@ Miles -
Did you come up with this yourself?
Sort of. Over time, I tend to grab bits and pieces from the Quantopian form and cobble them together. I think the ETFs came from someone else's post; I did not research them myself. I cooked up the idea of minimizing the variance with constraints (although I probably came across the general concept in a paper). The weighting by standard deviation is a concept mentioned on Quantopian and in a book I read.
I think this is just another algo that is highly overfit.
Possible. I don't claim to understand it. One thought, though, is that XIV, SPY, and TLT are all broad index-based ETFs, so that's a plus. You are suggesting that there is an underlying risk, so what is it? And how can it be managed/hedged? One advantage of Quantopian is that potentially bad algos can be exposed and discarded, or fixed, so that they are no longer bad.
@ Vladimir -
I was always wondering: why do you using minute data for weekly trading?
For short look-back periods (e.g. 5 days), there is very little data. Also, closing prices come from a single trade at the end of the day; intuitively, why not use more data in some fashion? There's a lot more flexibility in using minute bars, since smoothing can be applied. If there are too many data points, then the data can be decimated.