The Safe Withdrawal Rate: Evidence from a Broad Sample of Developed Markets
33 Pages Posted: 28 Sep 2022
We use a comprehensive new dataset of asset-class returns in 38 developed countries to examine a popular class of retirement spending rules that prescribe annual withdrawals as a constant percentage of the retirement account balance. A 65-year-old couple willing to bear a 5% chance of financial ruin can withdraw just 2.26% per year, a rate materially lower than conventional advice (e.g., the 4% rule). Our estimates of failure rates under conventional withdrawal policies have important implications for individuals (e.g., savings rates, retirement timing, and retirement consumption), public policy (e.g., participation rates in means-tested programs), and society (e.g., elderly poverty rates).
The data cover approximately 2,500 years of asset-class returns in 38 developed countries over the period from 1890 to 2019...
The dataset construction methods deliberately combat the survivor and easy data biases that impact prior studies.
We incorporate longevity risk into the simulation design using mortality tables from the Social Security Administration (SSA).
Our base case simulation focuses on the joint investment-longevity outcomes for a couple retiring in 2022 at age 65 who chooses a portfolio strategy of 60% domestic stocks and 40% bonds.
The 4% rule proves woefully inadequate for current retirees.
Anyone who actually understands the details of this stuff want to break down the reasoning behind this for me?
Like coles notes "here's why things are different now" sort of summary?
The concern is that US-centric analysis are assuming repetition of US returns that may not repeat. So they're trying to pull in data from all over the world. To add a cherry on top, they're also looking at expected lifespan in the US. They concluded that 4% rule is riskier than advertised (and them running US-only as a side-by-side lends some credibility as those numbers come in where you'd expect). But I think the analysis is likely flawed, but also don't understand stats well enough to say for sure.
I find the methodology a bit strange - they've got this big list with 29,000 sets of monthly real returns from all these countries (Argentina and Czechoslovakia make up about 1% of these months each). There are only 39 countries, and they've got data going back to 1890 for some of them (US, UK, Germany, other countries you'd probably expect), and then starting more recently as countries joined the ranks of "developed". So the overall set is skewed towards the long-time developed countries, but because of how they're assembling their sets of returns for simulation, the countries with less return data are going to be included at a higher rate particularly at the end of ~10 year periods. To test a given withdraw percentage and allocation, they run 1 million simulations. Each simulation is as follows:
1. Simulate lifespan based on SSA tables
Some details are easy to understand - for any given month, they take that country's returns only and blend them per the allocation.
But then rather than simply randomly selecting returns to fil the simulated lifespan, they do this:
2. Pick a block size ~10 years long "randomly from a geometric distribution with a probability parameter equal to the inverse of the desired block length (120 months)".
3. Pick a random starting point out of the whole 29,000 country-months. Then from that month they include just the one country's returns going forward until either they fill the whole block or they run out of data from that country. If they run out of data on that country before the block is over, then they pick another country randomly from the set of countries and start adding that country's returns <starting from 1 - that's what it says in the paper>.
4. Add the block of returns to the overall returns for this simulation and then go to step 2 - stop when you've got enough months of returns for the entire lifespan being simulated.
I think the "pick a random country" part of it when you don't have a whole block of data is going to lead to overweighting the early returns from the countries with less data in the dataset (a few don't even have the ~10 years of data to fill up a single block). I think this is probably a substantial part of why the global figure they found is so low - overweighting Argentina from 1947-1966 or Israel from 2010-2019 (actually all return data stops in December 2019 for this study). These are countries that came into and out of "developed" status - they might restrict the list of countries to ones that have 1000 data points to make up for this maybe?
Spending is almost a cop-out to me "we're using real returns so we just assume a constant monthly withdrawal at the initial rate". I might be wrong, but not sure handling inflation this way is all that relevant - one simulation might have 5 years of Argentina real returns and we're also assuming we got Argentina's inflation for the spending for that time. Then we switch to Italy's figures or whatever. I don't have the math skills to figure that out but it just strikes me as potentially a problem.