Fund style and policy combined with market conditions and characteristics is always a significant influence. Also, where did the higher performance grouping last in the rankings overall during the subsequent period. Returns quite good and may be "all star" caliber. Relevance and context is important. Mean reversion can be demonstrated virtually at will. As my colleague always says when queried regarding mean reversion; what do you want it show? Mean reversion is very malleable.

http://www.valuewalk.com/2015/05/hedge-fund-mean-reversion-2/ ;-)

Larry,

You raise good points about cherry picking data. The best defense against this is transparency and maximum data coverage with minimal intervention.

With this in mind, the piece analyzes all hedge fund long portfolios filed in form 13F that had <120% annual portfolio turnover. So there is no cherry picking of market conditions and little filtering of fund characteristics.

Alessandro,

Thank you for beating me to posting part 2.

Nothing! My pleasure! ;-)

Greg, seeking a little clarity. Selecting long only hedge funds with less than 120% turnover constitutes a qualifier rather than a random (no such thing in statistics) or full class application. Why select hedge funds? Why not run the study to cover the changes described and behavior patterns leading into and out of the change? If the sample is restricted to long only hedge fund holdings found retrospectively, is the sample fair, reasonable enough to constitute a statistically sound pool? Finally, how does the process avoid falling prey to the survivor identification syndrome? Not trying to be difficult but never saw anything with predictive capability during a few decades in the business. Assessing probability is hard enough.

Whether we’re discussing a sample or the universe is a matter of perspective: The dataset covers the entire universe of tractable hedge fund long portfolios. On the other hand, this universe is indeed ~1/2 of all hedge fund 13F-filers.

The above is mostly of philosophical interest. For the entire universe of 13F filers scraped from all historical SEC filings going back to 2000, the persistence of returns due to skill is even stronger. There is zero discretion in constructing of that dataset – we purchase it from a technology company that scrapes the filings without any knowledge of the intended end-use. Results for all mutual funds are also similar (though we don’t have the license to re-distribute survivor-free mutual fund findings).

Survivorship bias is indeed a significant concern (http://abwinsights.com/2015/03/26/hedge-fund-survivor-bias/, http://abwinsights.com/2015/03/31/large-hedge-fund-survivor-bias/). It is not an issue here since the universe includes defunct firms – they make up about 1/2 of the universe.

I regret that you have not seen predictive analysis of investment skill. I would venture a guess that you have not come across holdings-based analysis using a robust factor model? There is plenty of academic work on the persistence of residual returns (identified with proper use of a robust factor model), even under rudimentary analysis. Naturally, one can do much better with competent software engineering and larger datasets.

Why not simply cover the stocks subject to the changes and forget the hedge fund aspect? The fact that certain types of hedge funds own something seems irrelevant to the study. It is the stock's behavior that is the issue. As far as predictive analysis of investment skill, there is no such thing. Ask any of the most widely recognized skillful players over a long period and they will flatly state that they are not in the prediction business. They will also discourage people from assuming anything predictive regarding the nature of their work. Regarding survivorship I am referencing the statistical aspect which concerns the result not the entity. Survivor selection or identification is the greatest curse on back testing. Taleb discusses the subject in Fooled by Randomness and more can be found by accessing virtually any basic statistical text or lecturer. Skilled stock selection is a fine art. I am among the faithful believers in the value of sound selection methods of which they are many. Consistent application can produce superior results over time but all methods are far from predictive. I read the articles referenced and disagree with the premise. No matter how robust a system may be or how inclusive a dataset may be, no predictive model is possible. Probability is another matter but forecasting is largely a highly subjective, intuitive exercise. Forecasters are regularly wrong regarding events and the timing of events that perchance occur. Keynes made the only economic prediction that is proven fact. In the long run, we are all dead. I enthusiastically support any and all efforts to find methods to better understand the markets and support our activities. However, I never allow myself to believe that a predictive model exists. I have never seen one because they are none to see. I have seen many contenders and pretenders over the past 40 years but never saw one that paid the rent.

Not sure we are referring to the same thing using the term “survivorship bias” (https://en.wikipedia.org/wiki/Survivorship_bias). Analysis that includes all entities, surviving and defunct, is free of survivorship bias.

We seem to disagree on the semantics of what is “prediction” while agreeing on the existence of “skillful players” and "sound methods." As long as you believe that there are more entities/individuals/strategies/selection methods with persistently high performance and/or more of these with persistently low performance than one would get at random, it’s not terribly important how you choose to label this.

BTW, here’s a decent recent paper on predictive mutual fund skill measurement by Stanford and NBER: http://www.gsb.stanford.edu/sites/gsb/files/publication-pdf/RP3131.pdf. Note that it relies on returns-based style analysis so runs into statistical difficulties (http://abwinsights.com/2014/06/14/the-flaws-of-returns-based-style-analysis/, http://abwinsights.com/2015/01/04/returns-based-style-analysis-overfitting-collinearity/). Our analysis of portfolios is more intensive and robust, so explains wider dispersion in future returns.

It is very important how something is labeled. The term predictive should be used very carefully in all things, especially finance. When dealing with the capital markets there is no such thing as predictive. People offer predictions constantly. There is no basis for offering anything as a prediction but people do it all the time. It sounds more powerful than high probability forecast. Predictions pander to people that wish to believe that some analysts possess nearly clairvoyant powers. I firmly believe that learning from skillful players had extraordinary valuable. However, no predictions can be discerned from the information although reasonably probabilities can be assigned. The difference is huge. All the great mutual fund managers suffered through tough periods. Bill Miller did not become suddenly stupid or less skillful when his returns fell. Thanks for the literature references. Definitely widening my perspective. Also, I am impressed by your energy, inquisitiveness and enthusiasm. Keep it going.