Friday, May 10, 2019

Going it Alone with Retirement Planning Software

I love models. I build and research models almost every day and have for almost two decades. For the past year, I have explored online retirement models while co-authoring a paper with an econometrician. Modeling is pretty much all econometricians do.

I don't like bad models, or models used for the wrong reason, or models used by people who don't understand them, especially when people who don't understand them try to explain them to other people who don't understand them.

It also concerns me that many people attribute mystical powers to computers. A computer model is no better than the human that programs it. It's just a whole lot faster and doesn't get complacent or bored.

So, I certainly don't like all models.

If you're going to use computer models to help plan your retirement, there are many things to consider.

First, economic models can be very useful to study your retirement prospects and figure out the best bets but they in no way predict your future. As the saying goes, all models are wrong but some are useful.

A Monte Carlo model, for example, can test thousands of possible future scenarios for your household but your retirement is a one-time event. There is no way to know which one of a multitude of simulated scenarios might be similar to the future you will experience or if any of them will. The tendency is to guess that your retirement will be like the median model outcome but that means you will be overly optimistic half the time.

The output that models create is only as realistic as the assumptions we feed them. Unfortunately, we can't estimate with any precision what future market returns will be, how long we will live or even how much we will need to spend over the coming decades. These are some of the key assumptions that drive models and we are, for the most part, guessing at what they might be.

In computer science there is an old saying, garbage in, garbage out. What we mean is that the output of a program is only as good as the input. Make a wild guess at the input and the output will be a wild guess. Unfortunately, many of our guesses, or "assumptions", are by necessity fairly wild.

A reader recently commented that "retirement planning is an unsolvable problem with unlimited variables." You could say that about chess, too, but some players clearly solve it better than others.

Retirement finance is unsolvable if your definition of "solvable" is finding a single, optimal solution in advance for your individual household. But, there are lots of "games" in economics that are probabilistic — retirement planning can be considered a "stochastic game against nature" in game theory parlance — for which we can determine the best strategies even though we can't be guaranteed to win.

We should never expect a model to provide a single, optimal solution to the retirement planning "game", nor should we expect that from a human advisor. The optimal solution can only be identified with certainty after retirement is over and that isn't very helpful for planning purposes.

Our goal, like that of the chess player, should be to find and implement strategies that produce the outcomes we want and avoid the ones we fear more often than alternative strategies do.

Many so-called retirement models concentrate almost solely on investment results. Those are investment models, not retirement models. A comprehensive retirement plan will consider many factors including Social Security maximization, annuitization, life insurance, estate planning, taxes, and others.

It is possible to use multiple models (perhaps an investment model, a Social Security model, and a tax program) to solve these problems individually but a comprehensive plan needs to also consider the interplay of these factors. (See a sample list of free limited-purpose models below.) Change any of these factors and other factors will be impacted. We would pay a big price in the planning process if we didn't consider that. A comprehensive retirement planning model is a much better tool.

A major benefit of Monte Carlo models is that we can test changes to many factors and see how they interact in one model run.

Assume, for example, that we are planning retirement with a spreadsheet model, such as the Bogleheads spreadsheet below, that considers many retirement-funding options. Let's say that I want to run the spreadsheet to consider all possible combinations of market returns ranging from 4% to 10% in 1% increments, asset allocations from 0% to 100% equities in 10% increments, and annuitization from 0% to 50% in 10% increments. Considering just those three factors, I would need 462 runs to capture the combined effects with a spreadsheet model or I could capture them and many more with one run of a Monte Carlo model. There are actually several other factors I should include and note that the spreadsheet is not modeling sequence risk.

It is unlikely that a retirement toolkit provided by an investment firm will give annuities, life insurance or reverse mortgages consideration equal to equities and vice versa. Better to find a model with no agenda.

It's also important to know who built the model and their qualifications. I have a lot of confidence in the Bogleheads and more in Laurence Kotlikoff, who created MaxiFi. Wealthfront identifies the developer of their retirement model. But, unless you know the qualifications of the model builder, I'd steer clear. Anyone can build a model and post it on the Internet.

Building a retirement model requires an understanding of finance, modeling skills, expertise in the computer language used (even if it is only Excel), and a sound understanding of statistics and probabilities. If you don't have all four, then building your own model is a very bad idea.


Retirement planning software can answer a lot of questions but you have to know what to ask.
[Tweet this]


Models can answer a lot of questions but you have to know what to ask. A model is unlikely to suggest that Roth conversions might be profitable, for example, or that you should consider a combination of annuities, whole life insurance, and equities, as Wade Pfau and Michael Finke have suggested.[1] A good human retirement planner knows what to ask.

