Recently I’ve read up a lot on Safe Withdrawal Rates (SWRs). For those that haven’t come across the term, it’s a method used to find how much a retiree can withdraw from their portfolio of assets each year without running out of money before reaching the end of their life. The most famous SWR is the “4% rule”, which, in many respects, is a cornerstone of the FI Community.
A crumbly cornerstone?
Calculating a suitable SWR can be an important part of understanding how an individual can fund their income needs in retirement. Those eagle-eyed may spot that I used the term ‘can be’. At this point, I’ll warn the reader that I’m going to take a jump forward and delve straight into the nuts and bolts of calculating Safe Withdrawal Rates. If you’re new to this subject, I’d recommend reading a little more background on SWRs. I’ve recommended some helpful articles at the end of this post. Going back to ‘can be’, the reason SWRs may not be helpful is because: (i) predicting the future is hard; and (ii) the calculations have strong limitations.
Before I dive in
I want to make it clear that none of this post should be construed as a personal attack on the authors of various Safe Withdrawal Rate studies. It also doesn’t mean I think they are wrong. Most of the authors are far more qualified and far brainier than me. I also haven’t read every book, article or piece of literature on the subject. There is a huge body of work out there, and its not possible to read it all. So it’s entirely possible I’ve missed things (if I have please let me know in the comments). My aim, in sharing the following criticisms is to help readers understand the important limitations in using SWRs and to know when and how to take these calculations with a pinch of salt.
How you calculate a Safe Withdrawal Rate
Typically speaking there are two methods of calculating a Safe Withdrawal Rate:
- Historical Data approach – in this approach you use real historical data to calculate investment returns and other statistics required to generate a SWR. This is sometimes also called a ‘backtesting approach’.
- Monte Carlo approach – in this approach you use a Monte Carlo or other model to simulate a large number of investment return scenarios to test a SWR based on a set of assumptions. This is sometimes also called a ‘simulation approach’.
I strongly favour the Historical Data approach. Whilst I won’t go as far as saying Monte Carlo approaches are useless, for the specific purposes of SWR projections they have a huge fundamental flaw which makes the results from such analysis tenuous. [Note, in general Monte Carlo simulations are a great analysis tool, but for the specific reason I set out below be vary wary when they are used to model investment returns. There is also a third, hybrid, method called “bootstrapping”, which is a combination of a Historical and Monte Carlo approach. I don’t cover it here. But it too has some conceptual problems.]
Why the Monte Carlo approach is fundamentally flawed
To create a Monte Carlo simulation you need to create a set of assumptions. These are used to generate your future scenarios. The biggest flaw in the approach is that you need to make an assumption about future investment returns. The reason this is such an issue is that stock market returns do not follow a normal distribution. If stock market returns don’t follow a normal distribution, what do they follow? The answer is: we don’t know; or: it’s up for debate.
Back in the 1960s Benoit Mandelbrot, a mathematician and polymath, suggested returns follow a Levy Distribution. In 1963, future Nobel Prize winner Eugene Fama, in his first published paper, examined Mandelbrot’s work, and considered that equity returns follow a type of stable distribution. Research in the following 50+ years has concluded that equity returns don’t follow a Levy Distribution, rather they either follow some other Stable Distribution or, as is more commonly thought, some kind of Power Law Distribution. To put that in common parlance, equity returns have “fat tails”.
The most important thing about all these distributions is this: you cannot calculate an average. The reason for this is that under such distributions no value is impossible. You could have returns of 5% per year then BAM! the market drops 23% in one day. These huge outlier values make a calculated average on historical values meaningless – they will massively shift the next time there’s a huge market movement.
That’s why Monte Carlo approaches for calculating SWRs are flawed. The approach relies on an imperfect assumption of market returns, using an average and standard deviation that we know are incalculable. What makes it worse is that for calculating a SWR we are looking precisely for the points of failure – the so-called ‘black swans’. These one-offs can be one of the largest determinants in whether a SWR strategy works or not.
So if we’re ruling out using the Monte Carlo approach, that leaves us with the Historical Data approach. Let’s look at some of its limitations.
Limitations of Historical Data based Safe Withdrawal Rates
1. Successful SWRs are ill-defined: The thin line between success and failure
The first limitation is arguably the most abstract but, for me, the most important: How successful decumulation is defined. Success will vary from person to person. For most people, success is having enough income to cover living expenses and support an adequate standard of living level. Failure is the opposite, having insufficient income such that a retiree’s standard of living falls. But we are already adding vague terms into our definition. For example, what is an “adequate standard of living”? is there an “acceptable variation” in standard of living?
