Pandas Efficient VWAP Calculation

Quick Edit: Just wanted to thank John for the original post :)

You can get even faster results by @jit-ing numpy's version:

@jit
def np_vwap():
    return np.cumsum(v*(h+l)/2) / np.cumsum(v)

This got me 50.9 µs per loop as opposed to 74.5 µs per loop using the vwap version above.

Getting into one pass vs one line starts to get a little semantical. How about this for a distinction: you can do it with 1 line of pandas, 1 line of numpy, or several lines of numba.

from numba import jit

df=pd.DataFrame( np.random.randn(10000,3), columns=['v','h','l'] )

df['vwap_pandas'] = (df.v*(df.h+df.l)/2).cumsum() / df.v.cumsum()

@jit
def vwap():
    tmp1 = np.zeros_like(v)
    tmp2 = np.zeros_like(v)
    for i in range(0,len(v)):
        tmp1[i] = tmp1[i-1] + v[i] * ( h[i] + l[i] ) / 2.
        tmp2[i] = tmp2[i-1] + v[i]
    return tmp1 / tmp2

v = df.v.values
h = df.h.values
l = df.l.values

df['vwap_numpy'] = np.cumsum(v*(h+l)/2) / np.cumsum(v)

df['vwap_numba'] = vwap()

Timings:

%timeit (df.v*(df.h+df.l)/2).cumsum() / df.v.cumsum()  # pandas
1000 loops, best of 3: 829 µs per loop

%timeit np.cumsum(v*(h+l)/2) / np.cumsum(v)            # numpy
10000 loops, best of 3: 165 µs per loop

%timeit vwap()                                         # numba
10000 loops, best of 3: 87.4 µs per loop