For a time series sales forecasting task I want to create a feature that represents the average sales over the last 3 days. I have a problem when I want to predict the sales for days in the future, since these data points do not have sales data (NaN values). Pandas offers rolling_mean(), but that function results in a NaN ouput when any data point in the window is NaN.
My data:
Date Sales
02-01-2013 100.0
03-01-2013 200.0
04-01-2013 300.0
05-01-2013 200.0
06-01-2013 NaN
Result after using pd.rolling_mean() with window of 2:
Date Rolling_Sales
02-01-2013 NaN
03-01-2013 150.0
04-01-2013 250.0
05-01-2013 250.0
06-01-2013 NaN
Desired result:
Date Rolling_Sales
02-01-2013 NaN
03-01-2013 150.0
04-01-2013 250.0
05-01-2013 250.0
06-01-2013 200.0
So in case the a NaN is included, I want to ignore it and take the average of all the other data points in the window.
Here is on way adding min_periods
s=df.Sales.rolling(window=2,min_periods=1).mean()
s.iloc[0]=np.nan
s
Out[1293]:
0 NaN
1 150.0
2 250.0
3 250.0
4 200.0
Name: Sales, dtype: float64