ENH: Rolling Calculations Ignore Extended Window Specification

[daskol]

Is your feature request related to a problem?

Rolling windows API is not coherent to scipy.signal API.

Describe the solution you'd like

Support for weighted rolling window functions in pandas is based on scipy.signal. Precisely, pandas uses scipy.signal.get_window in order to get weights for a specific window specification. The issue is that scipy allows more accurate window specification. Namely, one can use non-symetric exponential window as following.

scipy.signal.windows.exponential(window_size, sym=False)

Moreover, pandas API specification requires paramter tau what is not necesaray in fact. So, API could be even simpler.

A possible solution to use win_type in order togetattrib of scipy.signal.windows module. If there is a function named as win_type then pandas should invoke the function with all passed paramters to aggregation function over rolling window. In the snippet below, .sum() implies invocation of scipy.signal.windows.exponential(2, sym=False).

df = pd.DataFrame({'B': [0, 1, 2, np.nan, 4]})
df.rolling(2, win_type='exponential').sum(sym=True)

API breaking implications

API won't be broken. However, I would like to relax some contraints. Namely, aggregations over rolling window should be calculated without explicit window parameter specification.

Describe alternatives you've considered

Alternative solution looks like the following. It is simple and a bit verbose.

window = sp.signal.windows.exponential(2, sym=False)

df.rolling(2, min_periods=2) \
    .apply(lambda x: (x * window).sum())