Great stuff. I saw that original paper and I had skewness on my list of things to look at. There might still be something in adding it with a low weight - I'm a big believer in trading things that make sense and add diversification even if their SR is close to zero. The other thing is it might work better within asset classes - like a lot of relative value trading. So for example within ags, within energy, within metals. My gut feeling is that skew is more interesting in financial futures, but until someone (probably you or I) checks this we won't know for sure. Here you'd definitely want to go within asset classes. Or you'd be persistently long stocks / short bonds. We know that trade hasn't worked. Correct me if I'm wrong but are all your results using daily returns to measure skew, albeit with different lookbacks? I wonder if using weekly / monthly returns would change things. I'm worried that bad data/ microstructure does weird things to daily returns. Of course we'd end up with estimates that are more random because of fewer data points for a given lookback. For this reason it might be better to use a more robust statistical estimate of skewness; eg a simple one is the 5% point from the returns distribution divided by the median estimate of returns minus the 95% point divided by the median. Or it might make sense to use the more robust estimate on daily returns to remove the effect of outliers, although I am still worried that there is a weird effect which will bias their skew which even the robust estimate won't take out. GAT
Nice findings. This probably would not change the results, but is there an advantage to have the weight of each future be the sign of its skew instead of the relative ranking of its skew within the portfolio? As you mentioned, the futures are all uncorrelated so there should not be an advantage in having a balanced long/short portfolio. Unlike stock portfolios where there is a common market factor, there is no common "commodities" market factor to balance.
Very nice work. It looks like the risk premium for skew (based on aversion to catastrophic risk and love of lottery outcomes) is not worth putting in as an alpha factor. I was wondering how hard it would be to modify your test a bit to look at the predictive value of the recent realized skew vs long term average skew in an instrument. I wonder if this ratio would be able to detect "smart money" accumulation or distribution. When building a long large players methodically clear overhead supply but ocassionaly shake the tree leaving a negative skew in their wake. The opposite happens when distributing or building a short.
Some destroyed traders in 1987, 2001 and 2008 would like to hire you to remove the outliers for them.
Yes I also like the diversification plus of this strategy but losing money for more than 30 years … that’s a bit of an ask for me Interesting, I hadn’t really thought of this as an RV trade … the key issue here is for me to think more about which markets within which sectors Yep this is on my list Yes all skews are based on daily returns. As you say, changing to weekly or monthly is going to significantly lessen the amount of data points and I’m pretty confident in the cleanliness of my data. So for the time being I’m comfortable with daily returns unless you think that there are maybe some microstructure issues that I haven’t considered. I like this Very much appreciate the feedback GAT … you’ve got me thinking. I have a few other projects on right now, but will post back here when I get a chance to do some further work on this.
So you mean, using the above testing portfolio and just simply go long all markets with a negative skew and vv? Sure, I guess you could do that because as you say there is little correlation amongst these markets but I think it’s unlikely to improve returns. In the past I’ve found that if you have identified a genuine return driver then you will improve risk-adjusted returns by weighting market allocations according to the strength of the factor that you’re measuring and if I were to trade skew that’s certainly what I’d be doing rather than a simple “long n top-ranked markets”, “short n bottom-ranked markets” approach.
That wouldn’t be difficult to test and is certainly worth considering. In fact I currently do something similar with a volatility measure that I employ for some shorter term strategies. The only thing that concerns me is that I’d now be introducing another potentially optimizable parameter into the mix … but I think the idea has merit so will take a look, thanks.
Just a small info relevant to this thread - we have performed an out of sample analysis of skewness effect in commodities (built on a research paper written by Fernandez-Perez, Frijns, Fuertes and Miffre - The Skewness of Commodity Futures Returns) and it still works well. You can read more about it in Quantpedia's research article, or on Alpha Architect's blog where it has been published ...