Safe withdrawal rate calculator

The withdrawal rate is the first year's withdrawal as a share of the portfolio; the dollar amount then rises with inflation every year. This calculator runs your rate against every retirement the record is long enough to finish, one for each starting year, and reports how many funded every withdrawal. Over 30 years that is the 69 that began between 1928 and 1996, taken from returns recorded through 2025. It also gives the rate that never failed in that record. Two other models sit behind the same control: years drawn at random from the record, and a single constant return.

Real returns for US stocks and 10-year Treasuries, 1928 to 2025 checked against Shiller's published workbook, last on 2026-09-12 Runs in your browser — nothing you type is sent to this site's servers or to its analytics, not even when you reload or share the link No sign-up Methodology and data status

What the record says

A 4% withdrawal over 30 years, rebalanced each year, funded every withdrawal in 65 of the 69 retirements that began between 1928 and 1996 with half the portfolio in stocks. The count is the same at three quarters in stocks and at all stocks: what separates the mixes is not how many retirements lasted, but how much they could have taken. The rate that never failed is 3.73% at a half-stock mix, 3.82% at three quarters and 3.69% with no bonds at all. More stocks is not monotonically safer in this record: the figure rises to three quarters and falls again with no bonds at all. Why it has that shape is not something 98 years can settle, so this page reports the shape and stops there.

The worst half-stock retirement began in 1966 and funded 25 of its 30 withdrawals. For an all-stock portfolio the worst began in 1969, which funded 23. Bengen names 1966 as the worst start year in his own data, computed on a different bond series over a shorter sample, and the record here agrees with him.

How a period is run

• The withdrawal comes out at the start of each year and is held constant in buying power, never rising or falling with the market.
• It is taken from both sleeves in proportion to what each holds, not from bonds first.
• What is left earns that year's real return on the chosen mix, and the mix is restored at the start of every year.
• A period counts as a success only if every withdrawal, including the last, is funded in full.
• Returns are real, already net of inflation, so no inflation figure is needed and none is asked for.

What the record cannot tell you

This is the part most safe-withdrawal pages leave out, and it matters more than the success rate itself.

• There are fewer independent tests than there are starting years. 98 years of returns yield only 69 distinct 30-year windows, and each shares 29 of its 30 years with the next. Those 98 years hold about 3.3 non-overlapping 30-year spans, which is a rough measure of how little independent long-horizon history sits behind 69 heavily overlapping windows rather than a strict count of independent trials. Resampling the record and redoing the whole calculation puts a standard deviation of 10.7 to 22.1 percentage points on a success rate — at 4% and three quarters in stocks, a 95% interval of 60.9% to 100.0% around the 94.2% the record gives. That is wider than the gap between the rates people argue about, and it is why this page gives counts rather than decimals. The experiment, its parameters and its table are on the methodology page. One country, one century, and a century that was unusually kind to it.
• The years are December averages, not calendar years. The price for a month is the average of that month's daily closes, so 2008 reads here as a 39.29% fall rather than the 37% usually quoted for the calendar year. Neither is wrong; they measure different intervals.
• The bond series is constructed, not investable. It is the holding-period return implied by the 10-year Treasury yield, with no bid-ask, no roll cost and no fund. A real portfolio would have earned less. Before 1953 the underlying yields are interpolated between annual anchors rather than observed monthly.
• The stock series is the S&P Composite, not the S&P 500, which did not exist before 1957, and its dividends are interpolated from less frequent anchors throughout.
• Nothing here is net of tax or fees. Both series are gross, and "real" means deflated by the consumer price index for all urban consumers, which is not a retiree's spending basket.
• Bonds did not always cushion the fall. In 2022 the bond series fell 20.16% in real terms, its worst year in the whole record, while stocks fell 20.14%. The two sleeves fell together.
• A backtest is not a forecast. The rate that never failed is the worst case of one sample, not a floor on the future.

Drawing the years at random

The second model keeps every convention above and changes one thing: instead of walking the record in order it draws 10,000 plans, each taking 30 years at random from the 98 in the record. A year is drawn whole, the stock return and the bond return of the same annual observation together, so the two sleeves still move as they did. In 2022 both fell about a fifth at once, and a model that can never do that understates how bad a year can be.

Nothing is fitted. A normal or lognormal curve would need an average and a spread that no document states, and every figure on this site carries the document it was read from. Resampling assumes that future years look like recorded ones, and it also treats those years as interchangeable draws that may be taken in any order. Both are assumptions you can weigh and reject.

At 4% over 30 years it puts the odds at about 56% with no stocks, 86% at a quarter, 92% at a half, 91% at three quarters and 88% with no bonds. Set against the recorded order, which funded 65 of 69 at a half-stock mix, those numbers run in opposite directions at opposite ends: drawing years at random is more hopeful than history for a bond-heavy plan and less hopeful for a stock-heavy one. In this model, then, drawing years at random is not consistently the more conservative of the two; which way it leans depends on the mix.

Why the two disagree, and what that is not

It is tempting to say the record's ordering did it. The ordering is not what does it. Pulling the comparison apart gives three separate changes, and only one of them reverses with the mix.

