Pages

Friday, August 15, 2025

ai-age, retirement

The question of our time is what is the “magic number” needed to finally not have to work anymore.  How do you know if you have made the right decision about your own number or number of years needed before you retire?  As working people get closer to age 50 it’s a question that becomes more acute, and drives people to anxiously work against a goal that was arithmetically consecrated by industry, or word of mouth.  Often using one-size-fits-none, Procrustean rubrics.  Below we unpack the formulaic nature of these math ideas, and examine the pros and cons of different approach estimates.  Let’s begin with a provocative poll seen by many thousands: if you need $2.5 million to retire at 65, how much do you need to early retire at 50?  Never-mind the simplistic $2.5 million precision because the overall framework can be scaled up down for your individual needs [eg, geography]. 

The poll however is more thoughtful and revealing than you might expect, leaving [as shown] most people including AI puzzled, or overly optimistic in their approach.  This is quite dangerous, as the ability to make up shortfalls later in retirement, including in the habits one needs to enjoy for a healthy retirement, become more challenging later.  But there are almost life-saving interlocking concepts that needed to be considered early in your retirement in order to be on your best path, and this complexity often is bypassed in anyone’s thinking or models [and they are probabilistic in nature]. 

Indeed it turns out you need more money to retire than imagine [and need to work 2-3 years more as a result], but neither should be considered by much if you plan correct.  So this article is not just theory; it's a vital reality check, offering practical takeaways that will empower you to reassess your strategies and maximize hard-earned savings.

 

poll question

First, let’s see the results from this poll, which incidentally mirror the many anecdotal conversations I have had from smart people in my own circle:


Worth stating there is a provided comment below the poll clarifies that withdrawal at age 50 is 3.5% at that time, and the nominal investment growth is perpetually centered about 4.5%.  Yet anyway look at the range of answers!  Nearly 45%, including AI actually don’t know at all.  The second most popular answer was the lowest stated value of $2.2 million, or -12% less than the age 65 goal.  What was yours?

It’s a little bit of a cop out to use or justify the wrong answer based on really extreme assumptions.  We can further ask secondary questions which is when and at what level does the portfolio peak, regardless of what is your chosen answer?  The answer will likely surprise you, and will be delved into a little further below.

 

simplified industry standards

The are popularized ideas on how to answer this.  Two in particular we cover below.  The first popularized method is simply use say apply a 3.5% withdrawal at age 50 to match needs at that time, and 3.5% is generally a little more conservative for retirees and more in the flavor of the popularized FIRE movement.  Recall for decades people have whispered the “4% rule” at age 65.  Recall again these are coded percents however that completely misalign with the long-run and serious acceleration instead of retirement expense drivers. 

We use Pt to signify principal level at a given age or elapsed time t, while using P to refer to principal level in a general sense.  The 3.5% of objective P65 [$2.5 million] expectation, is a withdrawal at that age of $87.5k.  We can then discount this withdrawal back to an age 50 needed value of $87.5k/1.03^15, or $56.1k.  Hence $56.1k/3.5% is a possible P50 of $1.6 million.  Note that this is far below any of the polling choices provided, so what happened? 

One large drawback of this method [and its one of the initial two models that leading AI is aware of and guides users with] is that it ignores your requirement of wanting to have a P65 of $2.5 million altogether, just to ensure that there is adequate expense coverage during the bulk of your retirement years.  The math just completely misses!  Instead your portfolio grows to only P50*[1+r-w(t)]^15 if we assume that W [withdrawal level] grows with inflation [growing at a slower rate than the principal’s gross rate or r].  w[t] is the withdrawal rate at future time t.  Or if we impossibly lock-in a w of 3.5% instead, then the portfolio grows to a far different value versus $2.5 million, completely missing the mark.  More importantly the W[t] and w assumptions in the end are flawed too, particularly over the long run stages of retirement as we’ll suggest further below. 