I am wary of models that use probability of ruin as their metric of success.[2] Probability of ruin measures the probability of portfolio failures but does not measure the magnitude of losses. For example, it will count a strategy that funds 29 out of 30 years an unequivocal failure. It will count a retirement strategy that successfully funds 30 years as a success but no more successful than one that funds 50 years.

As Zvi Bodie points out, probability of ruin doesn't consider utility. Presumably, retirees will be less satisfied coming up $100 short of paying the bills than they will be satisfied with a $100 surplus. Paying the bills is a necessity; having a little extra is a nicety.

Probability of ruin is a particularly bad metric for Monte Carlo models because, among other reasons, results can change significantly by changing nothing but the random number draw. If you use such a model, try running it several times with the same input and see if you get nearly identical results each time. If the results change a lot for each run and never converge then the model is problematic. If the results are precisely the same for each run with the same input it may be because the model always uses the same set of random numbers for every run to speed up computation, so we still can't say for sure that the model is properly constructed.

If you are going to plan with a "retirement toolkit", I recommend the following:

  • Be aware that today's retirement models are not a replacement for a human advisor. If you plan your own retirement with these models then you and not the model will be replacing the advisor.
  • Understand that no model can predict your future, certainly not for 30 years. You will need to recalculate periodically.
  • Know the credentials of the model builder.
  • Use models to explore the possible outcomes and better understand the economic forces at play. When you see bad outcomes, try to come up with a way to mitigate them.
  • Be aware that a model is only as good as the input we provide and the assumptions we make and that we can't make very precise assumptions. That means we won't get very precise results.
  • Understand that a model is not a retirement plan. It is one tool to help build a plan.
  • Find a model from a provider that won't profit from the sale of retirement-funding products.
  • Find a comprehensive retirement model that tests several key factors — spending rules, taxes, Social Security claiming, pensions, annuities, life insurance, asset allocation, etc. — and their interactions instead of trying to combine the results of single-purpose models.
I think the best and most comprehensive retirement planning software with a reasonable price tag for consumers at present is Laurence Kotlikoff's MaxiFi.com. It's not free but it is affordable. It can be tricky to use and, again, the more you know about retirement finance, the better the results you can expect. MaxiFi provides all the capabilities that I mentioned above and more and it completely avoids probability-of-ruin issues by maximizing lifetime consumption, instead.

What should you do with this information?

Use all "toolkits" with caution. I have a toolkit in my garage that contains all the tools needed to perform most household plumbing chores but for some reason, my wife still insists that I call a professional plumber.

Given that finding a great retirement planner can be challenging and expensive and that many of this blog's readers tend to do their own planning, it's easy to see the allure of finding a great software package and doing it yourself. Today's software, however, is much closer to a toolkit than to a "robo-advisor." Those who choose this path should avoid being overconfident in the results and should build plenty of safety margin into their plan.



Sample List of Free Limited-Purpose Retirement Planning Tools

  • Estimate a budgetary amount to spend from savings for the current year: Ken Steiner's How Much Can I spend in Retirement SpreadsheetNote: Ken Steiner mentioned to me that my original wording here, "safe amount to spend" should instead say "budgetary amount to spend", as we agree there is no way to predict a "safe" spending amount. Ken's goal is to provide a budgetary spending estimate based on sound actuarial principles.



REFERENCES

[1]Improving Retirement Outcomes with Investments, Life Insurance, and Income Annuities, Wade Pfau and Michael Finke.



[2] Toward Determining the Optimal Investment Strategy for Retirement, Javier Estrada.




Thursday, April 18, 2019

Black Holes, the Higgs Boson and Retirement Planning



The first image of a black hole. Credit: Event Horizon Telescope collaboration et al.

What do relativity theory, quantum mechanics and retirement planning have in common? Not a lot and that's actually an important point.

Black holes were implied by Einstein's work on general relativity in 1915 but the first one wasn't discovered until 1971[1]. Physics was able to predict the existence of one of the largest elements of the universe 56 years before one was discovered.

At the other extreme, on the quantum scale, Peter Higgs and five other scientists proposed the existence of the Higgs boson in 1964. Its existence was confirmed in 2012 based on collisions in the Large Hadron Collider at CERN. The existence of the Higgs boson was predicted 48 years before it could be confirmed.

In retirement planning, we do well to predict finances somewhat accurately more than a year or two in advance. Retirement planning is clearly not rocket science.