Using a proxy
It’s for this reason that SWR calculations typically use a proxy to this definition: success is not running out of money. This is much easier to calculate – you look at the balance in the bank – but it does mean some element of “true” success is lost. It’s possible to have zero or even negative net worth and still have an income. Conversely, it’s possible to be a paper millionaire but be unable to draw any of the wealth as income. But as using an income definition for success is very challenging, we have to fall back on the next-best which is using portfolio value as a gauge of success.
Adjustments to the SWR
Another element to success when calculating SWRs is the rate at which a SWR would not have reduced a portfolio to zero. Some calculations consider 90% (or lower) as an acceptable rate of success. Others would shiver at the thought that 1 in 10 times you’d end up broke. An answer to this is the SAFEMAX (or MSWR, Maximum Sustainable Withdrawal Rate): the highest historical rate that would not have run out of money.
Alternatively, a stricter definition is what the author at Portfolio Charts calls the Perpetual Withdrawal Rate: the rate which preserves the inflation-adjusted principal in the worst decumulation time horizon. However, this definition leaves us with a question: are we happy for our pot to dwindle to zero? or do we want our pot to stay the same (or grow)? (and potentially be passed down through inheritance). The answer of course will depend vary for each person. But neither of these definitions appear attractive. A portfolio inevitably falling to nothing creates a potential risk of running out of money. A portfolio where the principal stays the same (or grows) may seem like a waste, resulting in missed opportunities (FOMO).
There are lots of different definitions used to define successful decumulation when calculating SWRs. But none of these will be the ‘true’ definition of success for each retiree. Each has its own limitations which we should be aware of.
2. Inflation: Expect the unexpected
The second limitation comes from the assumptions we have to make about inflation. These pop up in two places: (i) investment returns; and (ii) future expenses.
Dealing with investment returns first. The main reason we invest is to protect against inflation. One of the big issues with some (not all) SWR calculations is that they use nominal returns. That is, they do not account for inflation eroding the value of the principal. Therefore, calculations that use real returns should be favoured. But there is an issue with this, and we’ve already said it: we invest to protect against inflation. By using the returns on investments that people have historically used to best inflation we are baking in an assumption that they may continue to do so in the future.
The trouble is, we can’t do much about it. That’s because most modern economies use Monetarism. This is where the Government controls money supply to try keeping prices stable. An underlying tenant of Monetarism is that it is impossible to know (or expect) inflation in advance. If inflation was predictable then Governments would not be able to affect inflation by controlling the supply of money. So in principle, if we fiddle with our inflation assumptions (from taking the guidance targets of the relevant central bank) we are making ‘bets’ on inflation that are not knowable. That’s fine, but can you predict the impact with any accuracy? We’ve seen with Quantitative Easing (QE) that predicting how inflation and asset prices react to money shocks is very difficult (impossible?) We can also see another issue in the graph below:
Since the early 1980s most Western Economies adopted Monetarism and have been successful in “controlling” inflation. But most SWR calculations will look at investment returns stretching back 100 (or more years), in periods where Governments did little or nothing to control inflation. How relevant are asset prices and returns from periods that had high inflation (and deflation) compared to our current (and future?) low and stable inflation environment? The answer is I don’t know, and even people far smarter than me don’t really know the answer.
In terms of future expenses, we’ve already touched on one issue with inflation that creeps in here: that is, future inflation is unpredictable. When withdrawals are deducted from the investment portfolio in a SWR model, these are usually done in a lump sum at the start of the year (month). But this implicitly assumes that the retiree was able to accurately predict living expenses in advance (for this reason, I prefer a calculation using monthly figures). For example, living from the 1970s trying to predict daily fuel costs was a nightmare.
EREVN summarises this issue better than I can:
When looking at simulations the difference between 92% success and 96% success (just to make up two numbers) is probably meaningless when there’s a 75% chance you have some unplanned spending shock that the simulation isn’t taking into account.
There’s one final issue with inflation. The calculations typically use aggregate measures of inflation – such as CPI. But these are likely to vary significantly to your own personal inflation and even to that of your cohort. Both EREVN and Karsten at Early Retirement Now have looked into this. But the trouble is, it’s very difficult to change the inflation rate baked into a calculation without making hidden assumptions (what your behaviour would be in certain economic scenarios; what relationship between your personal inflation and asset returns). We can calculate our personal inflation by tracking our expenses – but this is imperfect. We can also try to adjust using different measures of inflation (such as CPIH which includes housing costs. But, to be honest, I feel these are delving into an element of mathematical precision which doesn’t exist.
