Artificial intelligence applied by means of deep learning (neural networks) and by the traditional econometric method of ordinary least squares regression

Abstract

The field of application for instruments involving artificial intelligence (AI) algorithms is increasingly broad in today's world. In practically all sciences, this branch of learning advances in several areas and specificities. Thus, this research aimed to verify, in applied and didatic terms, whether the traditional econometric methodology of ordinary least squares (OLS) regressions and the neural network (NN) methodology have the same performance in different situations of empirical tests for overlapping data series in artificial intelligence algorithms. The research problem was to know whether the OLS method and the NN method, in general, present satisfactory results in statistical estimates and for certain types of AI algorithms. The research hypothesis is that traditional methods, such as OLS, work better for structured data, while neural networks excel with complex/random data. In advance, the quantitative econometric tests carried out demonstrated that the use of the NN and OLS methodologies are approaches that can be used together when applied to AI algorithms. The quantitative results also demonstrated that the RN technique, in general, presents better results for data with characteristics of random variables. And the traditional econometric method MQO presents, in general, better results when the observed data have characteristics of “more well-behaved” and less random variables in mathematical terms. Furthermore, the main criterion used to measure performance was the Mean Squared Error (MSE) indicator, in addition to the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination R² indicators.