Essays on machine learning for economics and finance

Econometrics and machine learning are quite close and related concepts. Nowadays, it is always more important to extract value from raw data, and distilling actionable insights from quantitative values as well as qualitative features. In order to deal with these topics, the first chapters (Chapter 1 - 4) are going to introduce the new wave called machine learning or big data and they will explain the most common techniques used in the field, respectively regression, clustering, model selection, and tree-based models (Chapter 2); time series analysis (Chapter 3); and eventually forecasting model with shrinkage methods (Chapter 4). Then, three applications are going to be provided. In Chapter 5, it is going to be shown an example of big dataset for the insurance vertical. Rothschild and Stiglitz ([30]) argued that people signal their risk profile through their insurance demand, i.e. individuals with a high risk profile would buy insurance as much as they can, while people who are not going to buy any insurance are the ones with a lower risk profile. This issue is commonly known as adverse selection. Even if their prediction seems to work quite well in a lot of different markets, Cutler et al. ([13]) proved that there exist some insurance markets in United States in which the expected result is completely different. In the wake of this study, we provide empirical evidences that there are some European insurance markets in which the low risk profile agents are the ones who buy more insurance. In Chapter 6, a second application is going to be provided. It has been studies the effect of behavioural biases on entrepreneurial choices to insure their firms against kinds of corporate risks. It has been used a large sample of Italian Small and Medium sized - finding that they under-insure themselves. The dataset allows to link corporate insurance choices with the personal traits of the entrepreneur and his household’s financial choices. In Chapter 7, finally, an application to financial markets is going to be shown. Bollen et al. ([10]) reintroduced the idea of formulating prediction based on the general sentiment of the investors, even if they originally exploited microblogging data. The purpose of this study is to verify whether social data may have a predictive power for the stock prices, returns, and volumes. The analysis has been implemented for different large technology companies, and the robustness has been tested through a ten-days rolling window. The evidence shows that there is some intrinsic value in these new features, and that both the sentiment and the amount of tweets posted online can improve the forecast given by a baseline autoregressive model. Some additional variations have been tested eventually with the same dataset.