Contrastando el Machine Learning y la Econometría en series temporales.

Abstract

This study compares the accuracy and complexity of econometric and Machine models. Learning in the prediction of household credit volume in Ecuador. Are analyzed ML models such as random forest and Xtreme Gradient Boosting along with models econometrics such as ARIMA. In addition, Google search data is incorporated related to credit. The initial results reveal that the random forest model outperforms econometric models, with a mean absolute percentage error (MAPE) of 1.5 percentage points lower than ARIMA. The inclusion of non-traditional data reduces MAPE in multivariate models. This study also demonstrates how Explanatory tools can help understand nonlinear relationships and the importance of variables in ML models. In summary, this work contributes to the discussion on the applicability of some tree-based ML models, in the economic forecasting, a task that is inherently non-deterministic.