It’s incorrect for P50 to be as low as $1.6 million, and see the withdrawals not even growing at the basic cost inflation rate in the initial years, let alone correctly completely eclipsing investment returns. At age 50 for example the 3.5% withdrawal is again $56k.  Leaving a $1.6 million portfolio at $1.61 million for age 51 [the $1.6 million grows slightly on a gross basis but then you are also withdrawing the $56.1k].  So now age 51 the 3.5% withdrawal amounts to less than $56.5k [so not even 3% expense inflation but a withdrawal inflation afforded of just under 1%].  So the “model” despite the low $1.6 million goal is severely broken and unusable even by the second year of retirement.   

The second major drawback with this simplistic modeling approach, and critically important tragedy of this framework is that you really need high a $70k’s withdrawal, and not $56.1k, so leaving beside w[t], the underlying flows W completely mis-calibrated for your level of needs on day one of retirement.  In other words you can't assume that just because w50 is 3.5% that w65 is as well, and the result is the entire discounting exercise as a result is mis-calibrated and highly sensitive to the values that you incorrectly assumed from a simplistic model. 

Now the second popularized modeling type to estimate your magic number is to start with a future horizon value and working to the present iteratively.  So right off the top we know our objective of P65, and t of 15 [15 years of early retirement].  We also assume we must start withdrawing as a result at age 50, if we are going to assume to be retiring then.  This can be expressed as a rate or amount at that age.  So start oscillating and experimenting to estimate a r of 4.5%, yet a w of 3.5%.  This gives us an initial effective growth factor [eg] we can also estimate as [1+4.5%]*[1-3.5%], or 1.0084 [or you can average the two versions shown so far, and simply use 1.009].

And consequently P50 =P65/eg^15, or $2.2 million.  Note this will indeed also likely zoom to towards the arbitrary P65 of $2.5 million, and keep growing towards infinity after that.  It also incorrectly maintains the withdrawals as a component of returns.  But doesn’t imagine a large and independently modeled withdrawal level for consistency.

The real challenge again here is that there is no way to mathematically create the proper second order rate on the expense inflation, and this is more and more needed over time as one begins to stop being able to work at all.  Also factor in illiquidity of domiciles as you reduce your portfolio to prefer simply liquefying it, and the health care costs that are very uncertain and convex.  But if you don’t stress over your health early in retirement and start too complacently you will completely be unprepared for life’s ultimate shortfall. 

Despite the popularity of the polling result, in the end we’ll need more than $2.2 million at age 50 for an early retirement.

 

ask AI

This question was posed in general ways to AI, with it almost never getting the correct answer despite peppering it with heuristic questions to clue it in.  This leaves you defenseless, only you probe it repeatedly and actually know the correct solution in advance [defeating the purpose of asking it for a solution to begin with!]  One generic example is show below [when I was stress-testing it for early retirement of 55 with an extremely liberal gross return of 7%, and a baseline P65 of $2X million], but the examples are littered throughout nearly every step among many.  They also presents different mathematical results based on the most subtle of word choice differences in the prompt, including the ordering of inputs described to it.  See one modern AI output below.


Notice in the answer above the early retirement age concept wasn’t even accepted by AI [even though most polling respondents above knew to do that] and the discounting as a result was erroneous.  And there were other classification issues with liquid versus non-liquid assets that are critical and needed to be streamlined for the results to make sense.       

 

the actuarial insight

The actuarial reserving concept is fresher and groundbreaking in dismantling misconceptions.  However as we’ll see we’ll still need to take it further.  We can at least start with the visible mathematical variables in the formulas for the first order impacts of expenses.  Like a ship captain taking into account wind or current, but not both.  Start by estimating both the rise of the plan assets P, accounting for continuous and growing withdrawals.  And up through the time of retirement.  Again we focus on a common goal, here set for a single person to achieve a more-than-sufficient, $2.5 million nest egg through age 65, the typical retirement age framework.

One way to start this is to decompose the plan gross growth and the future value of this actuarial withdrawal function.  This is because there is a progression in the level of your expenses in retirement and it's not merely an embedded function, but constrained when both the principal and the withdrawl expenses are maintained independently.  For example when there is a bear-market, do your mortgage bills or onion prices suddenly drop -20% in lockstep?  Of course not.  Here's the complexity of this math is shown below:

 

P65    = P50*1.045^t – sum[over k=0 to t-1] of W50*eeg^k

             = $2.5 million

 

Where eeg is the “solved effective growth” for all 1+w[t].  And sum[over k=0 to 14] of W50*eeg^k is therefore still close to the age 65 future value of annuity, or W50*[(1+3%)^15-1]/3% for shorter horizons such as 15 years.

