Author: Siqi Jiang
Advisor: Dr. Ionut Florescu
Date: 04/16/2021
Department: School of Business
Degree: Master of Science
ABSTRACT
Default probability prediction is a crucial topic because it is an influential factor in assessing listed corporations' credit risk in the financial market. While previous studies have focused on structural data like financial ratios to make predictions, this thesis examines the impact of unstructured social media data on the prediction of default probability. Specifically, we address this prediction problem in an attempt to model a joint-learning framework with multi-inputs. First, natural language processing was conducted to extract meaningful information from the social media news dataset using the transformer framework's encoder block as a feature extraction layer. Subsequent feature extraction of financial ratio was examined and a concatenation of vectors from different sources was performed. The data revealed that the Multi-Source framework has higher Accuracy than the model considering financial ratio only.