The increasing amount of information available in many countries generates opportunities to anticipate events related to economic behavior. One of the most important tasks is the prediction of economic activity in real time in order to anticipate slowdowns and possible recessions. The prediction of economic activity in real time is known as nowcast, and it requires a large amount of information. However, the recent pandemic has increased the lack of information and in consequence, it makes difficult the quantification of the economic context in the short-run. At present, dynamic factor models (DFM) are the preferred approach to nowcast in order to deal with the high dimensionality of data and the absence of information, however, recent research suggests to nowcast economic and financial variables by using deep learning approaches, including non-structured data in the modeling, which has the property of summarize the economic context in different ways. In this thesis, we propose a deep learning approach to nowcast for Mexican economic activity, and we make an extensive comparison with econometric models. We make use of traditional data, such as macroeconomic and financial data, and nontraditional data, such as indicators based on on-line newspapers. Additionally, we use variable selection methods to reduce the dimensionality of the data while keeping the relevant information for nowcast. As a result, we find that deep learning models are more flexible in terms of their lack of assumptions and that they are able to learn turning points in Mexican economic activity, and the non-structured data improve the nowcast, particularly, when standard information is not available.