Summary Thread: SWR Definition

Financial Independence/Retire Early -- Learn How!
User avatar
ataloss
**** Heavy Hitter
Posts: 559
Joined: Mon Nov 25, 2002 3:00 am

Post by ataloss »

I have coined a new term, the Historical Database Rate (HDBR) to draw attention to this distinction. This is essentially the same thing as the older term, the Historical Safe Withdrawal Rate (HSWR). The Historical Database Rate (and the HSWR) is not a Safe Withdrawal Rate calculation.


Have you cleared this with hocus?
This is where you and me part company. You are sneaking in by way of a parenthetical the old BenSolar idea of an hSWR. I reject that concept. I don't mean to sound harsh, but I view it as a nonsense concept. I don't see how there could ever even be such a thing as a hSWR.

The phrase "Safe Withdrawal Rate" means something. It is a number that provides an assessment of the probabilities of various future possibilities, based on what has happened in the past. That's my working definition. Those are my words, but I believe that the definitions used in the studies are in accord with those words.
hocus 8/2

I would hate to see anything sneaky....

That said, both raddr and BenSolar have made big mistakes in the course of this thing. The idea that <b<BenSolar was pushing for awhile, that we should put little "h's" and little "f's" in front of the capital letters "SWR" was a disaster. That lost us a lot of time. His intent was good. The initiative was a big mistake.
hocus 7/27
Have fun.

Ataloss
User avatar
ataloss
**** Heavy Hitter
Posts: 559
Joined: Mon Nov 25, 2002 3:00 am

Post by ataloss »

btw I want to again recommend the draft faq from 12/02

http://nofeeboards.com/boards/viewtopic ... t=faq#p725


Alternative Methods of Determining Safe Withdrawal Rates (12-10-02 draft)

