afan wrote:Some of the discussion in this thread has conflated the behavior of a given factor with the effectiveness of a particular mix of factors. Two different issues. The latter always raises multiple testing concerns. The former might or might not depending on how the factor has been tested and supported in the past.
Two different issues, but related and both of interest. Also, as mentioned before, the 'mix' being touted and tested is not the optimal one over the data set (if it were, this would be a clear sign of data mining and would strongly require more stringent tests for significance of outperformance over other combinations). That doesn't mean multiple different combinations weren't at least considered, though, but if you treat all four styles individually as significant or close to it then it is hard to imagine a scenario where you'd be skeptical of any combination of them, especially with the correlations seen.
afan wrote:Some of the discussion has also ignored concerns about the definition of performance. If the goal is higher expected risk adjusted returns, then one must define risk. Some of the discussion has assumed risk is captured by variance alone. For other comments it is not clear what is meant by risk. I think it is clear that real world investors care about at least skewness and perhaps higher moments. Analysis of the effect on an overall portfolio should take this into consideration.
Standard deviation may be close enough. But there's some minor negative skewness in the dataset, and I would not at all be surprised by more in real life. With around a 10% standard deviation target, because of the unknowns I think a close enough mental model is somewhere between 60/40 stocks/bonds and 100% stocks kind of risk, probably on the latter side.
afan wrote:The concern about running a fund with a mix of strategies, a mix of factor bets, is that there may be covariances not accounted for when the factors are treated as independent. Even if one models some level of interaction, the risk remains. That was what happened to LTCM. They thought they had modeled the covariances, only to discover that in an extreme event, their bets were much more highly correlated than they imagined. Given the Nobel - level IQ's involved, these events say that getting the matrix right is hard. Really hard. Probably too hard for yhe smartest people in the world. Unless the collective knowledge is much greater now than then, a difficult argument to make, it implies that one should be skeptical of claims to have solved this problem.
Then it's clear you haven't skimmed any of the relevant papers, because they're claiming something worse than independence: negative correlation. They show about -0.60 between value and momentum, with a t-stat of -13.14 on the monthly data series. Spuriously significant? You tell me.