Welcome back to the 20th installment of the Safe Withdrawal Rate series. Check out Part 1 to jump to the beginning of the series and for links to the other parts! This is a follow-up from last week’s post on equity glidepaths to address a few more open questions:
- Some more details on the mechanics of the glidepath and why it’s so successful in smoothing out Sequence of Return Risk.
- Additional calculations requested by readers last week: shorter horizons, other glidepaths, etc.
- Why are my results so different from the Michael Kitces and Wade Pfau research? Hint: Historical Simulations vs. Monte Carlo Simulations.
So, let’s get to work …
More on the glidepath mechanics
In last week’s post, we got a bit ahead of ourselves, simulating glidepaths without digging deeper into the intuition for why a glidepath should cushion the effect from Sequence of Return Risk. So let’s look at a simple case study to understand the benefits of an upward-sloping equity glidepath in retirement:
- A 10-year horizon, withdrawals are made annually at the beginning of the year. The initial portfolio value is $1,000,000, the initial withdrawal is $35,000, which is then increased by 2% every year to keep up with inflation.
- We look at one glidepath from 70% equities to 90% and one fixed 80% equity allocation.
- In the first case study, equities drop by 30% in year 1, then another 5% in year two before starting another nice 8-year-long bull market. Also, notice that the bond market returns are modeled to reflect a negative correlation with equities!
- The rebalancing to the target weights occurs every year at the same time as the withdrawals. In other words, post-withdrawal the portfolio displays exactly the target weights.
Let’s look at how the (nominal) portfolio values, withdrawals, and the rebalancing evolve over the ten years, see table below. The top panel is for the glidepath, the bottom panel is for the constant equity share.
- With the glidepath, you actually withdraw much more from bonds, especially during the first few years. Over 60% of your total withdrawals during the first 10 years come from bonds. On the other hand, with the fixed equity weight more than 85% of your withdrawals come from equities. That’s even higher than the target equity share! Because you withdraw so much more from equities while equities are cheap the fixed asset allocation is more exposed to Sequence of Return Risk.
- The benefit of the glidepath comes from the fact that we not only plow money into equities on the way down (two years of negative withdrawals!). But during the bull market, we withdraw only about $25k p.a. from equities and the rest from bonds! That gives the equity portfolio more room to enjoy the bull market!
- Compare that to the withdrawals with the fixed equity shares: You withdraw about $17k from the equity portfolio at the bottom of the stock market (ouch!) and then during the bull market, you withdraw more than the necessary consumption level from the equity portfolio to replenish the bond portfolio every year. During the bull market, the equity weight is constantly dragged above its target. Thus, you hamper the recovery of your portfolio when you constantly shift away from the well-performing asset (equities) and into the relatively low-performing asset (bonds).
- Also notice that after two years, the glidepath beats the static allocation by about $42k ($733,314 vs. $691,746). After ten years that gap has increased to almost $80k ($1,074,558 vs. $995,378). Almost half the advantage of the glidepath came from the bull market that followed the drop!
Of course, if the returns were to occur in the opposite order – a continued equity bull market eight years and then a crash at the end – results will look quite different, see table below:
- All assumptions are the same as before. I only reverse order of returns.
- Now the glidepath performs worse than the constant equity weight. But that’s expected: Because you start with only 70% equities you participate less in the bull market and you have the highest equity share when the market falls in years 9 and 10!
- Of course, even though the glidepath underperforms the static allocation ($1,089,990 vs. $1,162,099), you are still better off than with the glidepath when the bear market hits you in the first two years ($1,089,990 vs. $1,074,558)!
To summarize the case study results, let’s look at the final values for the glidepath and the constant asset allocation, see chart below. The variability of final asset values is lower with the glidepath. True, you underperform the constant 80% equity portfolio when you have a long bull market early in retirement, but the glidepath performs significantly better when it really counts, i.e., when there’s a bear market during the first two years of retirement!
Back to the historical simulations: more glidepaths
The table below is almost the same as last time, but with a few changes:
- I added eight more glidepaths. The first is inspired by the work of Michael Kitces who, relying on the Monte Carlo simulation study with Wade Pfau (see Table 6), suggested a 30 to 70% equity glidepath over 30 years, which optimized the success probability of a 4% Rule using historical average returns. So I used that glidepath (30->70% with a 0.111% passive slope). But I also use glidepaths with larger slopes (0.2%, 0.3%, 0.4% per month) and the same for a lower starting and end point (20% -> 60%).
- Instead of high CAPE vs. all CAPE scenarios, I split the percentile stats into high CAPE (>20) and low CAPE (≤20).
Results:
- When the CAPE is below 20, there is no benefit from a glidepath. Any 90-100% static equity weight will give you the highest, or at least close to the highest fail-safe withdrawal rate. The same is true when targeting slightly higher failure rates (1%-25%).
- But glidepaths are useful when equities are expensive (CAPE>20), as we already saw last week! The 60 to 100% glidepath had consistently the best withdrawal rates for all failure probabilities studied here. 40->100% and 80->100% are close behind. The active vs. passive glidepaths and the exact slopes don’t make that much of a difference if you get the correct start and especially the endpoint (100%!) of the glidepath.
- Quite amazingly, the glidepath recommended by the Kitces and Pfau study (30 to 70%) is consistently one of the worst. It not only underperforms pretty much all of the other ERN-designed glidepaths. It’s actually so bad that it even underperforms most of the static asset allocation paths in the historical simulations! At first, I thought this is because of my 60-year horizon, but as we will see in just a minute, the Kitces and Pfau glidepath is pretty universally inferior, even over a 30-year horizon!
