Bogle's 10-Year Forecast

[hnzw_rui]

Alternatives are worse. That's why Bogle said one needs to stay invested.

Who said negative to zero return? Bogle said 6% nominal, unless we have P/E reversion then it would be less.

Was referring to the linked article from the previous post:
Hussman Funds - Stocks For The (Really, Really) Long Run

Granted, it's quite likely the article was written just to peddle their (pretty lacklustre) actively managed funds.

 

[Aeowyn]

I will note that he used rolling windows on a monthly basis. I think monthly is poor choice as this greatly inflates the number of data points on the graph and makes his model seem much stronger than can be justified statistically (because of the heavy amount of overlap, the data points on the graph aren't anywhere near independent, etc. etc.).

I agree. I assumed there was a data point for monthly 10 year rolling periods as they were way too many data points for yearly 10 year rolling periods over the time frame the graph was supposed to be for. I'd like to see these graphs done annually (by calendar year) for 10 year rolling periods.

If you look at Bogle's graph for a bigger dataset (1915-2014) you can see that this relationship no longer holds and a Bogle prediction of 6% yields an actual return centered around 6%. The problem is that 1990-2014 is such a short time period that you can get a lot of noise and spurious results.

Why didn't he use this graph? Even though there is more scatter, I think it shows his prediction method is on average in line with returns. Of course Bogle couldn't have been doing predictions way back to 1915. Was this the data he used to create his prediction formula and back test it?

I should have specified in my prior post - I don't really predict a return of 11% over the next 10 years, but I think it is just as good a prediction as Bogles - meaning I don't put much stock in either prediction.

But over the truly long term (say 40-50 years) I think 11% is more likely than 6% - but I probably won't be around to see if I'm right.

 

[CaliforniaMan]

I found the paper where bogle actually describes his methodology used to generate graph posted above: An Error Occurred Setting Your User Cookie It's behind a paywall but available free with a trial registration.

To answer my earlier questions in this thread about how he generated his numbers, he uses the initial dividend yield and trailing 10 year average for earnings growth. His also assumes that P/E will revert to the prior 30-year average.

I will note that he used rolling windows on a monthly basis. I think monthly is poor choice as this greatly inflates the number of data points on the graph and makes his model seem much stronger than can be justified statistically (because of the heavy amount of overlap, the data points on the graph aren't anywhere near independent, etc. etc.).

In his paper, he also redoes his experiment from 1915-2015. Here he obtains an R^2 = 0.44 which is much lower but inline with experimental results for Schiller PE10. I can buy that Bogle's model is roughly as good as Schiller PE but I had hard time accepting an R^2 of 0.65 (which I think is too good to be true).



If you look at Bogle's graph for a bigger dataset (1915-2014) you can see that this relationship no longer holds and a Bogle prediction of 6% yields an actual return centered around 6%. The problem is that 1990-2014 is such a short time period that you can get a lot of noise and spurious results.

Very good analysis Photoguy thanks! You very well answered a question I had posted previously on this thread. A lot of people just don't realize that the observations used in a correlation MUST be independent for the statistical tables for that correlation coefficient to be valid. As the data becomes smoother (as with a time series, or smoothing) the tails of the probability distribution rise so that, depending on the amount of autocorrelation in the series, the tails of the distribution (around +/- one) rise while the middle (around zero) fall. And hence with autocorrelated series you very often get more values near 1 than near zero, even when there is no relationship.

It is a fun exercise to smooth random series and watch as you smooth them as the correlations tend toward 1.0. I think correlations between autocorrelated variables are often used (not that I am saying they are used that way here) to fool people.

Again, thank you for your thoughtful post.