Friday, April 27, 2012

Choosing a Retirement Income Strategy

This post picks up from a line of thought started in “Variable Withdrawals in Retirement.” I’m going to present a simplified scenario for two different withdrawal rates and explain about some quantitative ways to answer the question: How do you decide which spending path is more desirable?
The classic 4% withdrawal rate rule is about constant inflation-adjusted withdrawal amounts. The benefit of this approach is that it provides a smooth and predictable income stream for as long as wealth remains. But wealth can run out. 

Another rule is to spend a constant percentage of the remaining portfolio each year. This allows for spending to increase when wealth grows. As well, though spending amounts may drop to uncomfortably low levels, this spending rule prevents wealth from ever running out. The main disadvantages of using this rule are that the spending path is unpredictable and spending fluctuates widely over time.

The simplified scenario I will discuss consists of one set of paths for the real income and real remaining wealth provided by these two strategies. This is one simulation out of 10,000 or so that might typically be part of a Monte Carlo simulation study on retirement income. I will consider the case for a 65-year old female. In this simplified case, she can see the evolving path from these two outcomes when she is ready to retire. She wants to know which is better. 

She retires at 65. A market boom in the first year of retirement allows her spending to rise from the initial level of 5. Her wealth grew from a retirement date value of 100 to nearly 140 in the first year. After year 8, her wealth begins to decline. Though the constant inflation-adjusted amount stays the same at 5, the constant percentage spending amount fluctuates quite a bit. In the 34th year of retirement (her 98th year), she uses the last of her wealth with the constant amount strategy. Should she make it to 99, nothing is left. With the constant percentage strategy, she still has income through at least 105 (40 years after retiring) that is somewhere around 30% of the initial retirement value of 5. 

The big uncertainty she faces is that she doesn’t know how long she will live. She has to choose her income strategy in advance without knowing this. Also, though the constant inflation-adjusted withdrawals can result in no income at some point in the future, we can suppose that while this is a very undesirable outcome, there is a safety net in place such that she does not necessarily need to absolutely prevent this outcome. How can she decide in a systematic way about which of these strategies to use?

Failure Rates

Traditional safe withdrawal rate studies focus on failure rates, often over a fixed retirement duration. For the constant amount strategy, failure is represented by wealth depletion. In this case, if the retirement duration is 30 years, her failure rate is 0%, but with a 40 year horizon her failure rate is 100%.  Trying to apply this same concept to the constant percentage strategy doesn’t really work though, since the strategy never runs out of wealth. It fails in the sense that spending is below the desired real amount of 5, but it succeeds in the sense that it always provides some spending power and some remaining wealth. Basically, we need to throw the failure rate concept out the window when considering retirement income strategies with variable spending amounts.

Average Lifetime Spending and Average Bequest Value

Retirees may like a strategy that provides the highest average lifetime income and the highest amount of remaining wealth to leave as an inheritance. To calculate these, we need to incorporate mortality data.

Average lifetime income is the average of income in each year of retirement multiplied by the probability of surviving to that age.

Average bequest is the wealth remaining in each year of retirement multiplied by the probability of that being the final year of the 65 year old’s life.

These survival and death age distribution probabilities, derived from the Social Security Administration Periodic Life Table from 2007, are shown in Columns G (survival probabilities) and Column H (death probabilities)  in the spreadsheet at the end of this post.

For the constant amounts strategy, average income is the sum of Column C times Column G, divided by the sum of Column G. It is 4.97. Though wealth runs out in year 34 of retirement, the probability of the 65 year old female surviving to her 99th birthday is 3.9%. Earlier income receives a greater weight because it has a higher probability of being achieved. Likewise, for the constant percentage strategy, the average lifetime income is 5.23 (sum of Column C times G, divided by sum of Column G). The extra spending early in retirement accounts for a lot.

As for expected bequests, for constant amounts, this is the sum of Column D times H, the wealth remaining in each year times the probability that it ends up being the final bequest (i.e. the probability of death). It is 73.97. As for constant percentages, it is the sum of Column F times H. That is 69.75.
The constant amounts strategy has a slightly lower average income, but a slightly higher bequest.

Spending Shortfall Below a Desired Floor Level

Average bequest may be helpful since it gives some indication about the surplus provided by the strategy, but average lifetime income may not be completely satisfactory.

A risk averse retiree may instead choose to focus on how often and by how much spending power may fall under a minimally acceptable flooring level. This floor may vary from person to person, and it represents the spending level that you really hope to avoid broaching because it may mean a serious reduction in your quality of life, such as a need to move to a cheaper home or give up some important spending.

