2

I have daily data for 2 years starting from jan1 2014 till december 31 2015. I want to forecast for next 365 days using this data set. I am using below code.

PROC UCM data=Mydata;                                                                                                                                                                                                                                      
id date interval=day; /*set the interval as day*/                                                                                                                                                                                                                                          
model Sold;
irregular;
level;                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             
slope ;                                                                                                                                                                                                                                                         
cycle period=365.24 rho=1 variance=0 noest=(rho period variance);                                                                                                                                                                                               
estimate;                                                                                                                                                                                                                                                       
forecast lead=365 OUTFOR=Ucmm.DATA;                                                                                                                                                                                                                             
run;

There is a trend based on day of the week and month of the year. How do I specify that? I understand for day of week it can be set as 7. For month how can I specify that, as some months have 31 days while some others have 30 and yet february has 28/29 days. How do I solve this? I know it can be specified using season; command. But how?

Also, how to include impact of certain occasions like Christmas, new year, Thanksgiving etc? I believe model; command takes care of that as per this paper. How many of these dummy variables can we include? Is there any limit?

Richard Hardy
  • 54,375
  • 10
  • 95
  • 219
StatguyUser
  • 874
  • 3
  • 9
  • 27
  • You might also ask about Holiday effects , changes in daily patterns over time or the effect of certain days in the month which is quite common with daily data . – IrishStat Jan 03 '16 at 12:23
  • Thanks for pointing out, i believe command `model;` takes care of that. I will add it to validate though. – StatguyUser Jan 03 '16 at 12:25
  • You might share your data, your results , including the final error process often laden/bloated with unidentified structure due to deficient exploratory analysis and ask the list for some comparative/more comprehensive existing approaches. – IrishStat Jan 03 '16 at 12:34
  • Can you elaborate on exploratory analysis part please? For daily data, i plotted pattern by day of week, week of the year, day of month, month of the year. What more should i do in this phase? am i missing anything that i should be doing? – StatguyUser Jan 03 '16 at 12:48
  • For example the daily pattern for a Monday might be +20 based upon 700 observations ( really 100 samples for Monday). Now the Monday (statistically significant) pattern might be +10 for the first 50 and +30 for the last 50 , averaging 20 overall Going forward it might be silly to use +20 rather than the more correct +30 as the offset for Monday. You are trying to do visual analysis by plotting the data rather than statistical analysis to gleen the underlying significant patterns . It ight be time to move on and to rest your weary eyes . – IrishStat Jan 03 '16 at 14:51

0 Answers0