Design of a machine learning based system for pharmaceutical purchases.
Abstract
To obtain inherent laws from vast amounts of pharmaceutical sales data and to provide
valuable information to pharmacy managers, this work validates di erent methods and
approaches to perform a sales forecast. Part of the data is used to train a neural network
algorithm, with backpropagation for some methods, step by step, where shallow nets face
selected scenarios, with di erent space-time data considerations.
In each method, by using a sum of square di erences, and a peak search procedure, a reasonable
quality in the obtained abstract representations is pursued. First, an auto-encoder
is trained to develop in its hidden layer neural data abstractions about a random-moving
window. Thereafter by using the abstraction of the net plus recently captured information, a
second shallow net is trained to produce its own one-day ahead estimates, using new timing
and data procedures. After training, the whole stacked system's performance is compared
with the naive forecast scenario's mean square error and if it's a better value, the method
is used to produce stable daily forecasting for assorted products and periods. The system
has been tested in real-time with real data.