Strategic Management of Sales Assisted by Machine Learning: A Case Study in a Large Food Business

Abstract

This paper implemented Machine Learning techniques to predict sales success of a large Brazilian food company. The strategy explores the use of recorded data to apply a pattern recognition process based in Multilayer Perceptrons (MLP) algorithms and thus proposes an efficient management of business proposals. As a research methodology, a database with information collected over 4 years (23,093 records) in an Enterprise Resource Planning (ERP) system was evaluated and standardized. Then sales success prediction experiments were performed, using MLPs of the following types: Standard Backpropagation, Backpropagation Momentum, Resilient Propagation, Backpropagation with Weight Decay and Quick Propagation. Resampling by test and validation sets was used, with the proportion variation of 65-35% and 75-25%. The number of neurons in the hidden layer were {3,5,7,10} and the periods of interaction were {50;100;1,000;5,000}. The Receiver Operating Characteristics (ROC curve) and its Area Under the Curve (AUC) were employed to evaluate the performance, considering the normalization of the measured AUCs between 0 and 1. The experiments are focused on the annual and monthly prediction of sales success, according to the profile of the customer. The main achievements of this work were: a methodology for collecting and standardizing the data of the company; an analysis of the best MLP configuration, according to the profile of the customer; and a structure that has an AUC between 0.975 and 0.983, with a processing time between 4.2 and 4.8 seconds, capable of supporting the decision taking of the company and reducing the loss of potential contracts (which reached the order of more than $180,000.00 Brazilian reais). Besides, the implemented approach allows the prediction of renegotiations and success cases.