Juggling Trinity Study Data AOF Style

[Complete_newbie]

Im the last person here who'd just take word for it.

I've run some MCs but not on stocks etc, mostly just to learn the math and programming. I'll PM you my thoughts and some questions. I would say that most MCs generating stock forecasting simulations use normal or log normal distribution assumptions are they not ? I have no empirical data but whatever autocorrelation or serial correlation MC includes is minor difference vs a non parametric bootstrapping method. May be I'll run both and let you know or AOF can do it and see results.

[Hatton]

I used to think I understood math.

[Dont_know_mind]

hatton1 wrote:

I used to think I understood math.

Click to expand...

I thought I understood inductive reasoning. I looked at the wikipedia page and half the names are people I hadn't heard of.

When I was at university, all we were taught was the inductive problem (Hume) and Karl Popper's solution to that (hypothetico-deductivism and falsification i.e RCT's). It seems like there has been some development in the area, with mathematical formalization of Popper's soft rules.

https://en.wikipedia.org/wiki/Inductive_probability

https://en.wikipedia.org/wiki/Ray_Solomonoff

https://en.wikipedia.org/wiki/Algorithmic_probability

It will be interesting to see how the machine learning affects medicine.

 

 

 

[Donnie]

Complete_newbie wrote:

Im the last person here who’d just take word for it.

I’ve run some MCs but not on stocks etc, mostly just to learn the math and programming. I’ll PM you my thoughts and some questions. I would say that most MCs generating stock forecasting simulations use normal or log normal distribution assumptions are they not ? I have no empirical data but whatever autocorrelation or serial correlation MC includes is minor difference vs a non parametric bootstrapping method. May be I’ll run both and let you know or AOF can do it and see results.

Click to expand...

AOF isn’t using a normal or lognormal distribution.  That would be simplistic, but defensible.  He’s using a distribution consisting of an equal weighting of all historical returns with no other returns possible.  It’s a worthless exercise.  No takeaways possible.

[Complete_newbie]

Not talking about AOF. Talking about the MCs thrown on websites that many use. I bet they use normal distribution. If you have a more sophisticated MC somewhere let me know. I bet results of SWR don't change much. So while you call this worthless, it's useful still saying if assuming no correlation of stock price your SWR isn't much different.

Either way MC is model based on assumptuons. You can't fit enough variables out there for generative models

For the recrod if there is enough interest I can doachine learning on time series data (I'll take that challenge just to learn new skills ). Seems to me people really like SWR type of stuff; possibly Genesis of his post.

[LIFO]

 

Donnie wrote:

AOF isn’t using a normal or lognormal distribution.  That would be simplistic, but definsible.  He’s using a distribution consisting of an equal weighting of all historical returns.  It’s a worthless exercise.  No takeaways possible.

Click to expand...

Can you explain this thought a little more?  Do we know if annual stock market returns are distributed in a normal or lognormal fashion?  If this is all the data that we have, the distribution is the distribution. Are you concerned that he is sampling without replacement?

I'm not a statistics whiz but the real limitation to better analysis seems to be limited data.  Is it possible to do a better assessment of stock market returns with the current data we have?  More than a few of us here have experience with robust scientific methods and and statistical analysis.  Can we show them how it should be done?

[Donnie]

LIFO wrote:

Can you explain this thought a little more?  Do we know if annual stock market returns are distributed in a normal or lognormal fashion?  If this is all the data that we have, the distribution is the distribution. Are you concerned that he is sampling without replacement?

I’m not a statistics whiz but the real limitation to better analysis seems to be limited data.  Is it possible to do a better assessment of stock market returns with the current data we have?  More than a few of us here have experience with robust scientific methods and and statistical analysis.  Can we show them how it should be done?

Click to expand...

AOF is assuming that the historical returns represent the entire distribution of possible return results.  If a particular return didn't happen in the historical period, it cannot happen in the future according to that model.  This is obviously not true.

