Confused about independence and prediction power of data What is the correct way (if there is one) to think about when authors claim that stocks have produced some percentage annual return X over every 20 year period of time? They might calculate this by using the monthly starting price of the S&P500 and compare it to the monthly ending price of the S&P500 20 years
later to the month. 
They might measure the return from Jan 1970 - Dec 1989 but they will also use the return from Feb 1970 - Jan 1990 as another point of data and so on.
My questions are threefold: 


*

*Since the time periods used for the analysis overlap each other except for the 1 month on each end as we move forward in time, is this really an accurate way to measure returns?  Isn't the next 20 year period of time's rate of return highly correlated to the previous rate of return?

*Does it make more sense to use non-overlapping 20 year periods to measure this? i.e. Jan 1970 - Dec 1989 being one data point and the next one being Jan 1990 - Dec 2009?

*Do these measurements have any predictive capability or do they just tell us what stock returns have been in the past?
 A: *

*Yes. Return is return, just difference between endpoint and startpoint. There is nothing wrong in doing something like this. These things are correlated, but you get what you paid for so to say. Every of known 20-year periods produced this return and this is just one of the measures, certainly it's not enough to know because yes, they are correlated if they overlap. 

*Not really. If there is a 10-year downward slope in the end of first 20-year period and also negative return in first half of second 20-year period, it can still give you two periods with positive return! The (1) way of doing it would detect this 20-year negative return piece while this technique wouldn't. It is way weaker than the first one.

*Realized performance is not an indicator of future performance. Nothing guarantees you predictive power. Otherwise everyone would use the same thing which works and become rich (which is self-contradictory and means there is no such thing anyway). This is up to you to decide if the strategy is robust enough, maybe see what methods they use, what guarantees they give and how to compare to other return tests. I'd also say that 20-year period looks way too long to me, although I'm no expert and maybe this is such a special kind of strategy.