A. The Historical Sequence Approach
1. This method uses historical data directly.
a) It relies on minimal assumptions. Ideally, it would make none. But the historical record is incomplete in some respects. Best estimates fill in the gaps. For example, the SP500 index itself has not been around for long and index funds have been around even less. It is necessary to extend the historical record.
b) To a large extent the limitations of the historical record are common to all approaches. However, their effects are more pronounced when using historical sequences. Although it is possible to synthesize a reasonable (backwards) extension of the SP500 index, that is not true for many asset classes. Quite a few have extremely short histories. REITS constitute a major asset class that is very new. Income producing properties such as apartments have been around for a long time. But meaningful statistics are quite limited. Most records are very short. Most of the information available is strictly local in its application.
c) The historical sequence approach is inefficient in extracting information. The effective number of data points is quite limited. This is also an advantage. The historical sequence approach is inefficient in extracting false information as well.
d) The historical sequence approach has a credible basis. It cannot project future events that could never happen. After all, the sequences themselves have all happened once.
e) The credibility of the historical sequence approach is easily exaggerated. There are many qualitative changes that have taken place over the years. Some of these are dramatic enough to exclude certain projections from the historical record. Examples include the introduction of the Federal Reserve, the move away from the gold standard and both the introduction and numerous changes in income taxes.
f) The historical sequence approach produces a unique set of answers. They can be replicated. At a very detailed level there are differences among models. For example, results differ for different choices of days, months and years in making calculations. Results from investing each month differ from investing once a year. Results from investing at the beginning of a month differ from those at the end of the month.
g) The historical sequence approach is transparent. It is easy to identify assumptions.
h) The historical sequence approach often makes it easy to identify cause and effect. One can look into the general history of events to interpret what happens in any particular sequence. It is always a great advantage to know cause and effect relationships (as opposed to a simple correlation of events). It helps you estimate the reliability of your projections.
B. Deterministic Models
a) Deterministic models are based on information extracted from the historical record.
b) Examples include estimates of the periods and magnitudes of stock market cycles.
c) It is common to introduce a degree of randomness. In its simplest form randomness is added at a model's output and it is easily understood. Eventually, there can be enough random elements that the model becomes a Monte Carlo model.
d) Some of the elements in a deterministic model are based on cause and effect relationships. Others are from simple correlations. Their root causes are not known. That makes their projections are unreliable...even though they may turn out to be accurate.
e) There are many successful deterministic models. Often, the most successful are also the simplest and most transparent. One example is the projection that the stock market grows at 10 to11% annualized in nominal dollars or 6% to 7% in real dollars. Another example involves seasonal variations. There are good days to buy and to sell for each month. There are good months to own stocks and there are not so good months.
f) Transparency is important since it can help you avoid making serious errors. An example of a hidden assumption is the Federal Reserve's model for determining the fair value of stocks. It is based on limited data. During the short time history behind those data, the effects of inflation on the stock market were embedded in the treasury bond yield curve. Whether that will continue into the future is uncertain.
C. Monte Carlo Models
a) Monte Carlo models are based on information extracted from the historical record.
b) There are different degrees of transparency among different Monte Carlo models.
c) Monte Carlo models introduce randomness in a complex fashion. There can be many elements with random components.
d) With a Monte Carlo model a random number generator provides a value for each element with a random component during each run. Those values vary from run to run. After making a large number of runs, you look at the distribution of the results. They form your estimate of the true probability distribution.
e) Monte Carlo models can produce answers with as great a statistical confidence level as you desire.
f) Using Monte Carlo models does not guarantee that their results are accurate. Accuracy depends on how well information is extracted from the historical record.
g) Using a well-designed Monte Carlo model, you can make reasonable projections at the extremes of a probability distribution because of their large number of runs.
h) It is necessary to determine when the outputs from a Monte Carlo model can be relied upon. That determination is non-trivial.
i) There are several excellent Monte Carlo models.
j) Raddr's advanced Monte Carlo model emphasizes what he calls reversion to the mean. The definition of that phrase is somewhat ambiguous. What raddr's model takes into account is that the spread of actual stock market returns is much less than is usually projected. Projections usually decrease the spread (standard deviation) by a factor approximated by 1/SQRT(N), where SQRT means "square root"? and N is the number of years of the projection.
k) Gummy's (Professor Ponzo's) Monte Carlo model emphasizes the historical sequences of inflation. Year-to-year stock market fluctuations are treated as random (with a historically based non-normal distribution). Inflation sequences are selected at random but not year-to-year inflation fluctuations.
m) I would personally like to see a Monte Carlo model that uses a long-term cyclical variation to account for behavior of the spread in actual returns. We know that there are some factors in the stock market that vary slowly. The SP500 price to earnings ratio (using Professor Shiller's P/E10 numbers) has varied slowly. Most recently, that multiple increased over a period of two decades, the 1980s and the 1990s.
D. The Reliability of Projections
a) It is a serious mistake to rely on either the historical sequence approach or a Monte Carlo model by itself. Both approaches have something to add.
b) I use the dory36 calculator for my investigations. It is great for rejecting things that do not work. Not only is it fast, it is quite credible when indicating failure. It helps me understand the results since I know when effects would have occurred. By running a variety of conditions, I can gain insights as to possible cause and effect relationships.
c) I am not satisfied with the details of any positive result that I get from the dory36 calculator. I am comfortable with certain broadly defined conclusions when I have cause and effect. But the numbers themselves (such as for a safe withdrawal rate or a number of failures) are uncertain. There may be hidden sensitivities. Exact sequences can cause results different from typical sequences.
d) One thing that raddr has done is to match his Monte Carlo model results to the entire distribution in the historical record. It comes close enough to give me confidence in small extrapolations outside of the historical range. His results are traceable to a large number of small variations. Those results do not depend on one or two critically significant events in isolation.
e) The historical sequence approach has an inherent granularity built into it. There are only three or four completely independent sequences in the historical record. There are many partially independent sequences. There are 130 data points that form these sequences. The statistical behavior lies between those extremes. It is better than just having 3 data points. It is worse than having 130 data points. The historical sequences form an estimate of an underlying distribution. When you desire a high degree of safety, you are looking at a tail of the underlying distribution. You have very few relevant sequences with limited independence when you make your estimate.
f) There is another factor to consider whenever making a projection. It has to do with the underlying distribution. This question can never be answered entirely in advance but it must be addressed. What characteristics of the underlying distribution will remain the same and what characteristics will change in the future? For example, we have been well outside of the historical range for valuations since the mid-1990s. Should we ignore that fact? I think otherwise. Then, there is a question about the sequences themselves. To what extent is it necessary to consider the recent past? If you look at 50 year sequences, all of the sequences since 1950 (1952 actually) are incomplete. I can certainly believe the failures among the partial sequences. But I cannot believe the successes.

Have fun.

John R.
Have fun.

Ataloss
JWR1945
***** Legend
Posts: 1697
Joined: Tue Nov 26, 2002 3:59 am
Location: Crestview, Florida

Post by JWR1945 »

Not applicable, ataloss.

The fallacy that hocus was concerned about was not the use of the letters f or h. It was the part that included the letters SWR. The fSWR and hSWR describe things that are different from Safe Withdrawal Rates.

Have fun.

John R.
User avatar
ataloss
**** Heavy Hitter
Posts: 559
Joined: Mon Nov 25, 2002 3:00 am

Post by ataloss »

glad you got the approval for hdbr that was missing for hswr :wink:

We are making a lot of progress
Have fun.

Ataloss
Post Reply