pandas.rolling_std() gives nans where it probably shouldn't

[erg]

The output differs from bottleneck, which also gives nans, but fewer.

The Factor version never outputs nans and otherwise matches the pandas version.

https://gist.github.com/3624548

[erg]

Same bug report on bottleneck:
pydata/bottleneck#50

[erg]

Bottleneck is fixing their version of this bug...;)

[erg]

The bottleneck bug was taking the square root of a negative number. Probably the problem here.

[wesm]

thanks, fixed