To demonstrate this, I will use an income floor of 3. It is arbitrary, but our hypothetical retiree really hopes she can avoid experiencing spending shortfalls below the level of 3. Surely, she would like to be able to spend more than 3, and 3 is just her basic level of necessities. With this measure, we focus on downside and forget about the upside. Over the 40 year period, the total amount of spending shortfall below the floor of 3 with constant amounts is -18. There are 6 years with income of 0, which represents a shortfall of 3, and 6 x 3 = 18. The corresponding sum for the constant percentage strategy is -20.83.  If we add up the shortfalls weighted by the probability of surviving to the age of the shortfalls, we get a deviation with constant amounts of -.016, and the deviation for constant percentages is -.1051.  Constant percentages had more shortfalls earlier in retirement (the first shortfall happened in the 24th year), and there are higher probabilities for experiencing these shortfalls.
While this measure tells us about the downside, and in a sense it provides a modified form of failure rates that can be applied to variable withdrawal strategies, it does ignore any upside potential.

The Value of Spending (a.k.a. utility maximization)

A final measure I will look at provides an attempt to translate spending amounts into something that may more closely represent the value of that spending to the retiree.

A couple of issues will be at work here. First, more spending is better than less, but the additional value provided by more and more spending diminishes as spending increases. The additional value of increasing spending from 1 to 2 is very high, because this is helping you to fulfill same basic needs that vastly improve your quality of life. But the same 1 unit increase in spending from 9 to 10 may not provide much more value. That high spending might mean living in a home with 1800 sq. ft. instead of 1600, or it might mean 50 bottles of champagne per year instead of 30. The additional life-improving value provided by spending will surely be less as spending grows.

Another important aspect is that I will include the floor described in the previous section. Again, I will consider a floor of 3 as a reference point. Spending may fall below 3, but the value plummets for spending below 3. A spending drop from 4 to 3 is not nearly as bad as a drop from 3 to 2.

To discuss the value of spending, I will adopt the same basic function used by Joe Tomlinson in his February 2012 Journal of Financial Planning article, except I will reconfigure it to a new context. He calculated one value based on the lifetime outcome of how much wealth or total shortfall remained at death. I will translate each year of spending into a utility value and then add these up after multiplying them by the probability of surviving to that age.

In other words, this measure is similar to average lifetime spending, except that I don’t use spending from each year directly. Instead, I translate each year’s spending into a corresponding value based on the relationships seen in the following figure for two different loss aversion ratios (which measure the impact of how bad it is to fall under the floor of 3).

A few things to note about this: First, it will be hard to identify the precise shape of the function for real retirees. The shape depends on several parameters including one which explains the steepness of the curve, and also one which defines the appropriate floor level, and the loss aversion ratio that defines how bad it is to fall below the floor. More research is needed on this aspect. But we can still gain some insights by looking at outcomes. Also, note that this is only about spending power, and bequests are not incorporated into the calculations at all. But an important aspect here is that the measure does incorporate both downside risk and upside potential. Unfortunately, there is no connection between the value of spending in successive time periods. Real people may grow accustomized to a particular spending level, so that their frame of reference changes as spending changes. More research can eventually incorporate this aspect, as well as the possibility that the minimum acceptable floor could change over time. Finally, please keep in mind that the numbers defining the utility value are abstract and don’t have any direct meaning. What matters is the relative value of different spending levels, and not their absolute value. We don’t have to worry about utility being negative for levels below the income floor, it’s all relative.

This being said, we can calculate the “expected utility” of each spending path using this function which translates spending into a corresponding value, and then weighs these values by the probability of surviving to that age. With loss aversion of 2:1, the constant amount strategy provides expected lifetime utility of 0.387, while the corresponding value for the constant percentage strategy is 0.361.  Again, these numbers by themselves have no meaning. What matters is that 0.387>0.361 and the constant amount strategy provides higher expected utility in this case. With loss aversion of 10:1, a value function that more strongly punishes spending deviations below 3, the constant amount strategy provides utility of 0.364, which is more than the constant percentage’s strategy of 0.161.

As a final point about utility maximization, Michael Kitces recently wrote:

The reality is that utility functions are not new to the analysis of economic and financial problems, but these research papers are some of the first instances that we have seen in trying to apply utility functions in the context of financial planning, and potentially represent a significant step forward in determining effective solutions for clients. 

Putting it all together

For these two spending paths, we’ve considered 8 ways to quantify which strategy provides a more satisfying outcome. Failure rates were inconclusive, and the constant percentage strategy provided a higher average lifetime income. For the other six measures, the constant amount strategy performed better.

But understand that this is not the whole story. This was only one simulation. It doesn’t provide a final answer. We need to consider the outcomes for many more simulations, such as 10,000. Then we can look at the distributions of these measures and think more systematically about which strategy gives the best chance for the most satisfactory outcomes. 

