An oddball idea?

[sgeeeee]

Hey!

I knew I liked that guy.

P.S. - now that I'm sucking up - where's my curmudgeon certificate?

Preferably engraved on a framed dryer sheet.

heh heh heh heh

I sent that certificate to your Louisiana address. FEMA assured me they would deliver it.

 

[cute fuzzy bunny]

How much might that be? I'm beginning to waver...

Pretty much nothing since I bought two cars this year!

 

[dory36]

How much might that be? I'm beginning to waver...

Pretty much nothing since I bought two cars this year!

Dang! Back to kneeling at the alter of 4%...

 

[cute fuzzy bunny]

And just remember, CFB, if I weren't so ugly I would have no use for your South Dakota based escort service.

Just so you know, I have to pay the guys extra under the table to take the assignment.

 

[modlair]

You realize that this is not possible. The SWR from historical simulation can never be larger than it is today. It can only go down from additional history. Remember that the SWR is a worst case search of all possible historical sequences. Since we already have sequences that produce SWRs as low as 4%, it can never go up.

Of course not possible.

I confess I don't understand the worst case search comment. Certainly we see that at least one 30 or 40 year subset of our single 130+ sample series yielded failure, but why constrain to 100%? If an average of 1000's of data streams yields an SWR number using a 99% constraint that is significantly higher than that single 130 year series, would we not suspect the single failed 30 or 40 year subset to be an aberration? As soon as you say 95% or even 99%, then that worst case can be swamped out by probabilities -- and they are potentially compelling probabilities.

Another poster says it like so:

Over 4% fails for the data we already have now. Adding 40,000,000 years more data wont remove that failed sequence.

Although you might get 99.999% at 6% from the 40,000,000 years worth instead of the 80% you get now.

If 40 million years of series yielded a higher number at 99.999% success definition, and we had some faith that we had 40 million years of truly "market-like" data, I'm pretty sure the higher number would be get attention and that one failed series would be suddenly very focused upon in a search for why it is an aberration rather than for why it is truth.

But you're right, of course. For a 100% requirement there already exists a 30 year series that fails. What I'm looking for is a standard deviation (which is in raddr's files). Not a single failure.

I don't understand your issue here. Stock performance, bond performance and inflation rate are not random events. They are causal. The causes are complex and not readily described with mathematical models, but that does not mean they are random.

Malkiel would argue this, of course, but that isn't your point, I know.

If human behavior were entirely causal there would be no market. There would never need to be a determination of stock price from bid and asked because the definitive price would be set by causation, not mood. The question is can the sum of influences that define behavior over long duration be modeled so that a large number of market-like series of data are possible? Maybe.

And maybe not. In the absence of consensus from informed folks (and you clearly do understand DSP) there probably isn't enough justification for whatever workload dory would face. Oh well.

 

[dory36]

For a 100% requirement there already exists a 30 year series that fails.  What I'm looking for is a standard deviation...

But if we create a simulation that behaves like the 130 data points to date, and it doesn't have the same standard deviation as the original sample, then how is it akin to the real data?

dory36, who did signal processing and analysis for some unnamed government agency thoughout the Nixon administration...

 

[wabmester]

If human behavior were entirely causal there would be no market.  There would never need to be a determination of stock price from bid and asked because the definitive price would be set by causation, not mood.  The question is can the sum of influences that define behavior over long duration be modeled so that a large number of market-like series of data are possible?  Maybe.

Just because something is determinstic doesn't mean it's simple.    We have to use all the computational power we have just to model some relatively simple deterministic process like the weather, and we still can't predict it all that well.    What makes you think we could even get in the right ballpark with something as simple as a fourier transform of annualized market returns?

 

[dory36]

Don't make me get out my stick... it's still in the closet, just waiting for when it's needed again!

 

[modlair]

But if we create a simulation that behaves like the 130 data points to date, and it doesn't have the same standard deviation as the original sample, then how is it akin to the real data?

dory36, who did signal processing and analysis for some unnamed government agency thoughout the Nixon administration...

The theory offered is that there are core cyclical characteristics of markets and that the 130 yr actual data we have has those characteristics within it, but that the 130 samples presents us with the risk that the 130 samples were "taken" at peaks or troughs of cycles and thus might be extreme.

