Is modelling a structural change in a time series useful for statistical forecasting? A colleague of mine is arguing that we should look into using some structural time series tools for improving our demand forecasting (I work in retail). I am a little bit skeptical. 
I don't see any benefit to being able to model structural changes in our use case (relatively short series - 1 ~ 2 years of weekly sales data). 
But then it struck me: Does modeling structural changes help at all with forecasting? 
If there are structural changes to the series, then we are outside of statistical forecasting territory and in need of a domain expert to explain what happened. 
There's no point in trying to model the structural change as it won't help with any algorithmic improvement in forecasting accuracy. 
At most we can detect the structural change with the purpose of discarding the data prior to the change since it is no longer relevant - but other than that I don't see any purpose in trying to model structural changes. 
Is this line of reasoning correct? 
 A: The degree to which a structural change renders all data before the change useless for prediction depends on the nature of the structural change.  In my experience, most structural changes are not so disruptive that previous data is completely uninformative.  (Note: I also work in retail, developing sales forecasting tools.)  
For example, forecasting regional demand for an item at a distribution center; if the item is stocked in 20 stores, then is rolled out to another 10 stores, the mean demand will jump over the course of a couple of weeks, but the general trend and seasonal patterns of sales are likely to remain unchanged (only rescaled for the new, higher, demand level.)  Thus, past sales are still informative - but if you don't take the structural shift into account, your estimates of trend and level will be seriously messed up for some considerable time after the shift.
Another, more complex, example would be the rollout of a product that competes with an existing, highly seasonal, product, e.g., selling one type of spareribs (highly seasonal Memorial Day, 4th of July, and Labor Day sales) then adding another type of sparerib to the product mix.  Naturally the mean demand will shift, and possibly the seasonal effect as well, but short-term autocorrelations in sales and the long term non-holiday seasonal patterns (e.g., winter vs. summer) are likely to remain unchanged (albeit rescaled to the new, lower, demand level.)
