An Econometric Analysis of Unemployment in Ecuador: Classical Approaches and a Machine Learning-Based Proposal

Abstract

This study focuses on modeling and predicting unemployment in Ecuador using neural networks, comparing their effectiveness with traditional econometric methods. The dependent variable, unemployment (), defined as the percentage of the working population without employment, is analyzed based on seven macroeconomic predictors: Real gross domestic product (GDP, x₁), inflation (x₂, percentage change in the consumer price index), exports (x₃) and imports (x₄, valued in millions of USD), real minimum wage (x₅), public spending (x₆, percentage of GDP), and interest rate (x₇). Previous research, based on Okun's Law, identified linear correlations between GDP growth and unemployment reduction using Ordinary Least Squares (OLS) and cointegration models. However, these conventional techniques, including multiple regressions, showed limitations in capturing the nonlinear dynamics inherent in unemployment. In contrast, neural networks (NN), especially sequential architectures, demonstrated superiority in modeling complex interactions between variables. The study evaluates deep learning models with one and two hidden layers, identifying an optimal configuration of one hidden layer (16 neurons), achieving a loss of 0.2846, 75% accuracy, and Mean Square Error (MSE) of 0.2064 on unseen test data. The results highlight the ability of neural networks to model unemployment more accurately than linear methods, attributing this to their adaptability to learn nonlinear patterns in multivariate data. This study demonstrates that variables such as GDP, inflation, public spending, and interest rates significantly influence unemployment. Neural networks (NN) effectively capture these complex relationships. Their usefulness in predicting socioeconomic phenomena in developing countries such as Ecuador is highlighted, overcoming the limitations of traditional models.