How about a shorter retirement horizon?
Glad you asked! Here’s the same table but with a 30-year horizon:
- The same result as before: Glidepaths are of no use when equities are cheap to moderately valued (CAPE below 20).
- Notice how among the fixed equity weights, you achieve the most attractive SWRs between 65 and 75% when the CAPE is above 20, a bit lower than the 75-80% optimal equity weights over the longer horizon. But when the CAPE is below 20 you’re still better off using 100% equities, regardless of your failure probability!
- The glidepaths that did best over 60 years, moving from 60% to 100%, are still consistently very good performers. Quite intriguingly, the 40->100% glidepaths are now even slightly better!
- The Kitces and Pfau glidepath is still one of the worst performers. Both in its original form (slow transition over 30 years) and with faster transitions. The 20->60% specification is even worse.
Failure Rates of specific SWRs
Another way to slice that data. Instead of targeting a specific failure rate and then calculating the withdrawal rate, we can also look at the different withdrawal rates between 3 and 4% and calculate the failure rates, see table below:
- I do this only for the high CAPE regime (>20) to save space.
- Notice the unacceptably high 60-year failure rates for the 4% rule!
- Also, notice that the failure probabilities are lower with the glidepaths but the effect is only marginal. For example, even with the “best” glidepath will not miraculously rehabilitate the 4% rule. All you can hope for is to make the 3.5% rule a lot more secure!
Higher Final Value Targets
As requested by a reader last week, here’s the table with SWRs targeting specific failure rates but for different final value targets and using fail-safe and 1-5% failure probabilities. The reader asked for 1% steps, but I report only fail-safe, 1%, 3% and 5% to save space. If you want the 2% and 4% SWR percentiles you simply take the midpoints!
Results are roughly the same. But I noticed that the benefit of the 60-100% glidepath goes up vis-a-vis the static allocation. For example, the fail-safe SWR improves by 0.22% (3.47% vs. 3.25%) under capital depletion. But it improves by 0.29% under capital preservation (3.34% vs. 3.05%). Again, it doesn’t miraculously make the 4% Rule viable again but you’ll get a noticeable improvement in the sustainable withdrawal amounts!
Why do I get different results than Michael Kitces and Wade Pfau?
First, I thought that the main driver was the shorter retirement horizon in the Kitces/Pfau paper (30 years). But I showed above that even over 30-year windows their proposed rule, 30->70% linearly over 30 years (=0.111% monthly steps) is consistently one of the worst glidepaths. You can improve it a little bit by accelerating the pace of the glidepath to 0.2, 0.3, or 0.4% monthly steps, which gets you to 70% equities after 200 months, 133 months and 100 months, respectively. But even those glidepaths stink compared to some of the other paths that I proposed. They are even worse than some of the fixed asset allocations. What’s going on here?
The major difference between my work and the Kitces/Pfau study is that I use historical returns and they use Monte Carlo simulations. How can that make such a big difference? In my view, there are (at least) three features of real-world return data that are impossible to replicate with a Monte Carlo study a la Kitces/Pfau:
1) Short-term Mean Reversion: After each major drop in equities, we are bound to observe a strong recovery, see, for example, our post on the 2009-2017 bull market from a few months ago. The theory is that investors overreact on the downside (remember March 2009?) before a nice new bull market ensues. A Monte Carlo study will not replicate this feature. A Random Walk means that returns have no memory, i.e., the distribution of returns going forward after a 50% drop is the same as after 50% gain. But with real-world data, you’d benefit from a glidepath with a much steeper slope to better capture the bull market that will likely follow the initial drop. Remember, in the first year after the 2009 trough, the S&P500 went up by 72.3% (nominal total return, March 9, 2009, to March 9, 2010)!
2) Long-term Mean Reversion: The non-random-walk nature of equity returns is even more pronounced if we look at longer windows, say, 15 years. In the chart below, I plot the average annualized real S&P500 return over two consecutive (neighboring) 15-year windows. Notice the negative correlation? If the previous 15-year return was poor then the next 15 years had above-average returns! This has profound consequences on the glidepath design: It’s the main reason why the glidepath has to shift to its maximum much faster than over 30 years and it’s also the main reason why in the historical simulations, the preferred long-term equity weight is 100%.
If you get unlucky during the first 15 years of your retirement due to poor equity returns you benefit greatly from going “all in” during years 16-30 of your retirement!
In fact, that might be the only way to salvage an underwater portfolio that has been taken to the woodshed due to bad equity returns and 15 years of withdrawals. If you base your optimal glidepath design on Monte Carlo simulations you’ll find much lower optimal long-term equity weights!
3) Correlations: Kitces and Pfau have to pick one single stock-bond correlation in their Monte Carlo Study. However, in real-world return data, this correlation has been all over the map during the last few decades. We’ve had the 1970s/early-1980s where the correlation was strongly positive (both stocks and bonds lost value), but we also had the 2000s onward where stocks and bonds had a strong negative correlation and bonds were a great equity diversifier. The optimal glidepaths calibrated to that one single correlation are clearly suboptimal when using historical data.
What now?
