带有时间偏移大熊猫的移动平均线

Question

I looking for a vectorized solution to calculating a moving average with a date offset. I have an irregularly spaced times series of costs for a product and for each value I would like to calculate the mean of the previous three values, with a date offset of 45 days. For example if this were my input dataframe:

    In [1]: df
    Out [1]:
        ActCost OrDate
   0    8       2015-01-01
   1    5       2015-02-04
   2    10      2015–02-11
   3    1       2015-02-11
   4    10      2015-03-11
   5    18      2015-03-15
   6    20      2015-05-18
   7    25      2015-05-23
   8    8       2015-06-11
   9    5       2015-10-09
  10    15      2015-11-02
  12    18      2015-12-20

The output would be:

    In[2]: df
    Out[2]:
        ActCost OrDate      EstCost
   0    8       2015-01-01  NaN
   1    5       2015-02-04  NaN
   2    10      2015–02-11  NaN
   3    1       2015-02-11  NaN
   4    10      2015-03-11  NaN
   5    18      2015-03-15  NaN
   6    20      2015-05-18  9.67  # mean(index 3:5)
   7    25      2015-05-23  9.67  # mean(index 3:5)
   8    8       2015-06-11  9.67  # mean(index 3:5) 
   9    5       2015-10-09  17.67 # mean(index 6:8)
  10    15      2015-11-02  17.67 # mean(index 6:8)
  12    18      2015-12-20  12.67 # mean(index 7:9)

My current solution is the following:

    for index, row in df.iterrows():
        orDate=row['OrDate']
        costsLanded = orDate - timedelta(45)
        if costsLanded <= np.min(df.OrDate):
            df.loc[index,'EstCost']=np.nan
            break
        if len(dfID[df.OrDate <= costsLanded]) < 3:
            df.loc[index,'EstCost'] = np.nan
            break
        df.loc[index,'EstCost']=np.mean(df[‘ActShipCost'][df.OrDate <=         
                                           costsLanded].head(3)) 

My code works, but is rather slow, and I have millions of these time series. I'm hoping that someone can give me some advice on how to speed this process up. I imagine that the best thing to do would be to vectorize the operation, but I'm not sure how to implement that. Thanks so much for the help!!

Answer 1

Try something like this:

#Set up DatetimeIndex (easier to just load in data with index as OrDate)
df = df.set_index('OrDate', drop=True)
df.index = pd.DatetimeIndex(df.index)
df.index.name = 'OrDate'

#Save original timestamps for later
idx = df.index

#Make timeseries with regular daily interval
df = df.resample('d').first()

#Take the moving mean with window size of 45 days
df = df.rolling(window=45, min_periods=0).mean()

#Grab the values for the original timestamp and put the index back
df = df.ix[idx].reset_index()

Answer 2

如果我理解正确，我认为你想要的就是

df.resample('45D').agg('mean')