Validation of machine learning models through statistical field experiments

Objective - This article presents a practical application with the development of a statistical field experiment in a premium aluminum can industry in the United States, aiming to statistically validate results from machine learning (ML) models, built in a previous phase of the study. Methodology: This study uses concepts of interventionist research, which involves field experiments where the researcher and host organization work together seeking to experiment in the system under study, and through observation generate knowledge. Originality/Relevance: Regarding originality, it is not common to find ML models validated by planned field experiments, followed by rigorous statistical analysis. And the proposal relevance is due to its contribution to the literature and the possibilities of replicating the study on a larger scale, in the company itself or in any other company that faces similar challenges. Main Results: In a previous phase of the study, ML models identified the variables with the greatest impact on inefficiencies (scrap generation) in an aluminum can production process. These variables were validated in this phase of the study, through a statistical field experiment, confirming the statistical significance of the ML model results. Theoretical and Practical Contributions: The research contributes in practical and scientific terms, as the statistical validation of ML models by planned field experiments is a contribution to the applied science literature, in addition to practical possibilities. Likewise, despite being widely used in different areas, interventionist research still presents an important gap in applied social sciences, especially in the management of industrial processes.