A larger sample size, by virtue of its size, lessens the probability that all data points in all the series are taken at extreme points in underlying cycles within the composite waveform -- but the same cyclical characteristics from the 130 yr data may still be in the millions of new 130+ sample series. One averages the results for those millions and obtains a result (that is a mean) and one can compute the standard deviation too.

wab raises the valid point of confidence:

Just because something is determinstic doesn't mean it's simple. We have to use all the computational power we have just to model some relatively simple deterministic process like the weather, and we still can't predict it all that well. What makes you think we could even get in the right ballpark with something as simple as a fourier transform of annualized market returns?

A first order attempt at this is in raddr's files. It has already been done. It already produced results that differ from the historical case. His "enhancing" of random data is, perhaps, intuitive and one can embrace it or not -- but the one key thing that is done with that work is the concept of averaging additional series of hopefully market-like data. Whether or not the data is "market-like" is always going to be debateable. What I've suggested is a way to add support to that overall approach because what that approach constitutes is a low pass filter with an indirectly defined pass frequency and sharpness. An FFT might more carefully define what is "market-like".

 

[Yipee-Ki-O]

My head is spinning! Boy, am I glad that dearly(?) departed ***** was not a numbers guy. The double whammy woulda been unbearable No offense intended, mind you. Say, when y'all get this figured out could you devote just a couple of minutes of your time to determine who will win next year's Super Bowl? Forget SWR, with that in hand I can run down to the Strip, place my futures bet and come next February I'll be livin' large!

 

[modlair]

Say, when y'all get this figured out could you devote just a couple of minutes of your time to determine who will win next year's Super Bowl?

Can't help you win your bet, but could maybe help you feel better about the season.

All teams at this point in the year are less probable to win the SB than lose during the regular season and playoff process. Therefore, the odds are with you if . . . instead of being a fan and liking a team -- you chose to hate one instead. This is not unusual. Many fans have a particularly hated team.

If you watch the season and playoffs unfold, you are very much more likely to have occasion to be happy when that team loses enroute to the championship than you are to have to endure the hated team as SB champs.

 

[dory36]

Something isn't making sense here. I'm having a hard time following the logic, and an even harder time figuring out what conclusion we could reach following any hoped-for results.

Your original argument was that 4% was too conservative and additional data would perhaps demonstrate this (even though you previously said "I don't know if there are cycles to human behavior over 130 yrs of time").

SGeeeeeeeeeeeeee pointed out that since 4% was an actual outcome, any additional data could not increase the worst case result, only decrease it.

You then said we shouldn't look at the absolute worst case, only the ~1 - 5th percentile results.

By looking at anything less than 100% with the historical data, we're already eliminating the true outliers. Even if we set the success rate to 70%, we're still under 5% SWR, so those "outliers" would need to be quite prevalent in the data.

Now you're saying that your goal is to avoid potential outliers in the historical data that somehow happened to be recorded by more than 5% of the historical results but would be washed away by fabricating monte carlo data based on these same historical data. In your words,

The theory offered is that there are core cyclical characteristics of markets and that the 130 yr actual data we have has those characteristics within it, but that the 130 samples presents us with the risk that the 130 samples were "taken" at peaks or troughs of cycles and thus might be extreme.

If the idea is that the timing of the yearly withdrawals is creating an abnormally low result, one can look at taking the withdrawals in February, or March, or April, etc., based on historical data.  Intercst did so. At the 95% safe level, results varied a tenth of a percent or so (i.e., 3.95% to 4.05% withdrawals), but there was no hint that the 130 samples taken in January each year created any artifact. In the 30 year runs, averages across the year matched January sampling within 0.02% for the 95% level, 0.1% for the 99% level, and 0.03% for the 100% level.


Obviously there is nothing particular to be gained if the results don't change. Yet another confirming study to well-established findings is good for college exercises, but I'm not sure it advances anything else. So we have to be hoping for a significantly higher SWR. You mentioned a 6% SWR over 40 years, which seems like a good "target" for something useful, beyond the noise level.

