0

I have a dataset that contains data like so:

Customer, Quarter, TotalSales
Customer1, Q1, 2000
Customer2, Q1, 1232
Customer1, Q2, 432423
Customer2, Q2, 2222
Customer1, Q3, 242343
...

We would like to interpolate this into a more fine grained view, based on individual dates. Is it possible somehow to smooth the sales data out for the 3 months in each quarter, so that the total sale for that quarter is still the same, but still align the values to match the data of the days of the Quarters before and after?

Customer, Quarter, Daily Sales
Customer1, 2020-01-01, interpolated value
Customer1, 2020-01-02, interpolated value
Customer1, 2020-01-03, interpolated value
....
Customer1, 2020-01-31, interpolated value

                        ^ sum of these still add up to  Customer1, Q1, 2000
...
Roger Johansson
  • 22,764
  • 18
  • 97
  • 193
  • 1
    Don't you have access to monthly data? If not, which kind of behaviour do you expect for your "interpolated value"? Linear? Quadratic? Exponential?... There are quite a lot of possibilities. – Let's try Aug 26 '20 at 08:20
  • 1
    When You resample quarter to days You want TotalSales to be divided to aprox. 90 days? – ipj Aug 26 '20 at 08:22
  • "Don't you have access to monthly data? " obviously no. I don't think you can use any of those variations listed and still get the proper total of the entire period. if it declines in one part, it needs to increase somewhere else to keep the right total. – Roger Johansson Aug 26 '20 at 10:05
  • This method can help, especialy if You have access to additional daily data correlated to daily sales: https://stackoverflow.com/questions/60058095/temporal-disaggregation-of-time-series-in-python – ipj Aug 26 '20 at 10:25

1 Answers1

1

There is a class of methods known as "temporal disaggregation" methods. Look at this package on GitHub. Basically, you can specify a lower frequency time series and convert it (extrapolate) a higher frequency one. I've only used this package in R but it seems that the python implementation should be pretty much the same. You can specify whether the higher frequency observation should sum up to the lower frequency ones (or some other value).

Georgi Ganchev
  • 116
  • 5
  • Interesting method indeed. To really gain from it the another related daily series is needed that follows a similar movement to TotalSales. In connection to question for example it might be daily export data. – ipj Aug 26 '20 at 08:35