I have a data set that includes sales dollars by sales order and I want to perform a time series forecast on it. Low dollar sales orders have very little noise and after detrending and doing some prelim analysis and testing it looks like a forecast would work pretty well.The issue is there is a lot of variability in large orders that can skew the time series. I can remove any large outliers, but I wanted to know if there was a way I could optimize the threshold that determines large v. small orders so I can get a reasonable time series that doesn't have too much variance. Any ideas?
In this case, you will want to not adjust for outliers. If there is no way to forecast by customer then you will have to live with this the way it is.
Typically, forecasting dollars is not the best way to do things. Forecast units and then dollarize the forecast based on the price afterwards.