我认为你可以做到rolling_apply
在正常调用的函数内groupby/apply
。所以像下面这样:
def roll_wsum(g,w,p):
rsum = pd.rolling_apply(g.values,p,lambda x: np.dot(w,x),min_periods=p)
return pd.Series(rsum,index=g.index)
weights = np.array([0.1,0.1,0.2,0.6])
df['wsum'] = df.groupby('ID')['VALUE'].apply(roll_wsum,weights,4)
print df
Output:
DATE ID VALUE wsum
0 2012-12-31 A 100 NaN
1 2013-03-31 A 120 NaN
2 2013-06-30 A 140 NaN
3 2013-09-30 A 160 146.0
4 2013-12-31 A 180 166.0
5 2013-03-31 B 0 NaN
6 2013-06-30 B 5 NaN
7 2013-09-30 B 1 NaN
8 2013-12-31 B 3 2.5
9 2012-12-31 C 45 NaN
10 2013-03-31 C 46 NaN
11 2013-06-30 C 42 NaN
12 2013-09-30 C 30 35.5
13 2013-12-31 C 11 21.4
14 2012-12-31 D 18 NaN
15 2013-03-31 D 9 NaN
16 2013-06-30 D 13 NaN
17 2013-09-30 D 5 8.3
18 2013-12-31 D 11 9.8
19 2012-12-31 E 0 NaN
因此,我只是按“ID”对数据进行分组,然后将一组的“VALUE”列发送到我的 roll_wsum 函数(以及加权和和周期的权重)。这roll_wsum
函数调用rolling_apply
并将一个简单的 lambda 函数提供给rolling_apply
:“VALUE”和权重的点积。此外,这里强制实施min_periods=4
条件,因为我们需要数组的长度(权重和 df['VALUE'].values)相同。
鉴于我使用点积来计算加权和,它可能无法按照您想要的方式处理缺失值。因此,例如,您可能更喜欢以下内容(尽管它对于示例数据没有影响):
def roll_wsum(g,w,p):
rsum = pd.rolling_apply(g.values,p,lambda x: np.nansum(w*x),min_periods=p)
return pd.Series(rsum,index=g.index)
weights = np.array([0.1,0.1,0.2,0.6])
df['wsum'] = df.groupby('ID')['VALUE'].apply(roll_wsum,weights,4)