The Historical Data method requires implicit and explicit assumptions on inflation. Within these assumptions we are assuming that some element of inflation is knowable even though this is not the case. Inflation shocks can, and do, happen. We should be wary of false precision due to the unexpected nature of inflation and our imperfect estimates for it.
3. Historical data: An oldie but not a goodie
The third issue is that the Historical Data approach needs to use actual investment returns. The trouble is, we don’t have much of them. In fact, our sample is even smaller than you might think – remember that investment returns day on day, year on year are not independent of one another. As Karsten at Early Retirement Now puts it:
Strictly speaking, there are just under 3 truly independent samples of 50 year return streams: 1871-1920, 1921-1970, and we are charitable with the last one, 1966-2015, by ignoring the little bit of overlap in the first 5 years. So, if someone tells us their strategy has worked 96 out of 96 times, we would be impressed. But if someone tells us their strategy works 3 out of 3 times, we would be suspicious about that 100% certainty claim.
Unfortunately, this isn’t the only issue with historical data. Another issue is patchy data – missing in places – or we have to thread together data sources that aren’t strictly the same to create a continuous data source.
A third issue is that is, old market data is not very good for telling us about returns today. The four most well-known long-term data sources are:
- Ibbotson SBBI which goes back to 1926 (Ibbotson also ‘created’ some further sets going back to 1815 and 1825)
- Shiller/Yale which goes back to 1871
- CRSP (Fama French) which goes back to 1926
- Dimson, Marsh, Staunton/Credit Suisse/Triumph of the Optimists which goes back to 1900
There are a number of issues with data as we go back in time (lots of these issues are heavily discussed, so I won’t go into too much detail, EREVN has a good primer):
- There’s very few companies in some of the sets
- Lots of stocks are missing because not every exchange was counted
- The composition of industries was wildly different compared to today
- The data sources get patchy, there weren’t computers, so data was cobbled together from lots of sources
- Finally, the world was a very different place!
What this means is that the further we go back in time the dodgier the data. As a rule of thumb I get nervous when data is from before 1980 and I avoid data from before 1970 if possible. Unfortunately, this drastically reduces the amount of data we have to work with. As such, we need to give a wide berth to analyses that can tell us historical equity returns to 1 or even 2(!) decimal places. Remember: “there is uncertainty surrounding every estimate from those simulations!”
4. Withdrawing: Taking your money and running
There are roughly speaking three types of withdrawing methods:
- Fixed – a set rate, or set amount is withdrawn each year, adjusted for inflation.
- Variable – withdrawals vary depending on the circumstances.
- Valuation-based – withdrawals vary depending on certain valuation parameters.
In my opinion, a Variable withdrawal method should be preferred.
Firstly, whilst the most common in various studies its difficult to see how a Fixed method reflects reality. I think most people would struggle to keep spending constant year-on-year if there was market turmoil or a long-bull run.
Secondly, such a method forces a low withdrawal rate to account for the worst possible market scenarios. Not to mention, that it would be irrational for a person to continue to withdraw at fixed rate forcing themselves into destitution. A retiree would, if necessary, lower their withdrawal rate. But the Fixed method takes no account for this.
Finally, in a good case scenario, a Fixed method forces a retiree to stoically stick to a fixed withdrawal and prohibits them from enjoying any upside in investment returns. It feels philosophically wrong for a retiree to not enjoy the fruits of their labour (and arguably is a failure as the retiree ‘enjoys’ a sub-standard level of living).
Valuation-based methods involve increasing or reducing withdraws based on market valuation metrics. Most commonly CAPE (or CAPE10). I don’t want to write too much on CAPE because it is discussed enormously in the Finance community (and Prof. Shiller’s work is, by any measure exceptional), but there are several significant problems for using CAPE in a SWR calculation (I appreciate I’m very likely to be in minority on this, but bear in mind it’s just my opinion! If you’re aren’t familiar with CAPE ignore my boring bullet points below):
- CAPE has poor predictive power – whilst, in the best situations, a higher CAPE value gives a slightly higher probability of lower returns, the predictive power is mixed.
- CAPE is not very ‘actionable’ – a CAPE value of “x” tells us little about how much to reallocate between stocks and bonds or the likelihood of a market crash, we have to superimpose our own reallocation rules on top of it to make it actionable and this potentially leads to over-fitting or data mining.
