03-06-2017 09:53 AM
I have 50000 time series with seasonal trends and missing values. I would like to use seasonality to impute the missing values.
Please let me know if SAS has a proceedure for this.
The data is output from PROC TIMESERIES (interval = week).
The data covers 15 or 16 years (depends on which time series).
The seasonal patterns are all one year long (a period of one year, and correspond to spring, summer, autumn and winter).
Some of the gaps cover an entire season for a year, but other years have the corresponding season data (summer 2015 is missing; but summer 2014 and summer 2016 are not missing).
Because of the missing seasons, I am unsure if PROC EXPAND's methods will give reasonable results.
Would PROC MI's MCMC give solid results?
Initially, I used PROC HPSUMMARY to find monthly medians and imputed. Unfortunately, a collaborator objects to the four-week stair patterns.
03-06-2017 10:56 AM
One approach you can consider is to use a forecasting model to impute the historical missing values. For example, you can use PROC ESM with method = WINTERS to forecast the series. You will get not only the furture forecasts but also in-sample fit for the entire history including the missing value periods. You can use the in-sample fit to imput the historical missing values.