My Retirement-Finance-Model "Topography" Strawman

A model is best used as a decision-support tool rather than as a return-prediction tool.

- Patrick Collins 

I'll freely admit that the following figure might be naïve or reductive or reflect my biases but I thought I'd take it out for a trial run. Over 10 years of doing this kind of stuff, I might say -- for myself anyway -- that retirement finance models can break down into the kind of dimensionality I see in Fig 1...if we squint our eyes and don't look too close:

Figure 1. My opinion of Ret-fin models in 3 dimensions

My attempt to explain Figure 1:

X: Components to Integrated - is my shorthand for using individual chunks of analysis like a single formula or a spreadsheet vs going to highly sophisticated software that puts everything together in a million lines of code that integrates everything the modeler thinks to integrate.

Y: Simple to complex - this could be from deterministic formulas or even ad-hoc rules of thumb (even just judgement) to stochastic and dynamic^n models with autoregressive inflation or Cholesky decomps or regime switching or stochastic volatility or...

Z: Novel to Fit - In reverse order, fit here means a model that is highly tuned to our understanding of the world based on past history. Novel means the client, or the modeler, is wide open to having a model do something, anything, that does not have to hew itself to history. It is open or new or more or less unconcerned with exactly how the world worked in the past 

----------

A presumption I often see in the Ret-Fin universe is that it seems to be that movement from a to b in Figure 1 is considered "good," a tendency to which I could be accused myself.  Is it?  I mean, I might have set this post up a little tendentiously but I think a case can also be made for either "not so fast" or b->a might be "better" ... sometimes. This post is my attempt to verbalize to myself why that might be the case. I mean, in mathematics and economics, there are in fact more numerate and mathematically necessary ways of doing things, things that are often jumped over in even some of the scholarly papers I've read. In addition, I will stipulate that there is also (maybe) a simplicity on the far side of complexity (Justice Holmes if I recall). But sometimes we can get a little too smart for what is -- au fond -- necessary to our survival and prosperity. Let's ignore the super bright like Ed Thorpe where prosperity came from seizing complexity by the _____. 

I realize at this point in the post that while my original goal was supposed to be about a one paragraph long riff, I now have a few more thoughts on hand. So, in order to not go overboard, and certainly I have been neither scientific nor exhaustive here, I'll just freestyle some comments I have on Figure 1 and leave it at that: 

Opacity and bias. One problem with "point b" is that it can become a little obscure even to the modeler. The formulas have been put into a box and the windows have been painted black. How does everything interreact and what does it all mean? Has it really been tested well? Can it be explained successfully to Grandpa? Does it really have any (quantifiable) decision support value (since we know we can't predict the future)? To what degree does all this complexity create either a farrago or a biased view of the world utterly unrelated to the model user. Maybe call this kind of qualm "model risk" but for me it is even more than that. I have certainly made my fair share of very complex models. I create, I test, I add, I blog, I add, I add, I add, I test, I blog, I set aside. Then a year later I literally cannot figure out what I did or why. This says a lot about me, of course, but really, is that model really useful now? Probably not but I'll shield myself with the weak "I'm an amateur" thing. As an aside, this is why simple or closed form math is useful vis-a-vis a black box integrated monster. It is transparent and shows both the inputs and the shape of the interactions to the naked eye. It fosters discussion where the hyper-complex only cows us. 

As an example: Normal vs Fat tailed returns. A wise modeler will acknowledge that normal return distributions are not realistic and that they underrepresent the risk of chaotic, cataclysmic events. Reasonable critique. There are ways around this, though, of course. Things like alt-distributions like Student's T work a bit, as do mixed distributions of different flavors. In addition one can, like I did here once, add in a chaotic overlay using chaos theory or earthquake math. This nudge is probably necessary in order to disabuse one of the complacency inherent in the "normal" assumption. While this nudge takes head-on the ability of the world to deliver 1987 style events, I still think 1987 gets a little bit laundered in a model. It is in there but if I ran a fat tailed model in late 1986 I doubt I would have really been psychologically or financially prepared for the following October. Ignore that we recovered really really fast thereafter. Different discussion.  

So the Fat-Chaos nudge is good, right? Probably. Certainly it shows off some intellectual consideration and may provide a scholar with tenure or a professional with promotion. On the other hand no one ever really quantifies the net increase in the "decision support value" provided by the nudge. It is only implicit. The other side here, imo, is that the normal assumption is very easy to code and explains quite a bit of the underlying process. My go-forward approach for me is to stay more towards the normal assumption and (sometimes, when it is warranted and useful) add a very, very easy to explain "something" to reflect the nuttiness that can happen in the world ... and then still hold onto my chair in real life because the real world in real-time will never ever be as neat as a model.  

