# Sales forecasting to account for regression

I have a very beginner question. I am attempting to forecast total 2014 unit sales of a large number of products.

The data I have has 10 points for each individual product, which are the total unit sales for each of the previous 10 years. I also have the total yearly store revenue for the last ten years, which I can use to establish an overall average growth.

I don't want to use a simple arithmetic mean, because I want to account for regression on products that performed exceptionally well or badly in 2013.

I'm assuming the variation is random, and am not trying to account for promotions or seasonality.

I'm also assuming that the base growth rate for each product is the average growth rate of total store revenue.

I just want a reasonably accurate prediction of total unit sales for each product to set guidelines for inventory purchasing.

I am working in Open Office Calc, and I also have R, but am a total beginner in using it. What would be the best way to approach this problem, and how do I implement it in these programs?

• I'm not sure I understand what your model is/what your assumptions are from this. – Glen_b Mar 14 '14 at 1:09
• I assume what is being looked for is double exponential smoothing. I have no idea how to do it in open office or even R without writing out forumala – charles Mar 14 '14 at 2:17
• My model is that sales of individual items are increasing at the same rate as overall revenue growth. But I want to project our unit sales for this year to account for random variation within an item. For example, total revenue growth for 2013 was 9%, but item A's unit sales were up 25%. I want to predict how many units of A we will sell in 2014 assuming that the 16% difference between the two numbers is influenced by random chance. I hope that clarifies my question. – AYD Mar 14 '14 at 20:12
• If I understand you correctly, you have a large number of time series with ten points observed for each series. You could look into some (simplified) model for multivariate time series. Specifically, there will be competition between the products, so you expect negative correlations within each year! which adds to the regression effect you are talking about. I would start by grouping the products into relatively small groups of related products, for which competition is expected, and then try to learn about correlations and variances. Maybe differencing could take care of the growth? – kjetil b halvorsen Jul 9 '14 at 1:11