Nowcasting of the economic dynamic in the Dominican Republic under the Covid-19 crisis: An approach based in Big Data and machine learning techniques

Resumen

The availability of a huge amount of information, both quantitative and qualitative, structured and unstructured (big data), jointly with the emergence of machine learning techniques, appropriate for its treatment, has led to a revolution in various areas of knowledge, extending the data pool and analytical framework for decision making processes. In this sense, it has become possible to generate answers to various research questions that could not be addressed under a traditional empirical approach. Macroeconomic modeling has been considerably favored by the big data-machine learning binomial, mainly under the prevailing situation of the Covid-19 crisis, where the speed and greater scope of decision making processes have appealed to the use of granular information, available in real time. The objective of this document is to present a set of models, based on machine learning techniques, used to carry out nowcasting and generate leading indicators of the economic activity in the Dominican Republic. Through the system used (composed of a bayesian structural model of time series, a multi-layer perceptron and lasso and ridge regressions) the forecast errors for the monthly index of economic activity (IMAE) are minimized (measured in terms of the root mean squared error, RMSE) and has better performance than the benchmark, which is an autoregressive integrated
moving average (ARIMA) model. The potential of these techniques is emphasized in terms of macroeconomic modeling, to assess the balance of risks of certain variables and for the decision making processes, even in non-linear or chaotic scenarios.