Macroeconomic Variables Forecasting using Machine and Deep Learning Models

The work presented fits into the context of data science and is proposed as a study regarding the application of machine learning techniques for predictive analysis, focusing on time-series data. In recent years, there has been a considerable increase in the need to be able to analyze and understand phenomena of an economic nature and, for this reason, the development and application of methods capable of identifying the same phenomena and any unusual pattern. It can be essential, indeed, being able to predict anomalous behaviors, as they can turn into problems rather than significant opportunities. The main objective of this project is to try to develop an efficient method for forecasting economic variables, using models and techniques able to consider also the uncertainty of the forecasts themself. Machine and deep learning models are used, considering different structures and observing if some of their characteristics prove to be more suitable for achieving the set goal. Starting from historical information about the time-series we want to study, and other different variables, the models are trained and valuated on how good they perform into predicting future values and anticipating future anomalous behaviours. In the course of the research, also pre-processing methods for data are considered with the aim of obtaining a clean starting dataset for the analysis and at the same time facilitating and speeding up model learning. Different data scaling and feature reduction/selection techniques are tested; then the work is carried out with the best combination found. After testing more innovative models, also the use of more traditional and statistic ones is proposed, in order to have a fair comparison of the performance of two different approaches on the same type of data. The goodness of the results are measured using specific metrics, adapt to the time-series analysis. In conclusion, this master thesis intends to be presented as a contribution to the experimentation of more innovative techniques for the prediction of economic variables, described by data in the form of time-series, proposing a further comparison with models and methods already widely used and consolidated. Following a first attempt which has as its objective the punctual forecast of future values, the consideration of the uncertainty with which these values are produced through the construction of forecast intervals is proposed.