- Where CAPE does have predictive power, it is over longer periods of time 10+ years. This is a significant portion of any retirement period, and it would take some brass balls to keep under-withdrawing (in a low valuation bear market) for 10+ years waiting for CAPE to come good.
- The data CAPE is based on is patchy data (we’ve talked about dodgy historical data above). To add to that, earnings have changed over the years as accountancy rules have changed. Earnings in 1970 are not the same as in 2018 but for the purposes of CAPE they are.
- CAPE is noisy – even where CAPE has predictive powers, the range of potential outcomes is often very wide.
For those reasons, I conceptually prefer a Variable Withdrawal method. Being honest, I haven’t reviewed all the methods (there are lots of them), so I can’t (and I don’t think anybody can) say which method is ‘right’ or ‘best’.
5. Death and taxes: There are only two things certain in life
And they’re both missing in SWR calculations! Taking taxes first, SWR calculations almost always exclude taxes – that’s because it is very complicated to model tax charges and these charges are highly dependant on the individual and their circumstances. What this means is that a SWR before tax is likely to represent an upper bound to what your SWR is in practice.
Another big thing missing from most SWR analyses is death. What most analyses do is set a ‘retirement period’ typically of 30+ years, and then run the calculations. But by doing so this makes the assumption that end of retirement (death) is knowable. There are two big risks here: (i) we live longer than the estimated life span; and (ii) we don’t account for the likely changes we might make depending on our ongoing view of mortality.
Taking (i) first, this suggests we should consider using a longer ‘retirement period’; for my money 30 years seems very short, even for those reaching retirement age. Therefore, we should prefer analyses that use longer (such as, 60 year) retirement periods. It also gives credence to considering a Perpetual Withdrawal Rate; this implies an infinite retirement period.
Secondly, it informs us that analyses with shorter periods are more likely to over-estimate a SWR. However, the longer the retirement period the more old data we have to use – causing a trade off. Finally, a very long retirement period or Perpetual Withdrawal Rate can have the counter-intuitive result of giving a higher SWR. This is because over long periods of time the power of compounding rockets up the value of the investment portfolio. But this is no comfort for a retiree who may have to survive 20/30+ on a meagre withdrawal rate before things ‘come good’.
On (ii), we can think about some of the issues that we discussed under point 4 Withdrawing. If we knew we had only a few years to live would we change our behaviour? I think for a lot of people the answer is yes – we’d probably enjoy our money a bit more whilst we are with our loved ones. However, if our life expectancy slowly increased would we slowly reduce our spending to compensate? You might be saying “Yes I would!” But research on the general population generally finds we are poor at self-discipline. What this means is that as the length of our retirement period increases, our ability to support a withdrawing regime becomes fuzzier.
There are two further things that are missing from some SWR calculations and these are: (i) Social Security/Pensions; and (ii) Fees. I comment on these briefly, because these can and should be factored into the calculations. The possibility of state provision in retirement reduces our income needs and therefore reduces the amount we need to withdraw (i.e. increasing our SWR). On the flip-side, investment fees are a drag on portfolio returns and reduce the SWR.
So what can we do about it?
In the main I think we should still use SWRs to prepare for retirement. Now before you go crazy that you’ve read 3,000 words for nothing, let me quote Winston Churchill:
“it has been said that
democracy isSWRs are the worst form of Governmentdecumulation planning except for all those other forms that have been tried…”
Even though there are lots of issues with various parts of calculating a SWR that doesn’t mean we should scrap the concept. In fact, the SWR is helpful if we are conscious of the limitations and risks that come along with. In that respect, we should use SWRs in our planning, but not necessarily use them for concrete plans (or as former Dwight Eisenhower put it: “In preparing for battle I have always found that plans are useless, but planning is indispensable.”). So what does that mean in practice?
1. Use various Safe Withdrawal Rate analyses as indicators
Each SWR analysis has its own merits, but is likely to understate or overstate the appropriate rate in a certain direction. For example, calculations using a 30 year retirement period are likely to overstate the SWR. A very long retirement period may understate a SWR. In this way we can triangulate a range of SWRs. We can get a rough idea of what the bookmarks are for an appropriate SWR based on a set of circumstances.
2. Beware false precision
Given all the limitations it’s not mathematically possible to accurately assess a SWR to a high level of precision. Beware analyses that say results such as “our calculated SWR is 4.03%”. This is false precision.
