Sep

1

Any thoughts on the paper "The Two-Year Effect" by Graham Bornholt ?

This paper identifies a puzzling form of predictability in U.S. stock market portfolios. For the value weighted market index, those years that follow a low return two years earlier have an average return 11.6% higher than those years that follow a high return two years earlier. The difference in returns is economically and statistically significant.

Vic and Laurel reply:

Let us say it does not come as a complete surprise; the strong negative correlation between the current year return and the return two years back is mentioned in our book Practical Speculation on pages 210-211. See in particular Table 9.2 on Page 211,
9.2

We invite the submission of relevant analyses.

Kim Zussman follows up:

Using SP500 (w/o dividends) I checked Dec-Dec returns 1950-07 for years coming 2 years after down years (the year after the year after a down year). Comparing these with all yearly returns for the series showed them to be higher, but the difference is not statistically significant:

Phil McDonnell expresses skepticism:

One way to test for a two year negative correlation is to calculate a correlation coefficient with a lag of 24 months. For S&P adjusted [monthly] returns from 1950 to 2005 we get the following correlations at the various monthly lags from 18 months to 33 months:

Dr Phil autocorr table

Note that the correlation at 24 months is actually +1.55% which does not support the idea of a negative correlation at two years. However overall most of the monthly correlations are negative in the 18 month to 33 month range. Another indication of how weak any effect may be is that none of the correlations rise to the level of 5% significance. Given that there are 16 chances we might expect that at least one would be significant by chance alone.

Dr. McDonnell is the author of Optimal Portfolio Modeling, Wiley, 2008

Alex Castaldo replies:

I thought we were discussing yearly returns, not monthly. You lost me somewhere.


Comments

Name

Email

Website

Speak your mind

3 Comments so far

  1. George Parkanyi on September 1, 2008 7:29 pm

    Interesting, guys, but really, what would you do with that information even if it were statistically significant? The third year in the presidential cycle didn’t do a lot for us last year.

    A very successful business associate of mine always counseled me to follow the money (and we did very well by that together in an export business). Where is going? Why? Where will/did it come from? How long will it last?

    It might be interesting to measure aggregate money flows in and out of market sectors, commodities, and countries much like a remote sensing system does. Good proxies for countries are ADRs and country ETFs, while commodity ETFs now give you at least some current volume information. Long short ETFs also give useful information. It would be interesting to deploy “probes” all around the markets to measure the money flows of these various high-liquidity instruments, much like weather-forecasting systems collect met information from all over the world to monitor and forecast weather patterns. Sentiment indicators like VIX give you broad information, but your market “sonds” might give you more specific information about trends and trend changes.

    Cheers,
    George

  2. Barry Gitarts on September 1, 2008 9:33 pm

    I have seen some papers from the big money management firms that talk about effects like this but with the style box, i.e.: Small, Mid or Large cap — modified by growth, value or blend. I'm wondering if the style box is a good filter, rather then sector or something else which may create a class of stocks for such mean reversion.

  3. Greg Vinately on September 2, 2008 4:30 pm

    "Those years that follow a low return two years earlier have an average return 11.6% higher than those years that follow a high return two years earlier." The study does not claim to test an autocorrelation of monthly returns as performed by Dr. McDonnell. Autocorrelation studies are best geared towards volatility instead of price returns alone due to the effect of time-varying clustering effects. In fact, autocorrelation of squared returns is often a better measure when dealing with prices as commonly used in GARCH diagnostic tests. As per the abstract, the study also does not claim to test 2-yr. returns against returns of all years. It states that years that follow a low return two years earlier have an average return higher than years that follow a HIGH RETURN TWO YEARS EARLIER. On the surface the paper has more to do with mean reversion effects than price predictability alone.

Archives

Resources & Links

Search