As promised in yesterday's Foresight-SAS sponsored webinar on "Forecast Value Added: A Reality Check on Forecasting Practices," here is Part 1 of my written response to the over 25 questions that were submitted during the event. (Note: It may take a week or so to get through all of them.)
For those who missed the live webinar, the 28-minute recording is available for on demand review.
Also, please be sure to download the article "Forecast Value Added: A Reality Check on Forecasting Practices" that appeared in the Spring 2013 issue of Foresight.
Editor Len Tashman and the staff also invite you to request a trial copy of Foresight.
So with the backup support of Little Richard the Forecasting Gerbil (official mascot of The BFD blog), let's answer some questions.
*** Forecast Value Added Q&A ***
Q: You mentioned two traditional naive models, the random walk and seasonal random walk. How do I decide which one to use?
It's fine to always use the random walk (aka "no change" model) to generate your naive forecast. However, if you have highly seasonal data, you could consider using the seasonal random walk.
For consistency of the FVA analysis, I would suggest you choose one naïve model and stick with it for everything you are forecasting.
Q: Can I do FVA in a service industry?
Yes! FVA analysis can be applied anywhere you are doing forecasting.
FVA evaluates the forecasting process and doesn't care what industry you are in, or what you are trying to forecast. A manufacturer forecasts unit sales, an airline forecasts calls to the reservation desk, an insurer forecasts claims. Each type of organization has a process in place for doing the forecasting, and the process probably involves both statistical modeling and human overrides.
FVA lets you assess the overall effectiveness of the forecasting process (relative to a naive model), and potentially improve the forecast by eliminating negative-value adding activities.
Q: How do you do FVA for new product forecasting? What naive model do you use since there is no history?
There is actually no difference in doing FVA for new products.
Of course, until you have launched the product and start having actual sales, the random walk will forecast zero, but that's ok. (For new products that fail in the market place, zero may be a pretty accurate forecast!) If you are using a time-series model based solely on a product's history, your statistical forecast will also be zero until the actuals start coming in.
However, I expect there to always be management's forecast for the new item -- since presumably they wouldn't have approved the new product for development and release without some expectation of how well it will sell.
Just note that you probably don't want to use a seasonal random walk as your naive model for new product forecasting, as it will continue to forecast zero until you have a year's worth of actuals.