Physical sciences and their predictions are based on physical laws of the universe. Acceleration due to gravity on Earth is about 9.8 meters/second2, or about 32 feet per second per second as we Boomers learned in high school physics back when a meter in the U.S. was something one paid a quarter to park. On Mars, it's about 3.7 m/s2. Light travels at about 300,000 kilometers/second.

Drop an object from a height of 10 meters on earth and we can predict that it will reach the ground in 1.43 seconds traveling at 14 meters/second at impact. We can build models that predict such things with great accuracy.

Economics, however, is a social science, not a physical science. Finances can be modeled mathematically but actual outcomes are highly dependent upon the behavior of the humans involved. That makes the models far less predictive.

Unlike the universal laws of physics, the inputs for financial models are often unknown, so we make our best guesses. The most important factor of retirement finance, how long you and your spouse will live, is largely unknowable. Half of a group of people like you may live another 18 years but you might live twice that long or get hit by a bus tomorrow.

We know that stock markets have returned about 9% a year over the past 150 years but you won't be retired for 150 years. The geometric rate of market return you would have historically experienced over any single 30-year retirement during those 150 years could have been less than 3% per year or more than 10%, depending on the year you retired. The range of returns is broader for shorter periods.


Black Holes, the Higgs Boson and Retirement Planning.
[Tweet this]


Retirement models, whether mathematical, spreadsheet, or Monte Carlo simulation, can't predict the future the way models do in the physical sciences. Monte Carlo simulation was developed for the Manhattan project and was accurate enough to help develop the atomic bomb when the world had minuscule computing power. MC could predict how atoms would behave but it can't predict your retirement finances.

In 2001, William Bernstein published a blog post entitled, "Of Math and History" noting that "it’s the engineers who most often give me the willies."[2]
"The trouble is, markets are not circuits, airfoils, or bridges—they do not react the same way each time to a given input. (To say nothing of the fact that inputs are never even nearly the same.) The market, though, does have a memory, albeit a highly defective kind, as we’ll see shortly. Its response to given circumstances tends to be modified by its most recent behavior. An investment strategy based solely on historical data is a prescription for disaster."
(It's a great read, by the way, as is just about everything at EfficientFrontier.com.)

There is huge risk in believing that we can accurately predict our financial future, market risk or returns, a safe amount to spend annually from our savings portfolio, our optimal asset allocation, our probability of successfully funding retirement, or any such metric with any degree of accuracy for any period beyond perhaps a couple of years, let alone for a retirement that could last 30 to 40 years.

I recently told an audience at a retirement finance conference that the greatest risk of retirement is overconfidence. Believing you can accurately predict the things in the previous paragraph is a prime example.

You might rightly ask, given my perspective of uncertainty, why I spend much of my retirement days building simulation models. The answer is that I don't use the models to predict probability of ruin or to predict anything, for that matter. I use them to study many possible outcomes for hints for improving my retirement plan. I readily admit that I have no idea which simulated scenario, if any, ultimately will be similar to mine. I just want to find the bad outcomes and say, "Whoa! How can I avoid those?"

I was once asked what I think is wrong with spreadsheet models of retirement. My answer is that they only consider a single possible scenario and not a realistic one, at that. It is highly unlikely that you will live precisely 30 years in retirement, for example, and there is an infinitesimal probability that your market returns or expenses will be the same each of those years.

(You could run a spreadsheet model several times with different assumptions, of course, but you can generate tens of thousands of different scenarios in a few seconds with MC.)

The tools we use in the physical sciences can be useful in social sciences like economics but they cannot be as predictive because people, unlike atoms, are unpredictable. Computerized retirement models can look impressive when a computer spits out thirty pages of Monte Carlo simulated data but less so when one considers the huge assumptions that have been input into the program. Computers and simulation cannot remove risk from your retirement but they can help identify and understand it.

In an earlier post, I suggested that the most important retirement decision you will make is how much of your wealth to allocate to safer income assets and how much to risk in the market. Don't be overconfident in your ability to predict the risk and returns of the latter. Have a backup plan (a floor of safe income) in case of portfolio failure. Retirement models are simply the best estimates we can make of a largely unknowable future.

And, if someone tells you that you will die with $5.3M in your investment portfolio or that you can spend 4.27% of your portfolio balance annually with a 95% probability of not outliving your savings for at least 30 years, consider that with a large dollop of skepticism.

Remember when you're considering your retirement plan that all models are wrong but some are useful.



Zvi Bodie explains "America's best-kept secret", I bonds for inflation protection.


The American College of Financial Services has created a wonderful collection of brief video interviews with Rick Miller on topics of personal financial planning.


REFERENCES

[1] Who Really Discovered Black Holes?, BBC Science Focus Magazine.



[2] Of Math and History, William Bernstein.