I have a dataset regarding weekly (175 weeks data) demand. "week" is an atrribute ranging from 1 to 175 . do I need to normalise this column ? what is the significance of the values with output label ( "demand" column )
as there are many other attributes , model will give more importance where "week" has highr values ( more importance to week= 140 and less importance to week= 10). but this should not be done. am I right ?
I think it depend to phenomenon you are studding it.
You have t=1, 2,..., 175
You have s(t) for each t above.
What do you want to do?
Maybe you find seasonality, as in farming where seasons matter and return year after year.
A shift by one year, is a shift by 52 weeks, 2 years is 104, 3 years is 156, and you have slightly less than 4 years of data...
If your data is sales, maybe a recurring event, like Christmas in many countries or Chinese New Year in China, etc, will show a hint of quasi periodicity.
It may be interesting to find a "long term predictor", which is this shift T (in the examples of farming or sales, T=52 is a good bet...)
To find the most likely T, you compute an estimate of E((s(t+T)-a*s(t))**2) and find a and T to minimise this expression
To do this you can solve Min sum (s(t+T)-s(t))**2 for t=1 to t=175-T.
You get a first model with parameters a,T, where s(t+T) = a*s(t) + e(t+T).
Then you model e(t), using autoregressive filters, and you write e(t) = (sum a(i)*e(t-i) for i =0 to n) +r(t)...
Does it help you?
- Badges
- Science topic