I am desperatly searching for a solution with pandas. Maybe you could help me.

I am looking for a rolling mean with consideration of the previous mean.

df looks like this:

now, using the rolling(window=2).mean() function I would get something like this:

I would like to consider the mean from the first calculation, like this:

where,

row1: (4 6)/2=5

row2: (5 10)/2=7.5

row3: (7.5 12)/2=9.75

thank you in advance!

CodePudding user response：

We can use simple python loop for this , if you would like speed it up you can try numba

l= []
n = 2
for x,y in zip(df['count'],df.index):
    try :
        l.append(np.nansum(x l[y-n 1])/n)
    except:
        l.append(x)
df.loc[n-1:, 'new']=l[n-1:]
df
Out[332]: 
   index  count   new
0      0      4   NaN
1      1      6  5.00
2      2     10  7.50
3      3     12  9.75

CodePudding user response：

EDIT: There is actually the method ewm implemented in pandas that can do this calculation

df['res'] = df['count'].ewm(alpha=0.5, adjust=False, min_periods=2).mean()

Original answer: Here is a way. as everything can be develop with coefficient being power of 2.

# first create a series with power of 2
coef = pd.Series(2**np.arange(len(df)), df.index).clip(lower=2)

df['res'] = (coef.div(2)*df['count']).cumsum()/coef

print(df)
   index  count   res
0      0      4  2.00
1      1      6  5.00
2      2     10  7.50
3      3     12  9.75

You can mask the first value with df.loc[0, 'res'] = np.nan if needed