cohort who retired in the 2000s is still being counted.

Costs would move everything. Both Estrada's figures and the runs here are gross, and a total annual charge of a single percentage point on a 60/40 portfolio drawn at 4% is a quarter of the income taken. Whether the 25.6-year worst case holds up at realistic UK platform and fund charges is a different calculation, and it is the one a reader in a drawdown account is actually in.

The assumption that spending stays fixed in real terms is the biggest lever of the three. Every years-funded number above is the answer to a question nobody is forced to ask: what happens if you refuse to adjust. The figures for a retiree who does adjust are different, smaller in their worst case and harder to state in one number, and they are the reason the magnitude of failure deserves its own column rather than a replacement for the probability.