Fama-French should use multivariate regression

[twiecki]

https://github.com/quantopian/pyfolio/blob/master/pyfolio/timeseries.py#L563 shows we are using separate linear regressions which is problematic because the factors are confounded.

We should instead use multivariate regression: http://stackoverflow.com/questions/19991445/run-an-ols-regression-with-pandas-data-frame

CC @jlowin

[eigenfoo]

This is a pretty serious bug. We need to get this fixed ASAP, especially given the work that is in the pipeline from #400 (i.e. the four-factor "post-Fama-French" alpha).

[jlowin]

Indeed -- this could dramatically affect the attribution numbers that @georgh0021 is working on

[eigenfoo]

It turns out that this is not as easy as I originally thought. We don't want a multivariate regression, we want a rolling multivariate regression. Pandas used to support this sort of thing with pd.stats.ols.MovingOLS, but that has unfortunately been deprecated.

A solution is described on StackOverflow, but annoyingly that also does not work. We will need to write our own rolling multivariate regression, using a for loop... I'll open a PR for this.