Using Machine/Statistical Learning to Estimate and Forecast Sales


Estimating national and state sales for a class of products required models using biased data.


A machine/statistical-learning algorithm, “Elastic Net,” was used to optimize the choice of a model. This would extrapolate sales (from a biased sample of convenience) to the entire country and aggregate them down to sub-state-geographic levels. An extension of the method creates forecasts to future periods.


The predictive power and scope of applying an existing model was increased. The new one creates accurate sales estimates, along with statistical indications of their precision (in the form of standard errors).  The estimates are consistent over time, permitting useful year-over-year growth comparisons and market share estimates.

This illustration shows the distribution of one particular product among all the relevant hospitals in the forty-eight states. The colors and sizes of the circles indicate total sales in a given month. The gray dots indicate hospitals where sales must be predicted.