• The order within a period contributes almost nothing. Each recorded 30-year stretch sits close to the middle of every possible reshuffling of its own years, so history's particular sequences are unremarkable.
• Which years fall together in one period is the part that reverses. Two different things drive it in two different sleeves, pulling opposite ways: real bonds spent four decades below their 1940 peak, a stretch no random draw will reproduce, while stocks were less scattered over 30 years than random draws make them. Where those two balance is where the direction flips.
• And roughly two points of it is not about markets at all. A recorded period uses 30 different years; a random draw can take 2008 three times. That alone makes random draws look worse, at every mix from a quarter stocks up. At a half-and-half portfolio it is the entire difference between the two models.

What this model cannot do

• It keeps the risk of a bad decade landing early, but not the shape history gave it. A drawn path has an order, and a portfolio being drawn down is affected by it, so the failures here are failures of sequence. What the draw does not keep is the way one year followed another in the record: the runs, the regimes, the slow grinds. The recorded model keeps those. Neither number is the better estimate; they answer different questions.
• 10,000 plans are not 10,000 futures. They are that many rearrangements of the same 98 years. They hold no inflation, no country and no valuation the record does not already hold.
• Whole percentages, and no finer. Change the random seed and the figure moves about a point; the 98-year record it draws from is itself uncertain by eight to eighteen points. Running more plans narrows the smaller of those two and leaves the larger untouched.
• A plan that ran out counts as zero in the balances, which is why a low percentile can read as nothing. That is a count of failures, not a balance you might end with.
• It is not a gentler model, and should not be read as one. The worst 30 years real bonds actually delivered turns up in about one plan in forty-six here, against the one-in-sixty-nine share it holds of the record's own windows. On that measure the draw is if anything harsher than the record, not kinder.

For the same reason the calculator offers no "rate that never failed" in this model. The figure carries several basis points of seed noise, so it can only honestly be given to a tenth of a point: at a 90% success target over 30 years it is about 4.1% at a half-stock mix, 4.0% at three quarters and 3.7% with no bonds. It is an estimate, not a floor.

The formula

With a real return g (the nominal return less inflation, compounded: (1 + r)/(1 + i) − 1) and withdrawals at the start of each of n years, the first-year rate that exhausts the portfolio at the end of year n is w = g ÷ [(1 + g)(1 − (1 + g)⁻ⁿ)]. With no real return it is simply 1/n: 3.33% for 30 years. Years remaining for a given first-year withdrawal W and balance B solve the same equation for n.

What a constant return leaves out

Markets do not pay the average every year. Two retirements with the same average return but different orderings end very differently, because withdrawals taken in a downturn sell more shares. That is why the 4% rule (the 4 percent rule) was derived from every retirement start year since 1926 rather than from an average: in Bengen's data it is the rate that lasted at least 30 years even for the worst start year, 1966, whose portfolio ran out after 33 years. Against a constant-return calculation, that worst case corresponds to a real return well below the long-run average. Use this page to see the arithmetic; treat the rate that survives history as the conservative bound.

Sources

Historical real returns: verified · last verified 2026-09-12. Verification: Two independent derivations from the same primary workbook on 2026-09-12 agree bit for bit: the lead's (docs/verification/2026-09-retirement/derive-historical-returns.py) and a blind second one that did not see it (report V-HIST, data V-HIST-returns.json). Across 98 years and four figures per year the largest absolute difference is 0.000e+00. Independence here is in the reader, as it was for the AFR rulings, because only one publisher issues this series. A red team (report R-HIST) then attacked the engine that consumes the file with its own clean-room implementation and a third in Python, agreeing run for run across 360 input combinations. See the methodology page.

Questions this page answers

What is the 4% rule?

Withdraw 4% of the portfolio in the first year of retirement and raise the dollar amount by inflation each year after. In Bengen's 1994 study (retirements beginning each year from 1926 to 1976, half to three-quarters in stocks and the rest in intermediate-term Treasuries), a 4% first-year withdrawal raised with inflation never ran out in less than 32 years. The Trinity study (1926–1995, with long-term corporate bonds) found the same rate lasted 30 years in 95% of periods with a 50/50 portfolio and 98% with 75% stocks. It is a historical finding about worst cases, not a guarantee, and it assumes 30 years, no taxes or transaction costs, and spending that changes only with inflation.

Is 4% safe for 40 or 50 years?

Less so. At a constant 5% return and 2.5% inflation the rate that lasts exactly 40 years is about 3.85% and for 50 years about 3.40%; in Bengen's own data a first-year rate of about 3.5% lasted at least 50 years in every case.

How much do I need to get $2,000 a month?

$2,000 a month is $24,000 a year. At a 4% first-year rate that takes $600,000; at the 4.63% that a constant 5% return and 2.5% inflation give for 30 years, $518,788. Both figures are before tax; Social Security or a pension reduces what the portfolio must supply.

Related

• Retirement withdrawal calculator

  The fuller version: spending that rises with inflation, Social Security and a pension starting on their own dates, and federal tax on each withdrawal.

• FIRE calculator

  Turn the withdrawal rate into a target: the FI number it implies and the years to reach it from your savings rate.

• Coast FIRE calculator

  Whether the amount already saved grows to that target by your retirement age without further contributions.

• How the safe withdrawal rate is computed

  The annuity formula behind the rate, the years-remaining solution and why a constant return is the optimistic case.