We can also include more asset allocation and withdrawal rate choices, and more retirement income strategies such as partial annuitization with SPIAs or GLWBs, strategies which use buckets for different time horizons, the impacts of delaying annuitization, the use of bond ladders, the decision about when to begin Social Security, and so on.

Please stay tuned and keep reading, because all of this will be considered here in the coming weeks and months.


  1. Great info as always. It seems that 5%, either fixed or inflation adjusted, is a bit higher then what is usually recommended. I am curious how the results would change if a more typical withdrawal rate of 4% was used for the comparison of fixed % vs. inflation adjusted.

    1. This is an important question of course. The purpose of that blog post was to provide more in depth explanations about the different types of measures which can be used to compare retirement income strategies. I've been doing a lot of programming in the past week and I will indeed by discussing this in much greater detail.

  2. Two comments: First, it would be helpful to have a column for yearly investment return so the reader can see the pattern. It is possible to reverse engineer it, but that's unnecessary work when you must already have it. Second, to me one of the most obvious lessons is that one MUST rethink the withdrawal rule periodically. Unless this retiree has a very strong bequest motive (not suggested anywhere), at age 88, with over 50 units of wealth left, why should she let her withdrawals drop below her goal of 3? She has enough for 17 more years at that rate, so it would last until she is 105. Yes, that is too simplistic because she doesn't know future returns, but still ... Or at age 100, she takes only 1.5 from a portfolio of 30 that would supply her minimum goal of 3 for 10 more years?

    1. es, good points. About the returns, I could show those as well, but I think there may already be information overload. Especially as this showed just one simulation, but mostly I will be focusing on the results from many simulations.

      Your second point is particularly important. I've been looking at 3 different ways to deal with this: (1) a strategy in which the withdrawal rate is 1 divided by remaining life expectancy instead of always a fixed withdrawal rate, (2) buying a real SPIA to guarantee that spending will not drop below the minimum income floor, and (3) including a floor with the constant percentage withdrawals so that spending doesn't drop below the floor (unless wealth runs out). Any other ideas about this? Please stay tuned, I'm developing a lot more results to share.

    2. I have been thinkink along similar lines. One possibility would be for the retiree to buy an inflation-adjusted annuity paying 3 yearly at the beginning of retirement and then draw down the remaining money according to her lifespan. That's relatively simple to implement. Here's something a bit more difficult to implement, but that may be less conservative. Given a floor of 3, set a trigger point 10-15% higher. (I'm guessing at an amount; simulation would help.) Say we set it at 3.4. Then withdraw your constant 5% but the first year your withdrawal drops below 3.4, the trigger point, buy an inflation adjusted annuity for 3 (or as much as you can get). Because the annuity comes later, there will be more generous mortality credits and it is very unlikely that there won't be enough money to ensure an income of 3. You could even set the trigger at 3, but that might leave you short if you have a withdrawal of, say, 3.01 and then a big drop. Note that although these plans help with the downside, there should also be some way to harvest excess if the weath does extremely well. I haven't given much thought to that yet. Presumably, a 5% harvest rate is too low when a 20 year life span is no longer possible.

    3. Thanks. Yes, what you are describing I will also like to consider, but I need to add forecasts for interest rates to be able to do it. This sounds a lot like the Russell Investments Personal Asset Liability Model that I described a while back. I think this has a lot of potential.

  3. In some respects, looking at any withdrawal strategy with the word, "constant," in it is fighting the last war. Suppose I follow the sensible advice to build an income floor with pensions, annuities, and Social Security. For my optional expenditures, I'm highly unlikely to withdraw at any kind of regular pace; I'm probably not going to need or want a new car or boat or new kitchen every year.

    What this suggests is that rather than looking at annual withdrawals, it might make more sense to look at withdrawals over a longer interval; rather than an inflation adjusted 5% each year, how about a 15% withdrawal every three years or 25% every five years? I'm guessing that, by effectively averaging returns, that kind of strategy would support higher withdrawal rates, as well as providing a better match to actual spending patterns.

    1. That is pretty interesting, though it sounds tough to model. Because ultimately you will still withdraw some amount each year. What types of assumptions would you add to deal with this?

      More generally, one important point is that with any of these strategies you wouldn't necessarily need to withdraw the full amount in a given year if you happen not to need it. That would help sustainability, and in the context of the "value of spending" framework, it would just mean that the value of additional spending is zero so there is no need to spend it.

      ourbrooks, I think you are a Boglehead, right, and in the Bogleheads thread about the Scott Burns column, a lot of people were commenting that you shouldn't be required to withdraw more than you need. Of course not, and that fits fully into the framework when you define the "value" provided by additional spending.

