Forecast of the financial risk using time series analysis

The assessment of financial credit risk is a challenging and important research topic in the area of accounting and finance. Economic crises indicate that there is still no stable or globally valid solution for estimating the financial credit risk with sufficient accuracy. At the economic and banking level, credit institutions and credit information systems are looking for new methods of analysis on the data in their possession; in particular, the enormous amount of micro-transactions due to the advent of cashless transactions has not yet been exploited. In this dissertation, credit scoring models are proposed, using real payment data retrieved in a Payment Services PSD2 - Directive (EU) context. The data under analysis, i.e. a list of movements, is made available through the Account Information Service (AIS). But the data are not provided with a whole series of information to which credit bureaus and banks have access. Therefore, different solutions found in the literature have been extended and adapted to the available data. The goal of this thesis is to reproduce the models used in the state of the art and combine the results of supervised and unsupervised learning methods applicable to the dataset under analysis, and combine the results in such a way that it is possible to take into account the ability to spend, earn, save and invest, along with the probability of incurring fraud (Direct Debit Fraud), generating as output a numerical value that is representative in terms of risk score and can provide a more appropriate tool for credit scoring analysis. In an initial supervised learning phase, we use a fraud label extracted from internal research to train our model to classify users as fraudulent or not fraudulent. Then, different unsupervised methods, through the use of hierarchical clustering, have been explored to predict the credit risk score, because we want the algorithm to be generalizable to the context we are seeing of openBanking, where the fraud or non-fraud information would not be present. So the training of the algorithm cannot be supervised as it is not possible to rely on already calculated scores associated with the examples contained in the dataset. In parallel, results obtained from time series forecasting are also added, allowing the time series to be extended into the future. As a validation process, state-of-the-art research in banking risk management was compared also exploiting the additional fraud information to better control the clustering results. In particular, the models most used by credit bureaus to calculate similar scores (credit risk score, income risk score and consumer credit score) have been analyzed and adapted in order to be usable on the dataset under analysis, and therefore more usable in the context of open banking that we will see in the coming years. The results achieved and analyzed allow us to easily understand that there is no globally accepted method that has been shown to be better than others except on specific datasets that are not particularly significant at the level of structured research in this domain. Hence the importance of further analysis in this area by comparing different machine learning techniques to assess credit risk. In fact, we were able to obtain meaningful clusters by exploiting additional fraud information, but in a general context it would be appropriate to have financial domain experts able to validate and verify the cluster generation process on time.