I’m the first to admit the weaknesses of working with historical return data. We don’t know what the future holds. CAPE ratios are hard to compare over time, and I can come up with theories for why the returns going forward can be much more attractive than in the past. But I also have a theory for why they could be worse. So, using the historical simulations as a midpoint to gauge average returns is not a bad starting point.
In my personal view, a Monte Carlo study for retirement glidepath design is the worst of all worlds. You still have to make an assumption about future mean returns and there is no telling whether that assumption is better or worse than the historical return assumption. But you also lose all the interesting return dynamics that are due to equity valuations occasionally deviating and then returning to economic fundamentals. That’s why I will always stick with historical returns despite the limitations!
Conclusions
1: In retirement, an equity glidepath with a positive (!!!) slope helps you during an equity bear market. But not just on the way down! A lot of the benefit from the glidepath comes from better rebalancing dynamics during the subsequent bull market!
2: A glidepath can alleviate some of the sequence of return risk. But the effect is still relatively small. Don’t even start to think that a glidepath can miraculously make the “4% Rule” feasible again over the next 60 years! Expect an increase in the sustainable withdrawal amounts by about 5%, or a slight to moderate decrease in the failure probability of any given SWR.
3: A successful glidepath in retirement should ratchet up the equity share pretty rapidly and reach the maximum equity weight roughly over the length of one complete bear plus bull market. Dragging out the glidepath over 30 years or more is not recommended!
4: Historical simulations show that an equity glidepath is useful when the CAPE is high at the commencement of retirement. As it is today! If the CAPE is below 20, glidepaths are of no use and an aggressive static equity allocation (close to 100%!!!) has performed best in historical simulations!
5: Monte Carlo simulations miss important elements of real-world data, i.e., mean reversion of equity valuations and changing asset return correlations. Hence, glidepaths that were calibrated to do well in Monte Carlo simulations (Kitces and Pfau) tend to do poorly in historical simulations. Unless we believe that the past observed dynamics of equity returns no longer apply in the future, we should disregard the Kitces/Pfau glidepaths because they’d likely perform worse than even most static asset allocations.
Hi Karsten,
I’m doing a second read through of your SWR series. I haven’t gotten up to the glide path articles yet but I have a question that popped up in my mind.
Is the bond allocation 10-year treasuries or does it not matter? I remember you saying how Kitches used different bond allocations to sort of fudge the results so I wanted to clarify.
Also I never hear anyone recommend TIPS or other inflation-linked bonds as the bond allocation. What’s the downside to using inflation-linked bonds as part of the bond allocation instead of treasuries?
Big ERN,
Thank you so much for this series! The rigorousness of your analysis is very helpful.
I am about 2 years away from ER and have been looking for some insurance against sequence of returns risk. I’ve decided to implement a 60->100 glidepath with 0.4% active shift and could use your advice on a few practical questions:
1) In modeling how to implement this, should we use a “buffer” to decide if equities are at their peak? In other words, if I’m looking at the value of the S&P 500 on a specific day each month, to decide if some funds should shift from bonds to equity, it would be unusual to pick a day where it is EXACTLY at its high. Should I have a rule like “if the S&P is within 1% of its 52-week high, do NOT shift” to deal with day-to-day fluctuations?
2) My equity portion is 70% domestic, 30% international. This month, if I look at the price for VTI as a proxy for S&P 500, it says not to shift anything. However, VXUS for international is 10%+ below its peak. Do I shift if EITHER is below? Both?
3) Is there a safety valve CAPE value where it makes sense to change the allocation more quickly? You mention no value to the glidepath at CAPE < 20… at that point, does it make sense to start moving large chunks each month? Everything?
Thank you again for all your work!
I like the buffer method in 1).
2) I would probably shift 0.3×0.4% =0.12% in that case
3) so you want to be a market timer? Great for you! I doubt that the CAPE will easily drop below 20 again, but yes, absolutely if we get to the low 20s and below 20 I would accelerate that glide path!
Best of luck!!!
I came here to ask the same question #3 as Emily. It seems as though an analysis could be run with accelerated or deceleration glidepath % based on CAPE for best results.
Or perhaps more simply, it would be interesting to see results for every time the CAPE dropped to 20(?) or every time market dropped by 30%(?) and your entire bond allocation was shifted to 100% equities at that moment, how you’d fare relative to the static 0.3% or 0.4% glidepaths.
Have to leave that you though. I’m just a caveman, your world frightens and confuses me.
I’m very curious about how your bond portion is invested. I’m also concerned about all of the capital gains I would incur shifting my current allocation (mostly equities) into a 40% bond allocation. The equity-to-bond conversion would need to happen in taxable accounts, since I won’t have penalty-free access to retirement accounts for many years.
I have no bonds. Well, I have some Muni bonds to hold the margin cash for my option trading. But without that I’d have no bonds.
I assume, that for us non-US based investors, with a glidepath we would also want to reduce currency risk on the glidepath? Obviously a big exchange rate drop on the international investment side of the portfolio could be just as bad as a market crash (though often those are related).
I am an Australian and have about 50% in AUD denominated bonds/equity and 50% in non-AUD denominated bonds/equity. So I assume my glidepath should probably increase bonds pre-FIRE and then increase equity post-FIRE (as per the above), but also increase AUD denominated assets pre-FIRE and then slowly return to target weighting (eg. 50/50) post FIRE.
Does that sound about right?
Sounds about right. I would simply hold the bond portion in Aussie bonds. For equities you certainly want the diversification with non-Aussie stocks.
