JGinNISK wrote: Thu Sep 18, 2025 3:09 pm
I’ve often wondered about one aspect of the comments that the 4% withdrawal may not apply longer than 40 years.I rerun my monte carlo sims with the Flexible Retirement Planner often. In seven years my withdraw amount hasn’t substantially changed and my withdrawal percentage has slightly decreased, as my portfolio grows. At some point, the portfolio’s chance of growing is large enough over that time horizon that extending it longer doesn’t really cause any increase in the chance for it to fail. With a long time horizon, like 40+years it seems a more aggressive strategy actually provides more gains earlier, and increases the chance of crossing this threshold.
I can’t say conclusively but the fact that a more aggressive portfolio seems to have an improved likelihood of success on the N 30y rolling sequences known to date, may just be an anomaly of the historical sequences.
When I use a simple Gaussian model fit for stocks and a Chi-Square fit for bonds & cash, I get a result that is very close to the original Trinity Study (that used rolling 30y sequences, not a random draw from a distribution). However, this assumes stocks, bonds, and cash have zero correlation among them; that’s true for “all” data from 1926-2017 for stocks & bonds, but bonds & cash are probably correlated. Given those caveats, it’s still likely useful for a gut-check. As noted I get 4% for a 30y period with a 90% success rate (same as the Trinity Study), but when the period is altered, I get these results that differ from the 4% initial draw.
Table for 4% Rule Adjusted by Withdrawal Period (derivation link)
Some “models” don’t simply compare the the pass/fail metric on the N 30y sequences they have for a data set, they do a random draw from the index of historical returns, or a random draw from the distribution of the historical returns. I don’t care for a random index into the historical array of returns, compared to distribution modeling, as noted in this thread HERE.
I’m not wild about just using all the 30y rolling sequences as the test either since that seems to have similar limitations/anomalies (like the one you’re seeing for higher success with a more aggressive portfolio, which is counter-intuitive as noted in my post HERE). I have a post-graduate level understanding of statistics for my systems engineering background, but I’m no finance major nor am I a CFP or PhD in Finance/Economics like Drs. Fama & French, so I know a simple Gaussian model with no correlation to stocks nor bonds is likely inadequate, but it does fit the data from 1926-2017 (the data set I settled on when building my models), and up until the discrepancy with FiCalc’s rolling sequence approach it gave reasonably similar results to other models (e.g., FiCalc and PortfolioVisualizer’s Financial Goals). Note that for FiCalc’s data set there are 125 rolling sequences so the pass/fail metric is M successes / 125 trials and 125 is a pretty small sample size for reasonable confidence & power on a binomial proportion (another reason I don’t care for N rolling 30y sequences).
