Skip to main content

Our Methodology

SafeWithdrawls uses bootstrap sampling from historical returns instead of Monte Carlo simulation. This is a deliberate choice that produces more realistic retirement projections. This page explains why.

The Problem with Monte Carlo​

Monte Carlo simulation is the standard approach in most retirement calculators. It works by generating thousands of random return sequences drawn from a normal (bell-curve) distribution, typically calibrated to match the historical mean and standard deviation of market returns.

The core assumption -- that returns follow a normal distribution -- is the problem.

Real Markets Have Fat Tails​

In a normal distribution, extreme events are vanishingly rare. A daily market drop of 5 standard deviations should happen roughly once every 14,000 years. In practice, markets experience moves of this magnitude every few years.

The 2008 financial crisis is a stark example. Under normal distribution assumptions, the observed market declines would constitute roughly a 25-sigma event -- something so improbable it should essentially never occur in the lifetime of the universe. Yet it happened, and similar crises have happened repeatedly throughout market history.

This matters for retirement planning because tail events are exactly what destroys retirement portfolios. A retiree who starts withdrawing at the beginning of a severe, prolonged downturn faces the worst-case scenario -- and Monte Carlo systematically underestimates how often and how severely this happens.

What Monte Carlo Misses​

Beyond fat tails, normal-distribution Monte Carlo also fails to capture several important features of real market returns:

  • Volatility clustering -- Bad days tend to follow bad days. Crashes do not arrive as isolated events; they come in waves. Monte Carlo treats each day's return as independent.
  • Negative skew -- Markets crash faster than they rise. The distribution of returns is asymmetric, with a longer left tail. A normal distribution is symmetric by definition.
  • Cross-asset correlations under stress -- Stocks and bonds tend to become more correlated during crises, exactly when diversification matters most. Monte Carlo typically models them with a fixed correlation.

How Bootstrap Sampling Works​

Bootstrap sampling takes a fundamentally different approach. Instead of generating synthetic returns from a statistical model, it builds new return sequences by randomly sampling from actual historical returns.

Here is the process:

  1. Start with real data. SafeWithdrawls uses annual US stock and bond returns from 1928 through 2024 -- nearly a century of actual market history.

  2. Sample with replacement. To generate a 30-year projection, the simulator randomly picks 30 annual returns from the historical dataset. Each year's return is drawn independently, and the same historical year can be selected more than once (this is the "with replacement" part).

  3. Repeat many times. This process runs across many iterations, each producing a different sequence of returns. Your withdrawal strategy is tested against each sequence.

  4. Aggregate the results. The scores and projections you see represent how your strategy performed across the full range of generated sequences.

Why This Preserves Reality​

Because bootstrap sampling draws from actual returns rather than a theoretical distribution, it automatically preserves every feature of real market behavior:

  • Fat tails are included. The 1929 crash, the 1973--74 bear market, the 2008 financial crisis -- these are all in the dataset and can appear in any simulation.
  • Volatility clustering is preserved. When the simulator draws a bad year, there is a realistic probability of drawing another bad year nearby, because the historical dataset contains clusters of bad years.
  • Real correlations are maintained. Stock and bond returns in each sampled year reflect the actual relationship that existed in that year, including the breakdown of diversification benefits during crises.
  • No distributional assumptions. The method does not assume returns are normal, lognormal, or any other shape. The distribution is whatever history actually produced.

Crisis Scenario Testing​

In addition to randomized bootstrap simulations, SafeWithdrawls lets you test against specific historical sequences. These are not random samples -- they replay the actual sequence of returns starting from a specific year:

Starting YearWhat Happened
1929Great Depression -- catastrophic early losses followed by a slow recovery
1972Stagflation -- high inflation combined with poor stock returns for a decade
2000Dot-com crash, partial recovery, then the 2008 financial crisis -- a "lost decade" for equities
2007Global financial crisis -- severe drawdown with relatively fast recovery
2020COVID-19 crash -- sharp drop followed by rapid, unprecedented recovery

Crisis scenarios are valuable because they represent the actual sequences that destroyed real retirement plans. A strategy that survives the 2000-start scenario (which includes both the dot-com bust and the 2008 crisis within the first decade) has been tested against one of the worst return sequences in modern history.

The "Future Equals Past" Objection​

A common criticism of historical-based methods is that the future may not resemble the past. This is a fair concern, but it applies equally to all simulation approaches:

  • Monte Carlo also uses historical data. The mean return and standard deviation that parameterize a Monte Carlo simulation are estimated from the same historical dataset. The difference is that Monte Carlo then discards the actual distribution and replaces it with a bell curve.
  • Bootstrap preserves more information. By keeping the real distribution intact, bootstrap sampling retains the fat tails, skewness, and correlation structures that Monte Carlo throws away. If the historical data is our best available evidence about market behavior, bootstrap sampling uses more of that evidence.
  • No method predicts the future. Every simulation approach is a model. Bootstrap sampling is transparent about its assumption -- that the statistical properties of future returns will broadly resemble those of the past century -- and does not add the additional assumption that returns follow a specific mathematical distribution.

The question is not whether the future will exactly replicate the past. It is whether you would rather stress-test your retirement plan against the actual range of historical outcomes, or against a simplified statistical model that underestimates extreme events.

Academic Support​

The use of bootstrap methods in retirement planning research is well-established. Researchers who have advocated for or employed bootstrap and historical-simulation approaches include:

  • Wade Pfau -- Retirement income researcher and professor at the American College of Financial Services, whose work on safe withdrawal rates extensively uses historical return sequences.
  • Michael Kitces -- Financial planning researcher and practitioner who has written extensively about the limitations of Monte Carlo simulation for retirement planning.
  • Karsten Jeske (Big ERN) -- Author of the Safe Withdrawal Rate Series, which uses historical return sequences to analyze withdrawal strategies across hundreds of retirement cohorts.

Comparison Summary​

FeatureMonte CarloBootstrap Sampling
Data sourceHistorical mean and standard deviationActual historical returns (1928--2024)
Distribution assumptionNormal (bell curve)None -- uses real distribution
Fat tailsUnderestimatedFully preserved
Volatility clusteringNot capturedPreserved in crisis scenarios
Stock-bond correlationFixed estimateActual year-by-year correlations
Extreme eventsStatistically "impossible"Included -- they actually happened
TransparencyDepends on model parametersWhat you see is what happened

A visual comparison of return distributions would show the bell-curve shape of Monte Carlo (thin tails) versus the heavier-tailed distribution preserved by bootstrap sampling from actual market data.

Next Steps​