Great piece and series!
Another difference between the methodology of your study to the Pfau/Kitces one is that you calculate the SWR (or SAFEMAX as they call it) and they take the WR as an input/constant (either 4% or 5%) and calculate the success rate of that. I wonder if this can also cause the difference in results.
That shouldn’t make a difference. When I calculate a SWR <x% for one cohort it should happened if and only if the Pfau/Kitces calculations find failures of the x% Rule.
Maybe I’m talking nonsense, but it seems to me that they try to find the glidepath that minimizes failure rates at WR of 4% and you try to find the glidepath that maximize the SWR. Not sure if it should result in the same glidepath.
That may indeed make a (small) difference. But I showed that their optimal GP is still suboptimal for a 4% WR when using historical data. Monte Carlo vs. historical return patterns will have the biggest impact.
Sorry I missed that you showed this in the table for CAPE>20. Probably you had the same results for all CAPE’s. Keep up the good work!
Hi ERN, love your work. Thanks for the 30-year timeframes. Was wondering if you would consider breaking the CAPE categories into smaller ranges? E.g. 30… in my view there’s a big market valuation difference between a CAPE of 21 and a CAPE of 33. If I see a 3.8% withdrawal rate in the CAPE >=20 section, and maybe that applies when the CAPE is 22, I wonder if I should take it down 10% for a CAPE 30 environment? i.e. 3.8% x .9 = 3.42% when the CAPE is 30ish? Thanks!
I would advise against using too many CAPE ranges/categories. The number of independent observations will become so small that the SWR stats become unreliable.
Looks like some of my previous comment wording regarding CAPE ranges got cut off. I meant having CAPE ranges of under 20, 20-25, 25-30, 30-35. Unless there isn’t sufficient historical data for those ranges? Or to put it another way, how would we adjust the SWRs in a low 30s CAPE environment vs a low 20s CAPE environment? Recently it was in the low 30s.
Do you have any insight into Warren Buffett’s 90% stock 10% cash allocation? One study by a finance professor at IESE suggests this 90-10 allocation has a 2% failure rate using the 4% rule: https://www.ieseinsight.com/doc.aspx?id=1815&ar=7&idioma=2_blank.
Not a bad allocation. It increases the risk of some really bad outcomes, though. The 4% unconditional failure probability is lower than for a 60/40, but the failure probability is higher (>10%) when you condition on a CAPE>20. So, that’s a caveat!
When using the Glidepath and your SWR toolbox calculator together, should we use the SWR of the initial portfolio allocation? If so, do we change it when we have reached
For Example: at a 40/60 for 30 year horizon its showing SWR = 4.38% (CAPE+30), but when the glide is completed (assuming CAPE didn’t change) at an 80/20 its showing SWR = 3.54%. Should we recalculate each year based on current CAPE and new portfolio allocation?
If the bear market occurred during the time you went through with the glidepath then the CAPE will have very well changed. So, you definitely want to check how your current asset mix and your current WR performed when the historical CAPE was similar to your current CAPE.
Did you incorporate expense ratios into the total return computation in your model? If so, what values of expense ratios did you use? (Expense ratios would play a relatively big role in the final SWR, so curious to understand if/how they are incorporated here.)
In my Google Sheet (part 7 and 28) there is a field where you can enter your ER. I use 0.05% p.a. as a weighted expense ratio for the stock/bond portfolio in all of my calculations here.
I tried your SWR Toolbox v2.0 spreadsheet, and it seems to give different results than this article. For example, for 80% fixed stocks allocation over 60 years to 0% final target valuation, the failsafe SWR given by the Toolbox spreadsheet is 3.29% while this article lists 3.14% as the failsafe for the same set of params. Any idea why the discrepancy?
Set the additional cash flows in the “cash flow assist” tab to zero and you also get the same 3.14% failsafe in the Google Sheet
Would be intrigued to see research on optimal glide paths into the bind tent during accumulation as well!
Yeah, I’ve looked at some simulations. It will be a future blog post! 🙂
I’m very interested as well!
What’s your take on the following path-independent dynamic asset allocation strategy? https://www.bogleheads.org/forum/viewtopic.php?t=293469
I don’t interact/discuss with the Bogelheads. Too many trolls there, so I don’t try to respond to any posts there. The post starts with “The optimal asset allocation for a 4% withdraw rate with a 30 year horizon is 50% stocks.” which is not true. Not sure if I want to go through the rest after the post starts with an incorrect premise.
I have proposed path-dependent AA for the withdrawal path (=active glidepaths where the glidepath slope depends on the equity returns). They seemed to work pretty well.
Thanks for taking a look. Mainly, I was thinking that their high-level strategy of having a lookup table that maps (portfolio-size-divided-by-annual-spend, years-left-in-retirement) to (success-probability, optimal-allocation) seems like a very useful way to represent a (path-independent) strategy to guide rebalancing. I was also wondering if there’s some nuanced flaw in their constant shifting of allocation as the portfolio grows and shrinks over time. It “feels” similar to the fallacy about SWR that you pointed out in your recent post (i.e., no need to reset SWR when the market drops because of conditional expectations), but I can’t quite put my finger on it, so maybe I’m wrong about that.
Yeah same feeling here. Looks fishy, can’t put my finger on it.
Also, nobody can ever claim that they can provide a table where retirees can just look up their # in a table. Retirement is too personal. We all have other cash flows and spending constraints that will push the true # mich higher or lower.
