To my eye, one of the sharpest and most articulate of the new generation of retirement finance writers is Derek Tharp, now writing at Kitces.com. I will be watching his future contributions to the genre closely.
Derek has just written another top-notch and perceptive blog post, this one on simulation titled "Does Monte Carlo Analysis Actually Overstate Tail Risk In Retirement Projections?" This is a great post and demonstrates a solid command of the subject area. My interest here, though, is not to report on the article or summarize or even bullet-point what he is saying. You can read it yourself . I just wanted to highlight a couple things where I have a slight divergence of point-of-view. I suspect that one would actually have to read the article to understand where I am coming from but I'll leave that in your hands. With respect to simulators overstating risk via fat tailed distributions:
1. Whatever the underlying reasons that a simulator might overstate risk, whether it is due to the use of annual or more frequent return series, the presence or lack of serial correlation or mean reversion, or whatever else, the main reason that simulators overstate risk can be summed up like this: that is why they were created. Historical return series, especially those like 20th century US large cap data, under-imagine the breadth of possible future worlds that could unfold. Simulators are supposed to come up with worse scenarios than we can possibly imagine. It's the whole point.
2. While I like the article, Derek kinda falls into the fixed planning-age trap which many articles do as well. As an early retiree, where risk is palpably different than it might be at 70, this sometimes irks me. His examples use 30 years and assumes, I think, standard retirement ages. He appears to validate a 4% spend rate but doesn't consider someone retiring at 48 for whom a 4% recommendation, depending on how you look at the universe, might be considered malpractice. Nor is someone who retires at 75 considered for whom 4% would be a serious disservice to what little time he or she has left.
3. To quote the article: "Ideally, Monte Carlo analysis tools would allow a combination[*] – such as reduced real returns for 10 years, followed by normalized returns with mean reversion – but, unfortunately, no financial planning software is yet built to provide such regime-based Monte Carlo projections." I've commented on this before. While I freely admit that coding serial correlation and mean reversion into a sim is hard (I tried and failed), the statement that there is no financial planning software that provides regime-based MC projections is false. It would be more accurate to say that there is no commercial software or maybe there is "no software that I personally know of." First, as a comment-er pointed out Flexibleretirementplanner.com allows it. Second, it's not that hard to code a simulator that does it. I wrote a whole simulator with fairly robust features, including stochastic longevity two different ways, in about a week and the regime-based part of it, one of my first features added, took about a minute. It might not be perfect but it works.
4. If Derek wants a real world data series with more risk than the historical rolling model he uses as a baseline, he might want to look at Japan in the '90s. There are no doubt others.
5. I have a deep appreciation for the pros and cons and the good and ill uses of monte carlo simulators. I coded three of them on two different platforms. I am retired and use them as part of an ongoing triangulation process. The one mis-use of simulation I don't see highlighted brightly enough here -- above and beyond leaning too hard on the conclusions that come from over or understating tail risk -- is that real retirees only have one bite at the apple (or one whack at the cat as someone once called it). We do not live a life as a probability distribution. We only have one chance, not 10,000, to get it right. That's why the worst case matters so much: we must keep a weather-eye on the world as it unfolds while also having a lizard-brain-type memory of worst case scenarios.
Fat tails or not, simulate or don't...the advice to myself that I put in a prior post still stands even though it's a little self-absorbed to quote oneself: "a better game might be to play conservative early, test the wind each year, triangulate using whatever information is [and tools are] currently available and then maybe loosen up as the distance to the end shortens...even though it'll be less fun than it would have been 10 years before."
* combination of: 1) mean reversion/serial-correlation, and 2) regime based returns where they might be suppressed for 10 years or so.
No comments:
Post a Comment