---
title: "Is the 4% Rule Safe? K-Dense Web Stress-Tests 155 Years of US Retirement Outcomes"
description: "K-Dense Web simulated every 30-year US retirement since 1871. The 4% rule failed 3% of the time, and more realistic assumptions pushed that rate higher."
updatedAt: "2026-09-26"
author: "Yuhuan He"
authorTwitter: "yhhonx"
authorLinkedIn: "https://www.linkedin.com/in/yuhuanhe/"
tags: ["Use Case", "Finance", "Retirement", "Economics"]
canonical: "https://www.k-dense.ai/blog/four-percent-rule-155-years-retirement-outcomes"
---
**Across every 30-year retirement that could have started in the US between 1871 and 1996, withdrawing a fixed, inflation-adjusted 4% a year from a 60/40 portfolio ran out of money 3.05% of the time. That number sounds reassuring, but it sits at the optimistic end of the analysis, not in the middle.** When K-Dense Web added realism one piece at a time (more accurate volatility, resampled market histories, annual fees), the failure rate rose each time, to as much as 11.3%.

The "4% rule" comes from William Bengen's [1994 study](https://www.financialplanningassociation.org/sites/default/files/2020-05/7%20Determining%20Withdrawal%20Rates%20Using%20Historical%20Data.pdf) and the later [Trinity Study](https://www.aaii.com/files/pdf/6794_retirement-savings-choosing-a-withdrawal-rate-that-is-sustainable.pdf), and it is still one of the most quoted numbers in personal finance. I asked K-Dense Web to test it on Robert Shiller's public monthly US market data, going back to 1871, and to show how the order of good and bad years affects the outcome. This post is historical research, not financial advice, and the report says the same on its title page and in the footer of its main figures.

The session produced a 22-page illustrated report with 30 references. The feature I want to highlight is the **methodology review**. Between steps, a reviewer agent checks the analysis and can request enhancements, and in this session those requests changed how the headline should be read. You can browse the full session, including every script and result file, [here](https://app.k-dense.ai/share/session_20260923_150519_d46db96de78f).

## Step 1: Building 155 Years of Returns

Before any simulation, the agent had to build a clean monthly return series, and the first problem appeared at the download. The analysis plan named the Shiller data file on Yale's website, but the agent downloaded three candidate copies and compared their coverage. The Yale copy ended in September 2023, three years out of date, while the maintained copy on shillerdata.com ran to September 2026. It used the current file and kept Yale only as a fallback. The last three months had no dividend data yet, so the series was truncated at June 2026, giving 1,866 months with no gaps.

Each return series was checked by a second route. The agent's real stock returns matched Shiller's own total-return column exactly (it noted that this confirms the implementation but is not a methodologically independent check, since both use the same dividend convention). Bond returns from exact bond repricing matched a separate duration-based approximation with a correlation of 0.9996, and the small remaining gap is the convexity term the approximation leaves out. Historical sanity checks also held: the real stock drawdown from September 1929 to June 1932 came out at -75.8%.

## Step 2: Simulating Every Retirement

The simulation starts with $1,000,000 and withdraws a fixed real amount every month for 30 years. It covers all 1,506 complete 30-year windows in the data (February 1871 to July 1996), three withdrawal rates, and two portfolios: 60% stocks and 40% bonds, or 100% stocks.

![Every historical 30-year retirement path](https://www.k-dense.ai/blog/four-percent-rule-155-years-retirement-outcomes/all_retirement_paths.png)
*Each gray line is one retirement start date. The black line is the median outcome. Highlighted are September 1929 (pink), January 1966 (orange) and January 1982 (green). A line that reaches zero means the money ran out.*

| Withdrawal rate | 60/40 failure rate | 100% stocks failure rate |
|---|---|---|
| 3% ($30,000/yr) | 0.00% (0 of 1,506) | 0.00% (0 of 1,506) |
| 4% ($40,000/yr) | 3.05% (46 of 1,506) | 2.12% (32 of 1,506) |
| 5% ($50,000/yr) | 24.17% (364 of 1,506) | 16.33% (246 of 1,506) |

Two results stand out. Risk is sharply nonlinear: moving from 4% to 5% multiplies the 60/40 failure rate by about eight. And in this sample, 100% stocks failed less often than 60/40 at 4% and 5%, and ended with more money at the median at every withdrawal rate. The agent was careful to scope that second result: it follows from a fixed real withdrawal over exactly 30 years, and it would not necessarily hold for longer horizons, flexible spending, or a retiree who cannot tolerate large interim losses.

Just as important is where the failures come from. All 46 failing 60/40 retirements at 4% started between May 1964 and January 1969, so for that portfolio, failure in US history is not spread out. It is essentially one event. The all-stock portfolio adds a second, smaller cluster: 13 retirements that began between February 1929 and June 1930, around the 1929 crash.

![Which retirement start dates failed](https://www.k-dense.ai/blog/four-percent-rule-155-years-retirement-outcomes/failed_start_dates.png)
*Start dates that ran out of money within 30 years, by withdrawal rate. At 4%, the 60/40 failures all fall between 1964 and 1969, while 100% stocks also failed for 1929 and 1930 starts. Failure only becomes widespread at 5%.*

## Why the Order of Returns Matters

The retiree who started in January 1966 earned +4.19% a year after inflation over the full 30 years and still ran out of money in month 313. The agent then replayed the exact same 360 monthly returns in reverse order. With the same returns and the same average, the portfolio ended with $1,787,868. Only the ordering changed.