But a 6% withdrawal matches up to a 40% success/60% failure rate -- so 60% of the historical examples would have to be outliers...

Let's imagine we did exactly what you suggest, and we obtain a significantly higher result. We are so happy with our outcome that we publish it in the academic/professional journals. The abstract might say...

...
The researchers created a massive set of sample data intended to be "market like" for the purposes of a monte carlo analysis.  The techniques for creating this data set were selected to mimic market behavior with respect to any cyclic patterns. (No basis for these patterns is postulated.) Using this data set, the researchers looked at the results of a periodic withdrawal from a portfolio over an extended time. Contrary to the conventional wisdom and published results (see bibliography), the researchers found that 6% of a starting portfolio could be safely withdrawn for 40 years (p < 0.05%). While backtesting of this withdrawal rate shows that the strategy would have failed in approximately 60% of the actual historical periods for which data are available, the researchers believe those failures were not representative of the actual market behavior due to sampling bias or other unidentified reasons, and therefore should be ignored by those attempting to determine a portfolio balance necessary to last them the rest of their lives.

 

[Cb]

But which day of the month gives the highest SWR? Me, I'm taking my 4.0003654% draw the day after payday...

Cb

 

[Nords]

The SWR from historical simulation can never be larger than it is today. It can only go down from additional history. Remember that the SWR is a worst case search of all possible historical sequences. Since we already have sequences that produce SWRs as low as 4%, it can never go up.

Thousands of SWR posts and I finally learn something new. Thanks, SG! So, you're saying that as we get more historical data, the SWR will drop? Isn't that what ***** has been saying for years?

Rodmail, I hope you appreciate the results of your persistence: SG has uncloaked and agreed with *****.

Rodmail, I see you asking a lot of questions and arguing with people who don't perceive the situation your way, and I see you proposing amazing contortions of historical data to get it to conform with the way you think things should be, and I even see you postulating that people who produce results agreeing with your conjectures would be foolish to publish it.

What I don't see you doing is blowing your own code or even producing original research. Where's a practical application of your ruminations?

Don't get me wrong, I'm not going to do those things either. But you're the one claiming that everyone else is wrong without producing a functioning calculator or research that produces the results you claim. If I was a newbie to this board reading about SWR for the first time, I'd be just as likely to go with your numbers as I'd be to go with *****.

 

[dory36]

The muffled moan you heard was the sound of SG committing seppuku, as the only possible way of restoring his honor.

 

[unclemick]

Ooooooh! Now I don't feel so bad being labeled a bored 13, now 14? year old girl from Missoula.

Hey - MO is closer than LA.

Or get out my Norwegian widow soapbox and talk dividends, current yield and pssst - Wellesley.

Nah - I like FireCalc just like it is - old and advanced.

heh heh heh heh heh heh heh heh heh heh heh heh - BTY for my 13th year of ER at the old age of 62/63 gonna  try 5% varible for while. Mainly cause I'm not getting any younger.

 

[Dawg52]

BTY for my 13th year of ER at the old age of 62/63 gonna  try 5% varible for while. Mainly cause I'm not getting any younger.

Just curious, what you gonna spend your money on? Women and booze? 

 

[cute fuzzy bunny]

Wimmen with booze in 'em...

 

[unclemick]

Yep

Seemed to help a lot on the cruise - unless there's a cruise version of 'they all look good at closing time.'

Down to remodeling and perhaps some more travel.

Now a rich wider women - defined as richer than me and still healthy(I'll defer that definition to the reader) - that's worth looking for.

Not toooo many Norwegians in this part of MO.

heh heh heh heh heh - lot of openings in this post - I'm sure someone can't resist - he he.

 

[modlair]

By looking at anything less than 100% with the historical data, we're already eliminating the true outliers. Even if we set the success rate to 70%, we're still under 5% SWR, so those "outliers" would need to be quite prevalent in the data.

But a 6% withdrawal matches up to a 40% success/60% failure rate -- so 60% of the historical examples would have to be outliers...

Two persuasive points. I'm pressed for time right now and will revisit them in the context of what they imply, but for now just wanted to get a comment out that acknowledges these are strong points.

 

[sgeeeee]

The muffled moan you heard was the sound of SG committing seppuku, as the only possible way of restoring his honor.

