Sales Forecasting Models
The analysis of marketing databases, as well as spatial analyses are critically important for any company that sells products or services. They do however have limitations. They are bivariate in nature, in other words, they only take into account two variables at a time, namely, sales on one hand, and one other variable, on the other.
Each variable that individually correlates with sales at a given point of sale (household income, ethnicity, socio-demographic segments, etc.) must therefore be combined with other significant variables so as to take into account their combined effect, i.e. colinearity.
To do this, a multivariate equation must be created that will enable us to weight the effect of the different variables relative to one another and thus eventually provide us with a stable model that will help us predict sales at any given new point of sale.
The reliability of such a model can be defined in three ways: based on its R2 value (the goodness of fit of the model), the average error (the divergence between predicted sales and real known sales) and finally, the confusion matrix (the number of good decisions that the model would have us make). One always aims for an R2 value greater than 70%, an average error less than 15% and a confusion matrix whose reliability is above 90%.
We feel that any company that is planning to open more than 20 sites per year and for which a hundred or more sites must be analyzed will greatly benefit from using a multivariate forecasting model. For companies whose needs are more modest, it is possible to make use of an analogous model that, while less accurate, will nonetheless help orient decision making in this area.