    2. What I had in mind was periodic, larger withdrawals, with the money staying invested meanwhile. Every N years, take out N*X percent or dollars. For things like a new car or a new kitchen, people might not find it too difficult to delay the expenditure until the Nth year.

      I'm probably overlooking some vital complexity, but the calculations for this approach might be simple. Would it work to set the withdrawal rate to the cummulative periodic rate - X^(1+N) - but make the statement which subtracts from the portfolio value check whether the year of retirement is divisble by N and only do the subtraction if it is? The calculations would be similar for constant dollar withdrawals.

      I speculate that the benefit of this kind of periodic withdrawal might be sensitive to asset allocation, with riskier portfolios benefiting more.

      Yes, I read the Bogleheads forum and occasionally post there; I learn a great deal from my posts getting shot down.

    3. This does sound like more realistic behavior, but it might not end up changing the results all that much. Spending more one year but offsetting that with less the next year, and so on, might expose you a bit more to sequence of returns risk, but it is probably going to be hard to see much difference in the performance measures. But I'll keep this in mind, thank you.

  4. Variable Withdrawals: With regards to the stock portion of the portfolio one can do a) reverse dollar cost averaging (constant real withdrawals), b) dollar cost averaging (constant proportional withdrawals), or c) a sell high strategy (higher proportional withdrawal rate when real portfolio value increases or alternatively when PE10 is high, dividends only withdrawals when prices are low). Common sense and analysis of sequences of returns from 1926-2010 indicate that mean returns, maximum returns, and variability of returns all increase in the order of a, b, c. This makes financial sense as a constant real withdrawal rate is essentially a sell low strategy (more shares are sold in years when the market is down).
    With variable withdrawals one wouldn’t typically spend all the extra income during market peaks but rather pay down debts, add an aliquot of annuity, or stash it in a money market/short term bond fund/TIPS. This way the larger withdrawals during market peaks support later expenses.
    The question is: Can the extra anticipated average income from avoiding reverse dollar cost averaging (strategies b or c) be harnessed to provide a suitably stable and perhaps higher retirement income?
    One way that may work is to initially purchase a nominal annuity and then add to it when the portfolio does well (inflation and good luck adjustments). The nominal annuity allows initial income to be high without using the entire portfolio.
    Utility Function: Death is inevitable and one can approximate death probabilities. However, for an individual, it is prudent to always plan for living much longer than anticipated and never plan to fully go broke. The utility question should be: what is the highest spending rate that can be sustained for a period much longer than one can reasonably be expected to live. Actuarial calculations are for companies and governments that can account for real longevities. Individuals use maximum potential lifetime. Most people will estimate maximum potential lifetime as: (age at death of longest lived close relative + 10 or more years for improved medicine). For this reason safe withdrawal periods of various lengths (20 yr, 30 yr, 40 yr) are more useful to individuals than a mortality weighted utility function.

    1. Interesting stuff. Thank you for sharing. The first part requires more thought, but I would like to mostly focus on your last points with regard to using mortality data vs. planning for a fixed horizon.

      Actually, you've inspired me to write my newest blog entry on this subject. Thanks!

  5. Wade - thanks for the thoughtful analysis. Considering a proportional approach to withdrawals is really important, because it is more in line with actual behavior. People will spending more when they are ahead, and spend less when they are behind. As suggested by others, it would be really helpful if you considered expanding the universe of financing instruments to include immediate annuities - I think you will find that the results will be significantly better due to the mortality leverage that they provide. If you use a fixed immediate annuity, you can probably increase the % of the residual portfolio into stocks. In the analysis that I have done (I am a credentialed actuary), variable immediate annuities perform exceptionally well.

    One other point/question - I was not entirely clear on how you calculated the deviation amounts for when the income levels dropped below 3. I am glad that you weighted this for survivorship, but was not clear on whether you reflected this in your denominator when you converted these amounts into a percentage. I would suggest that this best be done with a denominator that is essentially equal to an annuity value that pays 3 per year with an interest rate of zero - as this is what would be a successful, survivorship-adjusted outcome.

    1. Thank you John. I do have fixed annuities and GLWBs programmed in now, and it is just a matter of time before I get around to writing up some results. It is nice to receive input from an actuary. Joe Tomlinson also visits here a lot. I'm definitely a very amateur actuary.

      About your question, I do think I am doing what you are suggesting, though I'm not positive. The denominator is just the sum of survival probabilities to each age, with no further discount factor included. So I think that is implicitly discounting with a zero real rate of return.

      I've also been looking at not multiplying by survival rates, but instead looking at the sum of real deviations over a given time, such as 30 years, without any further discounting.

      Thanks, Wade