04-05-2017 05:30 AM
As we know,we usually use time series model to forecast products who have historical time series data,but if historical data is short,for example new product,it usually less than 6 data,we can't get a better result through time series model,such as arima,esm,and so on. I usually use Gray model or just according to new product's attributes to find some similar product who has long time serie data to forecast new product,but some new products sometimes have no similar product.Does anybody has some new ideas for new product forecasting?
Look forward for your advise.
04-05-2017 09:18 AM
04-05-2017 12:05 PM
Forecasting new products is a complex and multifaceted problem. It has been subject to many papers and entire books. The general consensus is that no single model, or model class is sufficient to cover all cases. A process must be in place that includes a portfolio of tools, techniques and methods to address the various scenarios and stages of the introduction of new products.
Here are some pointers to literature. Hopefully they will be of some help:
This book gives an overview of new product forecasting as a process.
This paper is a recent survery of methods and models
And this paper describes the implementation of a structrured process in a SAS product that includes a component for forecasting new products (this is, of course, the most important reference since I contributed to it .
04-05-2017 05:03 PM
you can consider using regression models (linear regression, Tree, neural net etc.) across a set of similar products including the new product itself to pool information across multiple time series. You should also include the product attributes in the regression model to capture attribute specific demand features. Noticed that this regression model can be used to generate forecasts for products with different product attribute combinatiions. Using the Tree models will also give you the capability to generate forecasts with unseen product attribute values. Once the products become mature, you can use the time series models to generate the forecasts.
Hope this helps