The error bars around each spot estimate of a SWR are substantial, so when we are thinking about rates we should be ditching the second decimal point, and arguably even the first. A more suitable result for a SWR analysis could look something like: “3¼ ± 1%”. This means testing your decumulation strategy on a band of rates. The error bars around a spot estimate of SWR can be substantial (depending on the study). So when we are looking at a particular SWR we should consider whether it is possible to estimate that rate to the precision of two decimal places (and for some studies even one decimal place). To that end, it is more prudent for a result of an SWR analysis to give a defined range, something like: “3.25% to 3.5%” or “3.5% +/- 0.5%”. The breadth of that range will depend on the quality of the underlying data, and the statistical noise around the spot estimates. However, you can use the strategy under point 1, to narrow down wider ranges to improve their personal suitability. For example, say the range is 3.5% to 4.0% with 3.5% being commensurate with a 50/50 stock/bond portfolio and 4.0% a 80/20 stock/bond portfolio; but your personal risk tolerance is towards a higher bond allocation. Then aiming for something closer to 3.5% would be prudent. With that in mind, its worthwhile testing your decumulation strategy across a band of rates.
3. Be flexible
Where things are uncertain, we should counter-act this by being more flexible. Our behaviour can and should adapt as information and circumstances change. That doesn’t necessarily mean working a side-hustle in retirement (which can be a bad idea) but a worthwhile attempt might be acquiring skills that never go out of demand or skills that are at their most valuable during periods of economic downturn/low returns. An important enabler of flexibility is insurance. Insurance seems to be less discussed in the FI Community, self-insurance seems very popular. But having the right insurance policies can be an excellent way to protect against life shocks. Finally, whilst some may decry me for being wishy-washy, being flexible in your life outlook can also help add security to your withdrawal method. An example is being open to moving home (aka geo-arbitrage).
As I mentioned at the start, none of this should be read as a personal criticism. In fact, there has been a huge amount of great work done by lots of very smart people. But its important that we know the limits to the work done so far. All predictions of the future need a big health warning. Its important that you don’t blindly take a SWR from a study as a fact. You should mould a SWR to your own personal circumstances. I want to end with the following quote from Portfolio Charts (which was a very instructive source for this post) that encapsulates this post perfectly:
Just because something did great in the past does not mean it will continue to do so on your own personal time-frame. I believe withdrawal rate research is a wonderful way to help set financial goals and guidelines, but one should never put their life savings in the hands of a single back-tested number. Flexibility, intelligence, and determination will beat mechanical withdrawal rates every time!
Thank you for reading. Please share your thoughts (and criticisms!) with a comment.
All the best,
Young FI Guy
Portfolio Charts, an excellent site with explainers on SWRs and various calculators:
- Withdrawal Rates Methodology – https://portfoliocharts.com/withdrawal-rates-methodology/
- How Safe Withdrawal Rates Work – https://portfoliocharts.com/2015/11/17/how-safe-withdrawal-rates-work/
- Withdrawal Rates Calculator – https://portfoliocharts.com/portfolio/withdrawal-rates/
- How to Predict Withdrawal Rates Without a Crystal Ball – https://portfoliocharts.com/2017/03/21/how-to-predict-withdrawal-rates-without-a-crystal-ball/
Early Retirement Now, probably the number one source in the FI community on SWR research. Kartsen has done some amazing work, most notably his 23(!) part series:
- The Ultimate Guide to Safe Withdrawal Rates –https://earlyretirementnow.com/2016/12/07/the-ultimate-guide-to-safe-withdrawal-rates-part-1-intro/
- The SSRN White Paper on SWRs: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2920322
- He has also joined the Choose FI podcast twice, I highly recommend listening to both:
EREVN on Medium, has also done some great analysis on SWRs: https://medium.com/@justusjp
Retirement Investment Today, was one of the main catalysts for me achieving FI and thinking about SWRs from a UK perspective:
Living Off Your Money – Michael McClung – I’ll confess I’ve only read the first three chapters, but on the basis of Monevator’s recommendation I did buy the book at the weekend and plan to read the rest of it.
[Note: the post was edited on 02/05/2018 on the “false precision” point following feedback from several readers. I felt that the example I was giving was being unhelpful and suggesting that SWR analyses are noisy to the point of being useless. As I set out in the article, I don’t think that’s the case. SWRs can be a very useful tool if they are tailored to your personal circumstances. But we should be conscious that we can’t boil it down to one exact figure. Using a range, narrowed down through applying your situation to each analysis, is prudent way to get more mileage out of a SWR figure.]