No. It was the sigh of frustration at having a true statement undergo such an innacurate interpretation.

Nords, I've noticed you have been especially caustic and innacurate lately. Maybe it's your diet. Too much pineapple? Or maybe you haven't been getting enough exercise. Inferior surf?
:

 

[modlair]

By looking at anything less than 100% with the historical data, we're already eliminating the true outliers. Even if we set the success rate to 70%, we're still under 5% SWR, so those "outliers" would need to be quite prevalent in the data.

Quote
But a 6% withdrawal matches up to a 40% success/60% failure rate -- so 60% of the historical examples would have to be outliers...

Adding a third quote from your input, dory

If the idea is that the timing of the yearly withdrawals is creating an abnormally low result, one can look at taking the withdrawals in February, or March, or April, etc., based on historical data. Intercst did so. At the 95% safe level, results varied a tenth of a percent or so (i.e., 3.95% to 4.05% withdrawals), but there was no hint that the 130 samples taken in January each year created any artifact.

Dealing with the last quote first, varying the month of the sample and observing no change of result is powerful data and would surely seem to suggest that :

1) There is likely no evidence human market behavior occurs in cycles, i.e., history does not repeat regularly or

2) There is some weird and unlikely aliasing beyond Nyquist in cycles that are present and therefore are not detectable (and in this case what isn't detectable is irrelevant) or

3) There are cyclical influences of various sorts on human behavior/moods and they are of possibly different frequencies, but they are numerous, have random phase relationships (such that there is no point in time that yields a "peak or trough" in behavior, and near equal amplitude (with amplitude being defined as how decisive a particular cycle might be in causing urgency in buying or selling (it is not buying or selling that moves prices, it is the urgency to buy or sell that does).

Occam's razor would say #1 is the item to bet on.

But what must this mean for your first two points, dory.

You have found at least one 30 yr series that fails even marginally above 4% so another 1000 yrs of data will not make that 1 series go away and at 100% confidence you can never have higher withdrawls. But ignore your 2nd point for a moment and consider what it would mean if another 1000 points did exist and that 1 series that failed at 4% + delta was the only one that did. We would then have to deal with the probabilities that in the first 137 data points ever encountered/created by mankind out of possibly thousands of market decision moods to come that we happened to catch the one time that 4% + delta failed.

But yes, I know, you have your second point that shows even modest moves above 4% collapse the total number of 30 yr periods that are successful. What does this say about human behavior/moods in markets? It suggests that it's always the same, doesn't it? There just aren't going to be many 30 yr series at all that will permit 4% + X*delta or whatever because humans behave the same way and they don't allow the market to be strong enough over many protracted periods to support 4% + X*delta.

Does this seem reasonable? Is this not in conflict with the failure of technical analysis? Or indeed, with the first point above. If humans do the same thing over and over, that's . . . a cyclical behavior.

Anyway, no question about the strength of your points, but what seems like a proper extrapolation from them creates many conflicts/questions.

 

[justin]

I like the Occam's razor bit. It suggests taking 4% or so.

 

[dory36]

There just aren't going to be many 30 yr series at all that will permit 4% + X*delta or whatever because humans behave the same way and they don't allow the market to be strong enough over many protracted periods to support 4% + X*delta.

Or there is some sort of a feedback mechanism buried in there somewhere that has the same regulating effect.

My grad school advisor had a great one-liner... "argue with hypotheses all you want, but don't argue with data."

One of his great lessons was making us find numerous (often mutually inconsistent) hypotheses to explain data, while recognizing that the hypotheses were just that, but the data was a representation of the real world.

 

[modlair]

My grad school advisor had a great one-liner... "argue with hypotheses all you want, but don't argue with data."

One of his great lessons was making us find numerous (often mutually inconsistent) hypotheses to explain data, while recognizing that the hypotheses were just that, but the data was a representation of the real world.

Interesting. My grad school advisor had one, too. It was "I have more of a gravitational influence on you than Jupiter."

His point was that there is a county in Wisconsin whose weather has conformed each day for decades (p < .01) to a particular sequence that correlates with the phases of the moons of Jupiter.