![The 1966 retirement played forward and in reverse](https://www.k-dense.ai/blog/four-percent-rule-155-years-retirement-outcomes/sequence_risk_1966.png)
*Left: a simple worked example of why order matters once withdrawals begin. Right: the January 1966 retirement played in its real order (orange) and with the same monthly returns reversed (green).*

The comparison with 1929 makes the point sharper. Someone who retired in September 1929, at the exact market peak before the crash, survived with $331,478 left. Their 30-year return (+4.42%) was similar to the 1966 retiree's, but their first decade was +2.55% a year against -1.93%. The first ten years decide most of the outcome. The script that drew this figure also asserted that the plotted replay matched the stored result before saving it, so a mismatch would have stopped the run instead of producing a wrong chart.

## Feature Spotlight: The Methodology Review Changed the Headline

After the first step, the methodology review returned a pass with no critical issues and three optional enhancements, and the agent implemented all of them. The most important one concerned a quirk of the data. Shiller's stock price is a monthly average of daily closes, not a month-end price, and averaging hides some of the market's real volatility. The agent restored the volatility to a realistic level while keeping the average return fixed, and the 4% failure rate more than doubled, from 3.05% to 6.83%.

Further reviewer enhancements led to three more robustness tests in a later step. First, a stationary [block bootstrap](https://doi.org/10.1080/01621459.1994.10476870) generated 10,000 synthetic 30-year market histories by resampling blocks of real months, which breaks the historical clustering around the 1960s. Second, a fee test charged a constant annual expense. Third, a spending-guardrail test used the [Guyton-Klinger](https://www.financialplanningassociation.org/article/journal/MAR06-decision-rules-and-maximum-initial-withdrawal-rates) rules, which cut spending by 10% when the current withdrawal rate drifts too high and raise it by 10% when it drifts low.

![Robustness tests: bootstrap, fees and spending guardrails](https://www.k-dense.ai/blog/four-percent-rule-155-years-retirement-outcomes/robustness_checks.png)
*(A) Failure rates from 10,000 resampled market histories, compared with the historical frequency. (B) The effect of annual fees on the 60/40 portfolio. (C) Guardrails eliminate depletion, but only by cutting spending.*

Put side by side, the six treatments of the same 60/40, 4% plan tell one story:

| Treatment | Failure rate |
|---|---|
| Historical frequency (headline) | 3.05% |
| Withdraw at the end of each month instead of the start | 2.92% |
| Volatility restored to a realistic level | 6.83% |
| Block bootstrap, 10,000 synthetic histories | 6.68% (95% CI 6.19 to 7.17) |
| 1% annual fees | 11.29% |
| Guyton-Klinger spending guardrails | 0.00%, but worst-decile spending falls to $31,870/yr |

The report's summary is one sentence I keep coming back to: every adjustment that adds realism pushes the failure rate up, except the one that lets the retiree spend less. The withdrawal-timing row is a convention check rather than added realism, and it barely moves the number. The guardrail result also came with a clear warning. Its 0% is partly a definitional artifact, because a portfolio whose spending shrinks with it can hardly run out by construction. The cost reappears as spending cuts: in the worst tenth of retirements, spending averaged $31,870 a year against the planned $40,000. The fixed and guardrail plans fund different lifestyles, so their failure rates are not a like-for-like comparison.

Without the review, the session would still have produced a correct 3.05%. With it, the report explains why 3.05% sits at the optimistic end of the range, by how much the realistic adjustments move it, and which assumptions each adjustment relies on.

## The Writer Checked the Analyst

K-Dense Web writes the final report in a separate phase that reads the analysis output files directly. In this session that separation caught a real error. The analysis README stated that the 1966 retiree had a higher 30-year return than the September 1929 retiree. When the writing agent read the result files, it found the opposite: +4.19% for 1966 and +4.42% for September 1929. The report uses the corrected comparison and documents the discrepancy in its evidence section, instead of copying the claim from the notes.

The same care went into the references. Using the **parallel-web** skill, the agent ran literature searches on the safe-withdrawal literature, sequence risk, guardrails, the bootstrap and the Shiller data, then verified each reference's details against publisher pages. Two errors were corrected along the way: one widely cited paper was published in 2018, not 2017, and its commonly circulated DOI does not resolve. Three page ranges could not be confirmed at the publisher level, and they are flagged in the bibliography instead of presented as certain.

## What the Analysis Does Not Claim

The limits are stated plainly in the report. The 1,506 windows overlap heavily (neighbors share 359 of 360 months), so the effective sample is only about five independent 30-year periods, and the historical failure rates are frequencies over one realized path, not probabilities with standard errors. The bootstrap assumes the return process is stable over 155 years that span the gold standard, Bretton Woods and fiat money, and its 60-month block length was chosen on judgment. Taxes, trading costs, Social Security, annuities and mortality are not modeled, and bonds before 1953 rest on a spliced long-term rate instead of a true 10-year Treasury yield.

## Try It

A study like this usually means days of work: cleaning a 155-year dataset, writing a simulation engine, validating it, and running every robustness check a careful reviewer would ask for. In this session, K-Dense Web did all of that, and then wrote a general-audience report that is honest about which number to trust.

Try it yourself at [app.k-dense.ai](https://app.k-dense.ai), or browse the full session [here](https://app.k-dense.ai/share/session_20260923_150519_d46db96de78f).

Questions? Contact us at contact@k-dense.ai.