Back to search

Institutional Repository of Vilnius University

Kredito įsipareigojimų nevykdymo Baltijos šalyse prognozavimas /

Abstract

dc:description

This thesis investigates credit default prediction in the Baltic states loan portfolio using machine learning methods applied to monthly time series data. The primary objective is to develop and compare two early warning systems capable of identifying loan agreements at risk of default and estimating the remaining time until the default event occurs. The study is motivated by the growing regulatory requirement for proactive credit portfolio monitoring – the European Banking Authority guidelines (EBA/GL/2020/06, Section 8.5) explicitly mandate that credit institutions maintain early warning indicator systems for deteriorating loan quality. While traditional probability of default models assess credit risk at a point in time, the dynamic nature of loan data enables a broader question: can machine learning methods trained on sequential data predict not only whether a default will occur, but also when. The dataset used comprises monthly observations of loan portfolios from the three Baltic countries covering the period from 2016 to 2025. The data was sourced from a bank’s test environment containing synthetic and anonymised records. Extensive data preparation was performed, including removal of pre-observation events, correction of retroactively recorded defaults, removal of post–event observations, and feature engineering. A correlation analysis using the phi–k coefficient identified and removed 16 highly correlated variables, retaining the most informative features for modelling. Two machine learning architectures were selected based on a review of the academic literature on early warning systems and credit risk modelling: Long Short-Term Memory (LSTM) recurrent neural network, which processes sequential data step by step and learns temporal dependencies across the observation window, and XGBoost, a gradient boosting ensemble method widely applied to structured financial data. Both models were formulated as dual-task systems: a classifier predicting whether an agreement will default within a defined horizon, and a regressor estimating the number of months remaining until default. To ensure a fair comparison, both models received identical windowed sequences of the same agreements, with XGBoost receiving the sequences flattened into a two-dimensional feature matrix. Two prediction horizons were evaluated and selected for both models based on validation set PR-AUC performance. Both models achieved similar results on the test set. The LSTM model correctly identified 78.8% of defaulting agreements, while XGBoost correctly identified 75.9%. At the window level, LSTM achieved a ROC-AUC of 0.893 and PR-AUC of 0.513, while XGBoost achieved 0.877 and 0.508 respectively. For agreements that were correctly flagged, the average lead time before default was 10.3 months for LSTM and 10.5 months for XGBoost, with a median of 8.0 months for both models – providing sufficient time for preventive action such as loan restructuring or intensified monitoring. XGBoost demonstrated higher precision (0.720 vs 0.651), producing fewer false alarms per flagged agreement, while LSTM demonstrated higher agreement-level recall and is therefore recommended as the more suitable choice for risk management applications where missing a default is costlier than a false alarm.

Degree

thesis:*
Grantor dc:publisher
Institutional Repository of Vilnius University
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Lazickaitė, Eglė Ona,

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
lit

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Lazickaitė, Eglė Ona,. Kredito įsipareigojimų nevykdymo Baltijos šalyse prognozavimas /. Institutional Repository of Vilnius University, 2026. https://repository.vu.lt/VU:ELABAETD308334651&prefLang=en_US