How Deep Learning is Changing Corporate Finance Worldwide


Global businesses face a difficult task in predicting credit scores. In this case, Massaron uses deep learning techniques to predict the credit ratings of global companies.

Luca Massaron, Senior Data Scientist, Kaggle Master and Google Developer Expert on ML, spoke at Deep Learning Devcon 2021, hosted by the Association of Data Scientists, earlier this month.

In his talk on “Deep Learning for Credit Scoring,” Massaron covered topics such as how deep learning techniques can be used to predict credit ratings for global business organizations; Bonds are compared to the most widely used traditional machine learning approaches, such as linear models and tree classifiers.

(Source: Luca Massaron | DLDC 2021)

According to Massaron, in a paper titled “An Artificial Intelligence Approach to Ghost Ratings”, the goal of the study was to demonstrate that neural networks can be a more effective technique for calibrating and predicting ratings than other modeling approaches currently used in banking.

Risk factors in credit ratings

Luca went further by explaining the importance of credit scores. International rating agencies such as Standard & Poor’s (S&P), Moody’s and Fitch assign credit ratings, which are alphanumeric indications of credit risk. Different ratings correspond to different expected probabilities of default. He went on to say that while rating companies claim to use both quantitative and qualitative data to determine a rating, they do not disclose the methodology used for the assignment. He further explained the risk factors involved in credit ratings.

The discussion continued with the data and overall workflow, moving on to the “Balance Sheet Index” and “Balance Sheet Ratios” in detail.

(Source: Luca Massaron | DLDC 2021)

This is followed by macroeconomic factors, including country-specific measures and eurozone indicators. Finally, the category embedding model of an artificial neural network was discussed.

Massaron went on to discuss drivers and workflow. They used a sample of 2469 annual corporate credit rating observations to train the model and assess its performance. Because their analysis focused on corporate debt, they omitted financial institutions and sovereign debt and looked at various sectors.

The architecture of the proposed model includes a deep neural network with multiple layers of densely coupled artificial neurons when discussing the model. Likewise, the word embedding technique was used, which consists of modeling words and documents.

Prediction using SHAP

Also, he mentioned the original kappa coefficient when discussing the results, which is a randomly adjusted index of agreement between the algorithm and the ground truth. He went on to say that the SHAP (SHapley Addictive exPlanation) approach is used to assign each feature to a specific prediction.

Later, when discussing improvements, the speaker mentioned a complex nonlinear model against a few cases that don’t cover the whole state space. He explained the Double Deep Descent, which occurs in CNNs, ResNets, and Transformers while discussing regularization.

Similarly, while considering the increase in data, he said that one method to overcome the problem of limited data is to try to expand it. Mixup is an augmentation that seems to focus on dissimilarity rather than similarity. Moreover, he said that learning from a few cases leads to an irregular and discontinuous decision boundary in a neural network. It becomes smoother as a result of increasing the mixture.

The results of this work also demonstrated adequate accuracy across different scoring classes when applying categorical integrations to artificial neural network (ANN) architectures.


Comments are closed.