Thanks so much for your SWR series, Big Ern!
I am thinking of beginning an early retirement with a 45 year horizon. 3.25% covers my current living expenses.
Planning on following a 10 year glidepath from 60/40 to 90/10 equities/bonds, to hedge against sequence of return risk.
How do you feel about using an ETF that tracks the Bloomberg Barclays US Treasury Inflation Protected Notes (TIPS) index, for the bond allocation, rather than the non-TIPS index that IEF tracks?
I am a beginner and these posts have been a real help to me! You should really write a book about SWR. I’d buy it 🙂
TIPS should work roughly as well as nominal Treausry bonds. With TIPS you have the added benefit that they are a hedge against a potential future inflation shock.
Working on the book, but it’s hard to write it the way I want it. And I got too many distractions. Too busy in FIRE! But stay tuned! 🙂
When I look at glidepath ideas I see really different numbers on how low the equity allocation should be. But based on your work I’m thinking that 60% is the right value to target. I’m 50 now and hope to retire sometime in the next 5-10 years. I currently have 84% in equities (14% is in a cash balance pension, so I can’t change that until retirement). Should I begin moving out of equities to get to 60%? How fast – is a straight line from here to 60% over the next 5 years sensible? Should all of my new contributions go to cash or should I contribute to the 60/40 allocation and start moving principal?
Good question. And good timing! Please see the new part 43, published on March 2! 🙂
I’m having trouble determining what conclusions to draw. Is it still best to draw down to about 60% equities over the next 5-10 years? I would prefer to retire in 5 years but can certainly hold out for 10. Or maybe even go down to 3/4 time 5 years from now as a hedge. Given that and the fact that my cash balance pension (currently 14% of retirement money) MUST be in cash (at a 4% return minimum) is it too early for me to start moving out of equities?
Depends on you risk-aversion/flexibility. Maybe draw down linearly now to your target. Or hold back another 2 or 2.5 years and then draw down to the target. Whatever makes you feel comfortable.
A linear path seems reasonable. Whether you can do the shift with new contributions alone or whether you have to shift principal as well depends on the asset returns over the next years.
Dear ERN,
My question is about implementing a rising glide path (say 55% to 75%, step 0.2%/mo). Currently, equities are down, reducing my current exposure to ~51%. Assuming this holds at month end, is it recommended to rebalance all the way back to 55.2%, or maintain the 0.2% monthly step (51.2%)? I am 3 years into retirement with a 35 year window remaining. with investment/(projected needs) ratio >40.
The simulation assume that you rebalance your portfolio every month to the target. And target might be fixed or a moving target in a GP.
Dear ERN,
Thanks for you earlier response. The current CAPE(2)~33.7, the highest since the dot-com bust, and SP500 at ATH. Both of these may go higher due to Fed reducing rates and AI Hype. Have you ever considered a (Hyper-) active glidepath that would incrementally reduce equity exposure during risky periods, then resume the normal path thereafter? In contrast to the active GP presented here, that just holds steady and waits for the market correction? Any potential advantage for SWR?
That’s not a glidepath. That’s GP plus tactical asset allocation.
There is some merit to scaling back equity exposure when volatility is high: See Moreira and Muir’s 2017 Journal of Finance paper: https://www.jstor.org/stable/26652549
I haven’t simulated that yet, but I’m sure it works in the investor’s favor. Not a panacea, though.
Hi Big ERN,
I love the reverse glidepath–the concept, the results, the impact on SORR–everything about it, except for the mechanics of buying back in each month, calculating the real S&P to determine if it’s an ATH, and so forth. (Perhaps this is in part because I’m not fully seeing the mechanics play out.)
But, what I’m not seeing is how there’s not a simpler approach that intuitively seems really close to the same outcome: 100% equities until you hit your magic number, then sell 8 years of consumption worth of equities and buy into an 8 year bond ladder (mechanically, leave 1 year in a MM or HYS and buy a 7-yr ladder) and then spend down these bonds for the first 8 years of retirement (one could replace 8 with 9 or 10), leaving you at 100% equities at the end of 8 years. You can tweak this by adding 1% or 2% per year for spending increases.
Today, you can put this into TIPS and the inflation adjustment is take care of and it is less than your SWR x 8 x portfolio value because real rates are positive. Point being: you’re likely selling at a high and riding out the next bear market since high-value markets help people hit their magic number. You’re not “over-selling” (timing the market) and buying back in. You are doing this only at a time when the numbers work (i.e., you’ve “won”.)
Example: SWR is determined to be 3%, portfolio is $3M, so $90k/yr. At a TIPS building-ladder site (not sure if I can/should link) I see that today I can build a 7-yr TIPS ladder starting in 2026 and ending in 2032 costs $601,769 (a real return of 1.45%.) Add my $90k for the first year, and I sell $691,769 worth of equities [tax implications caveat], put $90k into a HYS and buy the $601,769 TIPS ladder. (This ladder throws off $2500 in interest in 2025, so I’d have either $2500 extra or only need to sell $689,269 for of equities.) My $2,308,231 worth of equities remain untouched for 8 year, addressing the SORR on a portfolio spend-down. After 8 years, I’m back at 100% equities without managing anything; at that point, I live off equities. For an extra $20k or so today I can add 1%/yr real spending increase. Like I said, we can do this for 10 years just as easily.
