Fama-French multivariate regression

[eigenfoo]

Fixes #379

We don't just 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 frustratingly this solution doesn't work since apply only works on Series data (see citynorman's comment on the top answer). So, we will have to write our own rolling multivariate regression...

[eigenfoo]

@twiecki @gusgordon the Python 3.4 build seems very sad without the statsmodels module

  File "/home/travis/build/quantopian/pyfolio/pyfolio/timeseries.py", line 24, in <module>
    import statsmodels.formula.api as sm
ImportError: No module named 'statsmodels'

[eigenfoo]

I also want to make sure that this computation is correct. Are these parameters the Fama French betas?

The -1 in the formula keyword means to set the intercept equal to 0.

[twiecki]

we should not force the intercept to be 0.

[eigenfoo]

Fixed!

[twiecki]

Yes, we need to add statsmodels to the requirements file.

[eigenfoo]

Added statsmodels to requirements file and fixed bug forcing regression intercept to be 0. The Travis builds seem to be passing, although the Python 3.4 build timed out: I'll try and implement this solution tomorrow morning.

Other than that, the PR is ready for review and merge! @gusgordon @twiecki

[eigenfoo]

@twiecki the Travis builds are not failing, but are timing out. See here for an example. I've tried implementing a solution using travis_wait, but that does not seem to help: builds still time out. @gusgordon and I don't really know what's going on. Any thoughts?

[twiecki]

@georgh0021 how slow is the regression when you try it locally?

[eigenfoo]

@twiecki

In [13]: %timeit rolling_fama_french(returns, factor_returns)
         1 loop, best of 3: 4.52 s per loop

If performance is an issue, it may be worth looking into Pythonic's answer to this forum post. He implements a numpy-only solution with linear algebra...

In [14]: %timeit rolling_regression(factor_returns.iloc[:, 0], returns)
         10 loops, best of 3: 61.8 ms per loop

[twiecki]

4.52 secs is quite slow for this simple functionality. How long a time-range did you check this on? The numpy version is not ideal but perhaps our best shot if we want to keep the functionality.

[twiecki]

Why do you test on only one factor? factor_returns.iloc[:, 0],

[eigenfoo]

Laziness 😛 the code on StackOverflow only took in one variable, I just copied and pasted to see what it would look like. Will refactor and test today.

[twiecki]

Not sure I trust that code. Have you tried sklearn.linear_model.LinearRegression?

[eigenfoo]

@twiecki sklearn runs much faster!
1 loop, best of 3: 479 ms per loop
And you were right, the regression matches the regression done by statsmodels, while the numpy solution was weird and incorrect.

Travis can't seem to find the sklearn package though. Any ideas?

PackageNotFoundError: Packages missing in current channels:
            
  - sklearn
We have searched for the packages in the following channels:
            
  - https://repo.continuum.io/pkgs/free/linux-64
  - https://repo.continuum.io/pkgs/free/noarch
  - https://repo.continuum.io/pkgs/r/linux-64
  - https://repo.continuum.io/pkgs/r/noarch
  - https://repo.continuum.io/pkgs/pro/linux-64
  - https://repo.continuum.io/pkgs/pro/noarch
            
The command "conda create -q -n testenv --yes python=$TRAVIS_PYTHON_VERSION ipython pyzmq numpy scipy nose matplotlib pandas Cython patsy flake8 seaborn sklearn runipy pytables networkx pandas-datareader matplotlib-tests joblib" failed and exited with 1 during .
Your build has been stopped.

[twiecki]

Great! it's conda install scikit-learn

[eigenfoo]

@twiecki @gusgordon back to square one: Python 3.4 build is still timing out. More help needed, unfortunately.

[eigenfoo]

@richafrank @twiecki bump. No idea why the tests are timing out.

[twiecki]

Hm, seems like 3.4 and 3.5 are pretty close to the limit. At this point we can probably drop 3.4 all-together.

[eigenfoo]

Not sure why the builds aren't timing out now... Something to keep in mind going forward I suppose. Once this becomes a serious problem we can look into finding a solution.

[twiecki]

Well, I think it's right on the razor's edge. We should probably just test over a shorter time-period.