I am working in a online supermakert , my current work is to predict daily sale count of fresh goods.
I tried to use time series model ARMA and xgboost, but both didn't fit well. The problem is every product may usually sold out/ stock out before a day end.
So the problem became restore the real daily sale count at first.
As you know, fresh things would easily decay, that's why we can't stock too much , which result in lack of normal data.
For example, I calculate the prawn stockout point(by hour) in last 147 days:
And average sale count(by hour) in last 147 days:
But I can't just use the
average sale count to restore daily sale count.
11.pm may only have 10 points and
10.am may have 140 points, very unbalanced.
Including the weather, holiday and total user count in that day and something else. There are too many thoughts in my brain, mess up.
Would you give me a proper adive/relatively correct way to restore the daily sale count?