It is worth noting that today we are in a different rate world than the ZIRP during which you originally did this analysis. If I were to buy a 30-yr ladder, $90k/yr real costs $1,999,104, has a real rate of return of 2.33% and a withdrawal rate of 4.71%. 4.71% is massively greater than 3% (has a 100% success rate over 30 years, and a 0% success rate for a day longer!), and my remaining $1,000,896 of equities grows untouched for 30 years. Because equity returns are greater than bonds in the long-run, I’m not arguing this is an optimal path, but it is illustrative of how powerful real bond returns are. In 30 years at a real rate of return of 3.75%, the $1M in equities have grown back to a real $3M.
Basically my question is how is this 8-yr TIPS ladder inferior to a 60% to 80% reverse glidepath? What is this approach missing–is it just the investment (buying back in) SOR, and how significant is this in terms of $?
Sorry so long!
I have simulated “Safety First” scenarios like that in part 61 and didn’t find anything better than the Glidepath setup.
Thanks for sending me to Part 61–lots of good comparisons playing out there.
I have gone through my own analysis of the 30-year TIPS, but not with a deferred annuity because I couldn’t find any quotes for inflation-adjusted deferred annuities, and over 30 years if equities don’t do better than annuities, your annuity is going to face serious counterparty risk, so you should cap it at your state’s insurance limit!
The results of the Safety First plus 100% equity portfolio is impressive during CAPE > 20 period and high real TIPS yields (the current environment), and it seems to be a winner. Also for conservative folks, it is not an expensive tradeoff for reducing Sequence Risk in worst-case scenarios (giving up 4% on the median in exchange for 2-3x that on the tail.)
Safety First vs. Glidepath is more along what I had in mind, but it shifts the weights linearly over 25 years, and your readers will know that that’s too long. It’s waiting too long to get more portfolio back into equity returns. Your comment (and this entire site) lead me to believe you’ve done a lot more simulations than what’s on that post. Have you done the Safety First vs Glidepath with 8-10 year period to get back to 100% equities?
I think we’re on the same page. TIPS yields are enough for a ladder. Annuities are usually nominal, so it’s hard to find anything with true CPI-adjustments.
My simulations in parts 19+20 use a GP with a roughly 10-year transition period. They will perform a bit better than the 25-year transition period. But in part 61, I used 25 years to be more comparable to the TIPS ladder setup. If the 25-y GP beats the TIPS ladder, the 10-year 60 to 100% GP will do even better.
Hi Dr. Jeske, thanks for the continued information on glidepaths. I understand your aversion to using Monte Carlo methods, at least in the Kitces-Pfau incarnation of it. Sampling randomly from the historical distribution to simulate future annual (not even monthly…) returns definitely feels suspect.
I do agree in general that memory-less machines like that cannot adequately simulate the future. However, what if the Monte Carlo was not drawing randomly but instead following a stateful process; for example, what if we had a 9-state or 11-state Markov chain? Or instead used a recurrent neural network (GRUs, LSTMs are quite popular and cheap to run with today’s hardware, at least at our scale of data). I ask because I work in artificial intelligence and I think this is the prime application of those probabilistic techniques: namely, we can generate panel data (say 1m parallel lifetimes) and analyze them. The key is to ensure that the sequence-to-sequence model is seeded correctly with historical returns so that the _average_ performance across those many parallel universes still roughly resembles the historical distributions.
What are your thoughts? I think this could be a lot more interesting than blind distributional sampling, albeit with substantial added complexity. It could be pretty fun to additionally compare the results with the back-testing from historical data. In fact, another exercise could be hold off on using some of the most recent historical data and see how well the average simulation process can generate the known recent returns and assign a reasonable perplexity score. (I figured that this kind of holdout set is necessary anyway when going the RNN route.)
Thanks again for your time — I really appreciate it as a fellow money nerd. 🙂
I can warm up to such block-MC simulations. I’ve done those as well. But how long do you want these windows to be? I noted in an earlier post that the market often goes through 30-year cycles, i.e., you find a strong negative correlation between negihboring 15-year windows:
https://earlyretirementnow.com/2018/01/24/random-walk/
So, if you need to include 30-year windows to capture such important correlations, why not just simulate your 30-year retirement with historical data?
Hi Dr. Jeske, thanks for your reply. Is the idea here that every sliding 30-year window thus constitutes a reasonable economic cycle? I guess I’m looking at about a 45-year retirement for FIRE, so it would be about one and a half such cycles. I guess if it’s all theoretical then back-testing works just as well as forward simulation, since we’re not trying to predict the future but instead develop a strategy that holds up even under Black Swan events (hence the 3.25% SWR etc.)
Tbh I’m mostly looking to answer the question of when to push the retirement button for myself. My budget now and my budget in retirement are relatively fixed already, so it’s more just a question of “when should the accumulation phase end?”, paired with “what is the associated safe withdrawal rate?”. It seems to me that regardless of one’s numerical target, the ceiling SWR is basically around 3.2-3.3% based on your own back-testing analysis. So only the question of the numerical amount in accumulation remains, and I was hoping to use a bit of MCMC magic to simulate future inflation and future returns for a bit instead of assuming fixed returns for forward projections.