You are right.  We do not have a lot of historical data, which is an issue.  However, (i) we can use historical returns to derive a statistical model to project future returns or (ii) we can use historical returns for scenario testing, i.e. how would my portfolio have looked in the 60's.  That's it.

Complete_newbie wrote:

Not talking about AOF. Talking about the MCs thrown on websites that many use. I bet they use normal distribution. If you have a more sophisticated MC somewhere let me know. I bet results of SWR don’t change much. So while you call this worthless, it’s useful still saying if assuming no correlation of stock price your SWR isn’t much different.

Either way MC is model based on assumptuons. You can’t fit enough variables out there for generative models

For the recrod if there is enough interest I can doachine learning on time series data (I’ll take that challenge just to learn new skills ). Seems to me people really like SWR type of stuff; possibly Genesis of his post.

Click to expand...

I'm not saying all MC is worthless, just this particular exercise.

 

[Zaphod]

LIFO wrote:

 

Donnie wrote:

AOF isn’t using a normal or lognormal distribution.  That would be simplistic, but definsible.  He’s using a distribution consisting of an equal weighting of all historical returns.  It’s a worthless exercise.  No takeaways possible.

Click to expand…

Can you explain this thought a little more?  Do we know if annual stock market returns are distributed in a normal or lognormal fashion?  If this is all the data that we have, the distribution is the distribution. Are you concerned that he is sampling without replacement?

I’m not a statistics whiz but the real limitation to better analysis seems to be limited data.  Is it possible to do a better assessment of stock market returns with the current data we have?  More than a few of us here have experience with robust scientific methods and and statistical analysis.  Can we show them how it should be done?

Click to expand...

Its been done, theres nothing further really needed. Why is there this strong overwhelming desire to pulverize the data into something new? I get there are a lot of blogs and everyone desires a new take on the same old same old, but just putting the same stuff through a different combination of data techniques doesnt make the outcomes any more applicable than whats already been done.

I already mentioned further up what is a very well known aspect to market returns, they are not perfectly normally distributed, have fat tails, mostly where you dont want them (the left). This is so well know and hashed out thoroughly you can easily get brilliant run downs on it and see it applied to asset allocations, draw down strategies, etc....Here is an example. Quick rundown on kurtosis, etc...http://www.statisticshowto.com/probability-and-statistics/statistics-definitions/kurtosis-leptokurtic-platykurtic/



Then you have skew, etc....here is a quick rundown of the terms. https://www.evestment.com/resources/investment-statistics-guide/assessing-skewness-and-kurtosis-in-the-returns-distribution/

A great example is the VIX, everyone always talks about it either being high or low or above/below average. Average is somewhere around 18 long term, which is actually short term since its only been around a while, not a ton of samples. However, whenever its less than that people start talking about it as if something is wrong, etc...VIX is very non normal with the fat tails on the right in this case. The most common reading on a daily basis is 12-13 however, way below average. Highly kurtotic and skewed, varying in different vol regimes.

[Dont_know_mind]

I would do this online course if I had time.

https://www.santafe.edu/engage/learn/courses/algorithmic-information-dynamics-networks-cells

It's only $50.

[Complete_newbie]

Think we have beaten this horse enough

Look all of these randomized sampling are based on assumptions. I understand MC is a generstive sampler based on a model and better but not perfect. Think AOF is just doing experiment of what it'd look like if returns were random event.

[Donnie]

Complete_newbie wrote:

Think we have beaten this horse enough

Look all of these randomized sampling are based on assumptions. I understand MC is a generstive sampler based on a model and better but not perfect. Think AOF is just doing experiment of what it’d look like if returns were random event.

Click to expand...

We clearly haven’t beaten this horse enough because what you wrote is likely wrong but certainly unclear.  Monte Carlo is a technique to predict the outcomes of deterministic events that contain randomness.  In this case, rate of return is random and the amount left in the portfolio after draw down and portfolio return is deterministic, i.e. if you knew the annual return and annual drawdowns for the next 30 years with certainty, you could predict the portfolio balance with certainty.