Japan > 1989.  Models that fit themselves to history are very naïve no matter how we design them. The assumptions are just that, assumptions: 1987 won't happen or, if it is modeled, the hit this year is under-appreciated; we assume that the equity risk premium is there forever; we unreasonably take expected returns as a prediction; we forget the single paths we might follow; we unreasonably assume the future to be like the 20th C American past; etc... Even if we model worst case scenarios or the 10th percentile outcomes I still wonder what it would have been like to have had those conversations with a Japanese retiree allocated to the Nikkei in Dec 1989. All that modeling and chatter probably did not capture the "lived experience" of the retiree over the next 20 years even though I suppose it was "modelable" on the front end. Collins again: "Although a good model will capture some statistical behaviors, no model can hope to replicate reality."  In the Nikkei example, the return premium failed and I am certain that every year, for 20+years, I would have been convinced that this year things were turning around, good or bad models notwithstanding. No complex-fit-integrated model would have saved me at all. 

Milevsky's Simple Approach. While I will freely admit that Moshe Milevsky can get a lil complex sometimes -- I once paged down for many minutes through something like 250 pages of advanced calculus on something he did in ret-fin and I myself have created some excruciatingly complex programs to analyze retirement in a not entirely un-aligned way -- he can also be very, refreshingly simple. Some of the greatest things in ret-fin that I've seen so far in the last decade, some of the things that got me into this blog stuff, were things like what I saw the first few chapters of his "7 Equations" book. To be reductive, it was like this: 1) use one simple deterministic formula to estimate portfolio longevity, 2) guess, or fantasize, about how long you might live, and 3) then have a conversation about the tension in the two. That's it. That might be more powerful in a retirement finance consultation than using the blackiest of the black boxes with super duper gradient hill climbing multivariate algos.  Really.

Ken Steiner's Simple Approach. Ken, at http://howmuchcaniaffordtospendinretirement.blogspot.com/ has a pretty simple approach. I was going to call it "blindingly simple" but it does take a little bit of thinking to get into it. He says: use an "actuarial balance sheet" (ABS). That's it. Basically ABS is a feasibility test which, when run dynamically, becomes a sustainability methodology.  It is exactly what I use myself more than any model I have ever built. There is no super-integration, no OCD historical model-fit, no excess complexity. We are way closer here to point a than b. Is a worse than b here? I seriously doubt it. Is there stochastic volatility or stochastic present value or Cholesky decompositions? Nope, and the ABS approach does not suffer for its lack. Me? I might add some other flourishes like an income statement and a spending control method (and periodic peeks at some of the weird stuff) but, my blog notwithstanding, this ABS thing is much closer to a than b. 

Optimization and Adaptation. As I mentioned above, in the quantitative disciplines there are of course better and worse ways of doing things. One critique, amongst several, of the Bengen 4% rule was that it is not really a mathematically "necessary" outcome. It is a rule of thumb. But I think even Bengen knew that. It was just all of us rubes grasping at straws that were the problem. On the other hand, the last 30+ years have seen a lot of advancement towards more numerate and formal ways to apprehend the problem with which we are dealing in Ret-Fin. I don't have the full list of names but think of Yaari, Merton, Milevsky, Irlam, Sharpe and others stretching out to the horizon. But here is my beef with sophisticated "optimal" approaches beyond just the obscurity and bias and complexity. It is that it is not always a practical, usable "decision support" mechanism for people like me. It is, rather, (sometimes, not always) an exercise in academic or professional sophism, ego and careerism (that is in no way pointed at those names, all of whom I respect immensely. This is pointed at the random whomevers layers and layers down). Also, in my opinion, the minute, the instant, an optimal solution is asserted, it starts to decay given the changes in the state of the universe. I think that turns an optimal "solution" into a process-methodology which is an entirely different beast altogether, a beast that can be aimed at the challenge we are discussing using both simple and complex tools. In addition, it is not just a simple vs complex binary, it is "one versus many," something I forgot to throw into Figure 1. In process-methodology world, imo, using many models to "triangulate" has way more power than a single-source construct even if it is "complex."  I can't really back that assertion up here in this post but I believe it. Maybe another post. 

---------

There is no doubt more to say on this but maybe later. My main point is that I personally moved, as many do, from point a to point b over the last 10 years. While that was "fun" (fun in quotes of course, heh) and useful and interesting and mostly correct, it was also a distraction sometimes. It did inform my thinking on this odd topic. But mostly I am moving now in a bee-line from b to a.