That said, one concern I have for historical back-testing is the lack of data. The issue is that if we use historical data, a 20-year accumulation phase followed by 45-year retirement is 65 years’ worth of market and inflation data. There aren’t that many cohorts at the monthly level when the window is this long, and I’m mostly worried that there isn’t enough data to analyze the outlier events, because stuff like 2000/2008/2022 happened relatively recently. I wanted to see the impact of those events during accumulation/early retirement in the future. But the only way to do that is to simulate it or assume fixed rates, right? How did you handle this issue when back-testing? More concretely, how did you make projections for, e.g., the retirement cohort of 2007?
This was the thing I was always stuck on when following along and doing my own analysis!
Lack of data: There are enough 45-year cohorts to test all the different macro disasters, e.g., demand shocks and supply shocks. I don’t see that as a problem. Even if that were a problem, creating artificial data out of nowhere or out of actual returns doens’t solve the original problem either.
I went through a numerical example in Part 58 to see how supplemental flows and other idiosyncratic parameters can raise the WR substantially. I recommend you check that out before you use only a bare-bones 3.25%.
Hi Dr. Jeske,
Thanks, that makes sense. I’ll be sure to check out your writings in Part 58. Am I right in understanding that if we limit ourselves to historical data and only consider 30-year cohorts, your last retirement cohort would be in ~1995 for the 1871-2025 data? Of course, that’s still plenty of cohorts. There would be one for each month between Jan 1871 and ~April 1995, which is around 1492 cohorts (already a lot).
(and similarly, for 45-year cohorts, the first cohort would be in Jan 1871 and the last one in ~April 1980)
As for data synthesis, I think one thing that I found to be quite interesting when working in the ML space is that constructing realistic examples from a generative Bayesian model can be pretty effective when analyzing these kinds of “hidden mechanism” processes. The question is whether there’s sufficient data in historical records to accurately develop such a model. I suspect “no”, and not with the appropriate level of longitudinal granularity either. So I can understand the restraint. Unfortunate — I think we’ll have to go without the fancy forward-looking techniques for this problem space then!
As I described in my entire series: you can simulate retirements up to 60 years/720 months. You can simulate them a little further than 2025-60=1965 because the most crucial time period is the first 15 or so years due to Sequence Risk. Simply fill in the last few decades with calibrated data. See Part 1 of the series for an explanation.
You can always draw more observations from some process. But the process has to be trained through actual data and for that we have only ~150 years. No additional insights are created by simulating a billion years of returns. And likely you are going to lose some features. So, I wish you best of luck generating your own personal return data, but I would not trust the results.
Hi Dr. Jeske,
Thanks again for your reply. I appreciate your patience. Yes, I’ve read Part 1. I saw you had this note written near the top:
> We extrapolate past the current history and append equity and bond returns after September 2016. To this end, we assume long-term average returns for equities going forward (about 6.6% real p.a.). […] We should note that these return assumptions are likely going to generate higher sustainable withdrawal rates due to the absence of return volatility.
This seems to suggest that a flat rate of return is assumed for the calibrated data. I’m a bit confused about how this gives us a strong degree of certainty: doesn’t this inherently not tell us anything about sequence risk? I believe you left a comment on that post where you say yourself that the average return is, in a sense, “not good enough”, and the full distribution is truly required (as well as the order in which the events occur).
I do agree that simulation is only as strong as your sampling distribution technique, so there’s not really any additional insight gained at all for doing 15 extra years or 15 thousand extra years. So are you saying that we have full 60-year retirement data for 1871-1965, and then the rest of the cohorts have at least some of their returns extrapolated using flat long-run real return rates?
As you said, since the majority of the risk is in the first 15 years, I guess technically speaking we’re good through around 2010, and after that assuming long-run is good enough. Maybe that’s sufficient then?
Based on my takeaways while I’m working through the SWR Series, if you’re concluding 3.2-3.3% SWR then you (like me) are conservative. Be sure to check out Part 61 Safety First. To me, Part 61 shows how us conservative folks can bump those 3.25% rates up a bit, particularly in this CAPE>20 / TIPS>2.2% real return world. Then you can bump it again in Part 58.
Have you ever looked at glidepaths for different periods (I am planning for a 40-year time horizon)?
Yeah. It still works over longer horizons as well.
Hi ERN,
Thank you for all your fantastic work in this space, I’m enjoying reading all your posts and watching related podcasts. One question I can’t find answered anywhere relates to the a and b parameters for the CAPE based rule. I understand that a = 1.75% and b = 0.5 is generally optimal, however how do I know if this is appliable to my situation and when should the a and b parameters change?
I display charts of the simulated retirement withdrawals and tables with stats on withdrawal volatility, drawdown, etc. Check if you find those palatable. Play with the parameters if you don’t like the results.
Does a high CAPE and/or high Debt-to-GDP suggest favoring a lower starting point on a glidepath? The 40-to-100 performs almost as well as the 60-to-100 over a long history of mostly lower CAPEs and Debt-to-GDP, so would you guess that 40 ends up being better for today’s cohort? Thanks.
I have never performed this conditional analysis with the D/GDP ratio. I don’t think it would show anything. It might even give you strange results, i.e., post-WW2 we had a high D/GDP ratio but also a nice equity rally.
But a GP is more useful with a high CAPE.
I’m trying to understand why the first “first 10 years” matters the most for sequence of return risk and for your glidepath. My reasoning is that if I successfully pass through my first 10 years and thus have passed the most dangerous timeframe with respect to sequence of returns risk, then I’m left in a scenario very much like when I started. I have a certain amount of investments saved up and a certain amount (or desired amount) of annual spend. And when looking at the chances of success going into the future sequence of returns is a key factor, and it’s the next 10 years that matter most. Nothing should be special about the first time I do this; the logic ought to be path independent. So what’s special about the first 10 years when I begin retirement compared to the next 10 years when I do the analysis again 10 years from now?
