Our latest paper, from our lab member Mahsa Tavakoli, is out. This was in collaboration with Prof. Rohitash Chandra at UNSW in Australia

When it comes to understanding how credit ratings are determined, most studies have focused only on numbers—like financial ratios and balance sheets. But in the real world, a lot of important information is found in written documents, such as company reports or news articles. In our study, we looked at how combining both numbers and text using deep learning models could improve the prediction of credit ratings. We tested different types of models and ways to blend the data, and we found that a model based on CNN (a type of deep learning model) that mixes information early and in the middle of the process gave the best results. Surprisingly, simpler models worked better than more complex ones, and the text information—like what’s written about a company—was even more useful than numbers in predicting credit ratings.

We also checked how reliable our model is when things change, such as during big events like the COVID-19 pandemic. The results showed our model stayed stable and still made good predictions, especially when using text data. Another interesting finding was that credit ratings given by Moody’s were more accurate over time than those from other rating agencies. This could help financial institutions trust those ratings more when making decisions. Overall, our study shows that using both text and numbers together can lead to smarter and more reliable credit rating predictions. It also opens the door for better tools that can help people make confident financial decisions—even during uncertain times.

The paper was published a bit ago at Applied Soft Computing. You can find the paper Open Access here, and its ArXiV version here.