But the eeg should really should closer to 1.035 for the initial stage to retirement [ie, 3%-4% average expense inflation].  And then about 1.05 if covering the full horizon [ie, 5% average expense inflation!]

Another way to see this is that P50 needs to be discounted slightly towards -2% for every ½ decade early retirement [just towards -15bps per year].  By the way, there is no condensed and meek formula for P50 as a function of either P65 or t, but you must simply rearrange the formula above, which is easy to isolate P50.

So AI uses the short-cuts described the earlier sections with no second order growth in the rate of withdrawal.  It completely misses the actuarial insight assumptions, and randomly too based on how you prompt it.  And last, delivers back from this reserving method of solving P50 at too low and impractical of a magic number, at well below $2.4 million. 

 

psychology and portfolio peak

Once your solution is calibrated [as it should for the longer horizon] to include the higher and accelerating withdrawal expenses, you get an answer far closer to just over $2.4 million.  In these scenarios the withdrawals are appropriate, and we indeed overshoot to a mid-50s peak of just over $2.5 million [a benign, just under a +5% jump in fact!]  While mathematically correct you should not take this initial ramp-up as a sign to be complacent or assume your health care and other in-fact convex withdrawal expenses are fine.  They may dangerously not be.  See output from another AI below. 


issues and lessons

The lesson from this article is that certainly we can have a $2.4 million principal number that grows to just over $2.5 million through age 60, but then start a slow crank ahead but before-you-know it rollercoaster plunges toward earth, say from unchecked medical exigencies in your much later future.  While it appears that these polling solutions are only +10% or so apart from one another, the market can be unforgiving as well during this time and so could your workplace sanity if you start to internalize the need to work longer right as you arrive at a miscalculated magic number. 

Semi-retirement is a different concept as well, and that involves simply treading to cover withdrawals only from an initial to a final retirement.  Allowing more of the weight of the 4.5% return to work for you.  And the concept of reserving applies to be able to match up to how you are doing against that latter target frankly, versus an on-track measure early in your career to simply work towards. 

The investments you should be making over the time should be as diversified as possible and still provide liquidity whenever needed.  Luckily this liquidity is not needed as much in the earlier years.  Be cautious with the strain of high mortgages still in your retirement, assuming you still have a large chunk of your mortgage term left.

Retirement should be a time for celebration, but also strict prudency.  This is the final nest egg you respectably worked your life building, and the thought of returning to work for lower wages and a loss of productive connection to the workforce can be daunting.  AI tools provide quick and seemingly confident answers, but with large errors, no different than the simplest of industry rule-of-thumbs.  This is a key issue for some time, and discussed a year ago in my book copula narratives.  Despite advancements in AI, the concept of higher order risk and the magnitude it has over simpler risks -such as return sequence risk or basic assumption risk- can not be overstated. 

We can note that higher initial withdrawal or gross returns instead closer to expense inflation have similar damaging effects on your retirement number.  This is where one could envision the higher polling options: the $2.5 million or even $2.6 million variety.  Else significant superannuation or even mild longevity risk, with complete depletion, frighteningly a decade at least before death. 

So last, make sure to audit results in this case from AI models, and test from various perspectives, such as marginal stock and flow, and including at different time frames of your overall life [eg, age 50, 65, 80, 95, etc].


addendum

On the one-year anniversary of copula narratives, and the eight-year anniversary of Google banning me, it is interesting to now see how their top AI describes this book:

A well-regarded collection of statistical essays and articles by Salil Mehta, a prominent statistician, blogger and author. 

The analytical collection is a conceptual grouping of his insightful, profound analyses that break down complex statistical concepts through the lens of real-world, and often newsworthy, situations. 

copula narratives cross-roads is his metaphor exploring the often-unseen connective pattern between different events and variables, in the urgent world around us.



salil statistics [10k+ books sold, 36m reads, 1/4m follows]
nova consilium: analytica iniustitiae
follow via RSS or e-mail or Amazon

No comments:

Post a Comment