| SEMINAR ŞTIINŢIFIC MONEDA, FINANŢE, BĂNCI – SEMINAR ȘTIINȚIFIC |
AI in portfolio optimization: challenges and opportunities
Joi, 21 noiembrie 2024 – ora 16:40 – sala 3207
Walter Farkas
University of Zürich and ETH Zürich; farkas@math.ethz.ch
Abstract:
Artificial intelligence is the automation of statistical analysis – aiming to produce highly accurate outcome distributions. In order to be reliable, statistical predictions crucially depend on whether the future observations are similar to the past. For AI to be effective, identified patterns must be reoccurring but on efficient financial markets identified performance patterns are unlikely to persist. In this talk, drawing on an ongoing research project with Andreas Zimmermann, we will show the new opportunities AI brings to portfolio optimization. Specifically, we will discuss how AI enables to quantify complex dependencies of financial instruments performance and the integration of their diverse outcome distributions in different financial market states to enhance portfolio optimisation.
Zoom:
https://ase.zoom.us/j/88975840175?pwd=Q0ZnbHl0TjhqcEtzVnhMUWRmdWdPQT09
Activitate de socializare: N/A
Balancing an unbalanced word: can deep learning models predict credit defaults without bias?
Marti, 26 noiembrie 2024 – ora 16:40 – sala 3M4 (etaj 1, clădirea Moxa)
Alexie Alupoaiei
National Bank of Romania; alexie.alupoaiei@bnro.ro
Abstract:
We investigate potential biases related to using AI/ML models for default predictions or credit scoring, choosing gender (female) as protective attribute. We propose a new methodology based on the principle of „equal risks, equal rights”, using Deep Learning Networks. Specifically, we investigate the probability of AI/ML models generating Type I errors (false positives) depending on the borrowers’ risk levels. We use a database containing all consumer and mortgage loans with value higher than EUR 4000 (nearly 900,000 debtors) granted by a European Union banking sector. The results show that proper use of ML model feeded with unbias dataset do not generate bias in outcome.
Zoom:
https://ase.zoom.us/j/88975840175?pwd=Q0ZnbHl0TjhqcEtzVnhMUWRmdWdPQT09
Activitate de socializare: La Radu
Cerc vicios sau virtuos? Analiza interacțiunii dintre riscurile bancare asociate deținerii titlurilor de stat și riscul suveran
Marti, 3 decembrie 2024 – ora 16:40 – sala 3M4 (etaj 1, clădirea Moxa)
Leonard Uzum
Bucharest University of Economic Studies and National Bank of Romania; uzumleonard@yahoo.com
Abstract:
We propose a rotated dynamic Nelson-Siegel model for modeling the Romanian sovereign yield curve and a tailored Lasso VAR model with macroeconomic and financial predictors for forecasting it. As expected, inflation and the key interest rate emerge as primary drivers of the yield curve. Nonetheless, our results also show that other factors significantly impact the yield curve, including credit growth, the current account balance, the market operations balance, interbank interest rates, international financial stress (proxied by the VIX), and local banking sector risk (measured by a stock volatility index). Furthermore, our results suggest that domestic bank risk is partially responsive to the spread component of the yield curve but remains unaffected by other yield curve characteristics. This study offers a valuable tool for fiscal policymakers aiming to comprehend yield curve dynamics in an emerging economy and for institutional investors considering borderline investment-grade sovereign bonds in their portfolios.
Zoom:
https://ase.zoom.us/j/88975840175?pwd=Q0ZnbHl0TjhqcEtzVnhMUWRmdWdPQT09
Activitate de socializare: Coolinart
Comparative Analysis of Base Line, Deep Learning, and Large Language Models in Credit Scoring Using Synthetic Datasets
Marți, 10 decembrie 2024 – ora 16:40 – sala 3M4 (etaj 1, clădirea Moxa)
Andreea Bozagiu
Bucharest University of Economic Studies; andreea.bozagiu@fin.ase.ro
Abstract:
This study compares traditional credit scoring methods, deep learning models, and large language models (LLMs), using synthetic data to protect privacy and ensure consistency. Credit scoring has traditionally used methods like logistic regression and new AI models which may improve prediction accuracy. In this paper it was tested and evaluated these models baseline methods (logistic regression), deep learning (Gradient Boosting Machine and Neural Networks), and LLM-based models for feature extraction and prediction looking at performance in areas like accuracy, precision, and recall. The results show that deep learning and LLM-based models perform better with complex data, while traditional models still work well with lower computational demands. This paper provides valuable insights into balancing accuracy, interpretability, and computational efficiency when developing credit scoring models.
Zoom:
https://ase.zoom.us/j/88975840175?pwd=Q0ZnbHl0TjhqcEtzVnhMUWRmdWdPQT09
Activitate de socializare: Zăganu
Democracy, Pandemic Crisis and Dividend Policy
Marți, 17 decembrie 2024 – ora 16:40 – sala 3M4 (etaj 1, clădirea Moxa)
Victor Dragotă
Bucharest University of Economic Studies; victor.dragota@fin.ase.ro
Abstract:
Using as data 120,690 firm-year observations from 12,069 companies from 59 countries, we tested whether different indicators for democracy, pandemic crisis, and their interaction explain dividend policy. Our results are mixed. Propensity to pay dividends is positively influenced by democracy, but only for the entire sample and for the developing countries; for developed countries, democracy has a negative impact on this indicator. On the other hand, for dividend / assets, the impact of democracy is reversed. This result can be explained by the different informational content of these two indicators. Dividend policy was negatively affected by the COVID-19 pandemic. Finally, we have interacted democracy with epidemic crisis and analyzed its impact on dividend policy.
Zoom:
https://ase.zoom.us/j/88975840175?pwd=Q0ZnbHl0TjhqcEtzVnhMUWRmdWdPQT09
Activitate de socializare: Old Kitchen