The use of textual data analysis and machine learning in bankruptcy prediction : evaluating the predictive power of sentiment scores and ratios from news articles for bankruptcy prediction in the Norwegian market using machine learning

Abstract

In this thesis, we investigate whether there is predictive power in sentiment scores and ratios derived from news articles with regards to bankruptcy prediction of Norwegian private limited companies. Our analysis is based on Norwegian news articles and annual accounts from the Brønnøysund Register Centre. We derive sentiment scores and ratios by performing lexicon-based sentiment analysis on the news articles. The sentiment scores and ratios are averaged for four different time observation periods and are then matched with their belonging companies. Furthermore, we utilize Altman’s five financial ratios to form our financial variables. Our models including both Altman’s financial ratios and sentiment variables are in our analysis compared to a reference model only including the financial ratios. In order to assess the problem we develop models using two different techniques, Generalized Linear Modelling and xgboost. Our emphasis is on comparing models with sentiment variables to reference models without sentiment variables in order to examine the potential predictive power of sentiment. We assess different model configurations, taking into account both different news observation periods and bankruptcy prediction horizons. The scores and ratios from the news observations are included on different time lags, ranging from 1 to 12 months prior to the announcement of annual accounts. The performance of the models is measured in AUC and balanced accuracy. In addition, we examine the average marginal effects in the developed GLMs and variable importance in the xgboost models. The results of the applied methodology indicates that there is no significant improvement when including sentiment variables. The reference models utilizing only financial ratios tend to perform better than the models including sentiment variables in terms of AUC and balanced accuracy. In terms of marginal effects and variable importances, the financial ratios also tend to outperform the sentiment variables. Furthermore, we provide a nuanced discussion based on the presented approach and results, and point to further research approaches that we find promising. Keywords – Bankruptcy Prediction, Textual Data Analysis, Sentiment Analysis, Predictive Analytics, Machine Learning, Big Data, xgboost, GLM