Find "fastest growing" products from a dataset of all product sales? Apologies for the simple/newbie question. I have a dataset of product sales by month across six years, which looks a bit like this:

I want to find the "fastest growing" sections by cost, for all products over the  baseline level of 1000 items sold in the past month. 
There are obviously lots of ways I could do this - I can think of some:


*

*For each product, look at the sales in the past quarter, vs same quarter 2 years ago, and find the products with the biggest percentage increase

*As above, but over a 5 year span

*For each product, look at the rate of change over the past quarter, and find the products with the fastest accelerating growth month on month


From a statistical point of view, is there a "right" or "wrong" way to do this? Or is there a good or bad way to do this? Or even an interesting or uninteresting way to do this?
My instinct is to pick the first of the bullet points above, but that would obviously miss any products that have grown a lot over the past three years, say. 
 A: The possible approaches to analyzing this data and answering your question are virtually limitless. The first thing you need to decide on is a relevant timeframe for the work. You've mentioned two: 2 and 5 year percent change. These will give you very different answers, regardless of methodology. This decision is not something anyone on this forum can advise you about as it's totally a function of the objectives of the analysis given to you. Secondary considerations include the technical sophistication of the audience for whom this analysis is being done as well as your own understanding of statistical analysis. 
One obvious approach is to put together a panel data model as in pooled time series regression. Since you have a fixed time and sample size, the classic approach outlined by Lee Cooper in his book Market Share Analysis (free download here ... http://www.anderson.ucla.edu/faculty/lee.cooper/MCI_Book/BOOKI2010.pdf). Forget the "market share" part, it's simply a great, lucid summary using linear, OLS regression of this class of models applied to marketing data like yours. "Pooled" refers to stacking up the static cross sections (your items) while "time series" refers to the monthly information. Your target variable is cost and there are a number of possible predictors (features, independent variables) that could be included in such a model, e.g., items sold, items (categorical), month, year, a proxy for price (cost divided by # of items), and so on. Cooper recommends normalizing the target variable with a log-centering transformation but taking the natural log is simpler and less subject to retransformation bias. The key outputs would focus on the magnitude of the slopes (rate of change in cost as a function of items sold) and acceleration (the rate of change of the rate of change -- the slope). 
For a more sophisticated approach, a set of nonlinear mixed models in R are discussed on this website containing lecture notes by Weiss... http://www.unc.edu/courses/2008fall/ecol/563/001/docs/lectures/lecture27.htm
Other possible methodological avenues include:


*

*% change, as you've noted

*CAGR, compound annual growth rate, formulas for which are all over the web

*Parametric t-tests for trend detection (http://journals.tubitak.gov.tr/engineering/issues/muh-03-27-4/muh-27-4-5-0206-6.pdf)

*Nonparametric Mann-Kendall test For monotonic trend (http://vsp.pnnl.gov/help/Vsample/Design_Trend_Mann_Kendall.htm)
These suggestions only begin to scratch the surface of possibilities.
