pandas - 循环乘以 Pandas 中的先前值

Question

我在 pandas 中有一个 DataFrame，如下所示:

df = pd.DataFrame({'origin_dte':['2009-08-01','2009-08-01','2009-08-01','2009-08-01','2009-09-01','2009-09-01','2009-09-01'],
                   'date':['2009-08-01','2009-08-02','2009-08-03','2009-08-04','2009-09-01','2009-09-02','2009-09-03'],
                   'bal_pred':[10.,11.,12.,13.,21.,22.,23.],
                   'dbal_pred':[np.nan,.25,.3,.5,np.nan,.4,.45]})

    bal_pred   date   dbal_pred origin_dte
0   10      2009-08-01  NaN     2009-08-01
1   11      2009-08-02  0.25    2009-08-01
2   12      2009-08-03  0.30    2009-08-01
3   13      2009-08-04  0.50    2009-08-01
4   21      2009-09-01  NaN     2009-09-01
5   22      2009-09-02  0.40    2009-09-01
6   23      2009-09-03  0.45    2009-09-01

我想遍历并替换 bal_pred 的每个观察值，其中 dbal_pred != NaN 为 dbal_pred[i] * bal_pred[i-1]。例如，bal_pred 的第二个值将变为 0.25*10=2.5。当 origin_dte 发生变化时，意味着 dbal_pred 再次为 NaN，计算将跳过 NaN 观察并计算下一个 bal_pred。所以 df 看起来如下所示。我有一个 while 循环来执行此操作，但问题是循环遍历大型数据帧需要很长时间。非常感谢更简单/更快的方法来做到这一点!

    bal_pred  date       dbal_pred  origin_dte
0   10.000    2009-08-01    NaN     2009-08-01
1   2.500     2009-08-02    0.25    2009-08-01
2   0.750     2009-08-03    0.30    2009-08-01
3   0.375     2009-08-04    0.50    2009-08-01
4   21.000    2009-09-01    NaN     2009-09-01
5   8.400     2009-09-02    0.40    2009-09-01
6   3.780     2009-09-03    0.45    2009-09-01

Answer 1

另一种方法是标记每组数据，然后取每组的累积乘积

group = df['dbal_pred'].isnull().cumsum() 
df.dbal_pred.fillna(df.bal_pred, inplace=True)
df['bal_pred'] = df.groupby(group)['dbal_pred'].cumprod()

输出

   bal_pred        date  dbal_pred  origin_dte
0    10.000  2009-08-01        NaN  2009-08-01
1     2.500  2009-08-02       0.25  2009-08-01
2     0.750  2009-08-03       0.30  2009-08-01
3     0.375  2009-08-04       0.50  2009-08-01
4    21.000  2009-09-01        NaN  2009-09-01
5     8.400  2009-09-02       0.40  2009-09-01
6     3.780  2009-09-03       0.45  2009-09-01