MC’s predictive ability rests solely on how accurately the probability distribution used represents the actual probability distribution.  If you had a perfect distribution, you would have perfect predictive ability using Monte Carlo.

AOF is using Monte Carlo techniques, he’s just using the wrong probability distribution, which invalidates the conclusions.  He’s trying to provide nuance to the 4% rule, which he can’t do unless he uses a better probability distribution than the one he’s using.

 

[Complete_newbie]

Donnie I understand full well what the techniques are but that's not his goal. Predictive capability is NOT is goal (as far as I can tell). And you're telling me there is a perfect distribution model on which MC draws random samples from ? Show me. You have to have a model that defines that distribution first. As far as I know normal or log normal is used in majority of MCs - please refer me to other ones. Model only as good as inputs and variables which NO model can take in.

What AOF did is an experiment thinking what if distribtuik is equal and random and Drew samples. Who cares. I saw it as am approach to think about of not really applicable.

[Donnie]

Complete_newbie wrote:

Donnie I understand full well what the techniques are but that’s not his goal. Predictive capability is NOT is goal (as far as I can tell). And you’re telling me there is a perfect distribution model on which MC draws random samples from ? Show me. You have to have a model that defines that distribution first. As far as I know normal or log normal is used in majority of MCs – please refer me to other ones. Model only as good as inputs and variables which NO model can take in.

What AOF did is an experiment thinking what if distribtuik is equal and random and Drew samples. Who cares. I saw it as am approach to think about of not really applicable.

Click to expand...

He is precisely quantifying safe withdrawal rates to two digit precision on the withdrawal %.  How could predictiveness not be the goal.  But if you are going to do better than a SWR of ~4%, you need a better, more accurate probability distribution, not a worse one.

Yes, there are numerous papers written on the distribution of stock returns.  People run quantitative hedge funds based on these models.  No, these guys aren’t writing blogs for the FIRE community and shipping content out for free, but that doesn’t mean they don’t exist.  Why would you think normal distributions is the “cutting edge” here? Just google “Monte Carlo stock market returns” and “fat tails”, “autocorrelation”, “mean reversion”, or anything else you want to learn about.

You can think about AOF’s work however you want, but the analysis is improper and it’s surprising that it is coming from a professional statistician.

[Complete_newbie]

Donnie I didn't read his article as if he was making a predictive model; rather just an experiment with base assumption of equal random events being stock returns. When I read it he quoted some economist saying "drawing paper out of a hat" that's random selection.

If that was his intention then it's bad statistics ofcourse.

Who said normal distribution is cutting edge ? It's actually laughable to me that we are approximating stick distributions that way...but I bet if you do MC using gaussian curve or "insert your proprietary model" your SWR is probably not that dofferent .

I don't think he's a statistician ... Seems like an actuary

[Complete_newbie]

LIFO wrote:

 

Donnie wrote:

AOF isn’t using a normal or lognormal distribution.  That would be simplistic, but definsible.  He’s using a distribution consisting of an equal weighting of all historical returns.  It’s a worthless exercise.  No takeaways possible.

Click to expand…

Can you explain this thought a little more?  Do we know if annual stock market returns are distributed in a normal or lognormal fashion?  If this is all the data that we have, the distribution is the distribution. Are you concerned that he is sampling without replacement?

I’m not a statistics whiz but the real limitation to better analysis seems to be limited data.  Is it possible to do a better assessment of stock market returns with the current data we have?  More than a few of us here have experience with robust scientific methods and and statistical analysis.  Can we show them how it should be done?

Click to expand...

I don't think that is the issue that much. I think just using historical returns doesn't give us many other scenarios that monte carlo can since it is not only based on data we have but it generates random returns/scenarios based on the model you created. Historical return resampling/bootstrapping doesn't have any such model. So yea its more advantageous to use MC. BUT Monte Carlo simulations are only as good as the input assumptions. Do financial planners really input even mean reversion in it? Doubtful.