Brief description

The project proposes Early Warning Systems (EWSs) based on the joint use of Hidden Markov Models (HMMs) and Machine Learning (ML) methods as well as penalized likelihood and Bayesian Inference (BI) techniques. These EWSs can be used to generate meaningful warning signals of some risky events in a timely manner so as to allow policy makers to prepare appropriate responses. In Economics, tools of this type are crucial to preserving financial stability. Applications of the same relevance may be found, for instance, in primary and secondary health care.

Traditionally, EWSs are based on data about previous events and mainly rely on classification tools, such as logistic regression and k-NN, which are very popular in ML, whereas HMMs have an important role in Econometrics, as they allow capturing time-varying unobserved heterogeneity, which is crucial in the analysis of panel data. 

Our proposal is focused on different HMM formulations depending on the type of application. In particular, we distinguish the case in which past events may be clearly identified through an output binary variable from the case in which this output is not directly available. Using the terminology of ML, in the first case we are facing a supervised learning problem and an unsupervised learning problem in the second case. The main difference is in how the covariates (inputs) are included in the model. New HMM versions will also be proposed; these versions are based on a hierarchical structure of latent variables. The estimation/learning techniques of the proposed models will be aimed at regularization so as to avoid overfitting and obtain stability of predictions. In proposing such methods, we will also pay particular attention to possible problems arising in the applications, such as large data size and missing and sparse values. Moreover, on the basis of the adopted HMM for a specific application, the EWS will be calibrated by using appropriate measures of predictive power.

A crucial part of the project is the development of applications aimed at the prediction of:

1. banking crises at the macro level;

2. financial market crashes;

3. emergency health care utilization.

All applications will be based on suitable datasets, most of which are already available to the research team; emphasis will be given to the particular implications in terms of preventive policy measures.

The research team includes members belonging to four units (Perugia, Ancona, Milano-Bicocca, Palermo), who will closely collaborate during the 24-month period of the development of the project. The project has clearly an interdisciplinary nature, as testified by the variety of backgrounds of the researchers involved.

We plan to disseminate the results through publications and presentations at national and international level and packages for open-source software such as R that may also be exploited in Python and Stata. Moreover, we will recruit young researchers during the development of the project.