Correlation and related pseudo-periodicity
|
In a paper by
Grinblatt & Moskowitz, it is noted that future returns seem to be
negatively correlated in the short term (1 month past returns) and long term
(3 - 5 year past returns), but positively
correlated in the medium term (3 - 12 months past returns).
This seems an intriguing observation, so I did
the following, using monthly S&P 500 returns, starting in Jan, 1960 (for example):
- I calculate the recent monthly returns by
averaging the monthly returns over the previous 12 months.
- I compare this recent return to the monthly return 1
month in the future.
- I move ahead one month to Feb, 1960 and repeat steps 1 & 2 (calculating the average of
12
previous months and the return 1 month in the future).
- After umpteen years of "moving ahead one month", I have two sequences of returns:
the 12-month
recent averages and the returns 1 month in the future.
- I calculate the
Pearson Correlation coefficient
for these two sets of returns.
- Then I repeat steps 1-5, comparing recent returns with the return
2 months in the future.
- Then 3 months in the future.
- etc. etc.
I get this interesting result:
>Why?
Yes. That's the question ...
>And what's the answer?
I have no idea. Yet. However, it has nothing to do with "the market" or investor
psychology or Greenspan or ...
>So what does it have to do with?
I don't know ... yet. But if we replace the monthly S&P 500 returns with 'invented' returns,
generated by a random number generator, we get Figure 2
Now, I'll have to repeat all this with, say, 6-month past returns,
or 24 months or ...
>And what does all this mean?
I suspect there's a mathematical explanation so I'll ...
>Mathematical? Uh ... can we change the subject?
Pay attention.
|
Fig. 1
Fig. 2
|
Fig. 3
|
>Okay, but the paper by Grinblatt & Moskowitz talks about future
returns compared the average return over the past 1 month, 12 months, 36 months ...
Right, so I did that, too.
Figure 3 shows the correlation between future returns (which I take as the average monthly
return over the NEXT 3 months) and past returns
(which I take as the average monthly return over the past M months).
>It still looks sort of periodic.
Let's call it pseudo-periodic, okay?
|
>What about your invented, random monthly returns? Does it ...?
In Figure 4, instead of S&P 500 returns, I go through the same ritual, but I use a set of
randomly generated monthly returns.
>So, what're you going to do with all this?
Let me think on it.
|
Fig. 4
|
See also Correlation: does it matter?
for Part II
|