All I can come up with is that first, my life expectancy wont have changed all that much in 10 years unless something catastrophic happens, so I’ll have 10 years less that I need to be able to support. And the first ten years will see with high probability either a) my portfolio grows such that my existing fixed real withdrawal amount is now a lower withdrawal rate, or b) the CAPE ratio will be lower (because the equities didn’t go up). So this leads me to wonder whether your starting position on the glidepath should be conditioned on your withdrawal rate and the CAPE ratio.
My context here is that I was recently laid off, and I’m considering turning this into early retirement. My portfolio is currently around 90% equities and 10% bonds and has regularly been rebalanced to maintain this. I have a bunch of cash coming in soon from a pending house sale. If I look at my annual spend then excluding income taxes I’m currently withdrawing and spending about 2%. I’m pretty comfortable with that amount of spending. I have a financial advisor who is suggesting buying treasuries or CD in a ladder expiring in 1, 2, 3, 4, and 5 years as a way of reducing sequence of return risk. I think that’s a worse version of the glidepath concept outlined in this series. But I have the practical question of where I should be on the glidepath. I should add that I would prefer both for piece of mind and for my children that I prefer to think of the failure condition as ending below current portfolio size rather than hitting 0.
Correct, after you’ve lived through the first 10 years, you now face another Sequence Risk problem, where the subsequent 10 years matter the most. I’ve written about the “self-similarity” isse in part 38. See here:
https://earlyretirementnow.com/2020/07/15/when-can-we-stop-worrying-about-sequence-risk-swr-series-part-38/
But the narrative is still true: if you’ve survived the first ten years unscathed, you are OK. You now face the Sequence Risk of “surviving” vs. “mega-rich” where the worst case scenario is much better than when you first retire.
Part 38 does indeed address the topic I was thinking about. There conclusion there is that sequence risk remains a factor as long as you are withdrawing. Now I’m thinking of what to do with that information. I think a takeaway is that where should be on a glidepath shouldn’t actually have anything to do with time since you retired. It should be a function on retirement years remaining, withdrawal rate as a percentage of current portfolio size, and current CAPE. That means that the glidepath would be affected not just by the passage of time, but what the market (or non-market effects like a windfall or a sudden unanticipated cost) has done to you over time.
Assuming that actually makes sense, you could take this a step further. If you have some appetite for risk, say you a withdrawal rate and allocation that is 99% likely to succeed, you can periodically recalculate not just your asset allocation (place on the glidepath), but also your withdrawal amount. With time-since-retirement not entering into the math, it would be just like you are retiring again today with your current portfolio and market conditions. There’s a real chance you could end up being able to increase your annual budget. Of course your risk would go up when compared with the risk you’d experience if you kept your lower budget, but it would not exceed the risk you chose to take when you retired initially.
A final thought, and this is a bit more speculative, but I think this naturally feeds into strategies that do note purely optimize on success probability as defined by not running out of funds or maximizing a fixed SWR subject to success probability constraints. If the above allows you to increase budget later in retirement, there’s value in that. People generally experience a decreasing marginal utility on an increased budget, so the goal would be maximizing the area under the curve on some sub linear utility function. You could calculate that at various failure probabilities and then see how they trade off.
Yes, SoRR remains an issue. But it’s likely worse at the beginning of your retirement when the worst-case scenario would draw your portfolio down and/or your retirement budget falls below comfort level.
Later in retirement your often SoRR decides between living lavishly vs. living luxuriously. It’s still risk, but much less of a headache.
Hi Karsten – I’ve read Part 19 and Part 20 and wondering if there is a way to arrive at the SWR for a 45-year horizon for rising equity glidepath (40%>80%)? Part 20 includes tables for 30-year horizon and 60-year horizon assuming capital depletion. Is it simply taking the average of the SWRs for these two horizons to arrive at a 45-year horizon SWR? In addition, for the 50% Final target and 100% Final target (CAPE >20 and 60-year horizon), is there a way to back into a 45-year horizon using the 60-year horizon table (no 30-year horizon table provided for comparative purposes)? Thank-you.
You can simulate your own glidepaths with any length retirement you specify: https://saferetirementspending.com/
Thank-you. Much appreciated.
Hi Big ERN. Thank you for the great series.
I’m trying to piece together what would be the strategy for maximizing average final asset value for cases where someone accumulated too many “one more year”s. For example:
* 50-60 years of life expectancy ahead
* Assets are x40 of expenses (aka a WR of 2.5%)
* Expenses are flexible, with flexibility to comfortably cut 40% of these (aka get to 1.5% WR)
* Retiring now
In such a case, if the goal is to maximize the average final asset value, would you recommend:
a) Keep 100% equities, and use a CAPE-based withdrawal rate
b) Build a 60/40->100/0 active glide path, and use a static SWR
My guess is that 100% equities have a higher average final asset value, but not sure this applies in current high-valuation market conditions.
Thank you!
Impossible to determine without knowing what your static WR and the CAPE parameters are. But my suspicion is that in a direct comparison, 100% equities will give you higher chances of a high final value but also high chances of massive, painful retirement budget cuts along the way.