Research

The Association for the Advancement of Artificial Intelligence (AAAI) conference is recognized as an elite conference in the field of artificial intelligence (AI), gathering a community of academic and industry experts engaged in both theoretical and applied AI research. The 38th AAAI conference was held in the city of Vancouver, Canada, from February 20^th to 27^th. It offered an excellent opportunity for sharing research work in AI, bridging potential collaborations, and offering the guidance of the direction of AI innovation.

Both Cristian and Jet attended. Jet presented our latest research: “Breaking Barriers: Unveiling Gender Disparities in Corporate Board Career Paths using Deep Learning.” This research was showcased in two workshops: the AI in Finance for Social Impact, organized by J.P. Morgan, and the AI for Financial Services, highlighting the cross-disciplinary impact and relevance of our work. In the former, the lab was also represented in the work “Extreme climate events and credit risk: stress testing approach for loans to SMEs based on network analysis”, presented by the PhD candidate Camilla Carpinelli, part of the group led by Prof. María Óskarsdóttir, at Reykjavík University.

Celebrating Recognition: Best Poster Award

Our collective efforts were recognized with the Best Poster Award at the workshop, a moment of great pride and joy for our team. We extend our deepest gratitude for this honour to the organizers, reaffirming our commitment to contributing meaningful insights and solutions in the realm of AI and social impact in banking.

Networking, Collaboration, and Future Horizons

The AAAI conference provided an excellent opportunity for engagement with the community of AI in Finance. Our team had a great time to meet, network, and exchanging ideas with fellow researchers and practitioners from both academia and industry. Notable interactions included enlightening discussions with peers from prestigious institutions and companies such as CMU, J.P. Morgan, and Capital One, among others.

As we reflect on our experiences at the AAAI conference, we are filled with gratitude for the opportunity to contribute to and partake in the global dialogue on AI. We look forward to attending the next one in 2025!

HouseNetwork-e1707236743536.jpg?fit=200%2C200&ssl=1

Our latest preprint, titled “Attention-based dynamic multilayer graph neural networks for loan default prediction”, introduces a novel model that could enhance the accuracy of credit risk assessments.

Credit Scoring and Correlated Default

The inspiration for this work comes from our previous studies, which clearly show that borrowers are not isolated entities, but part of a complex web of connections that can influence their probability of default. This interconnectedness suggests that a borrower’s risk of default may be impacted not just by their financial situation, but also by the network of relationships they are part of.

Our study leverages these insights, proposing a sophisticated model that combines Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) to assess credit risk. This way, the model can use dynamic multilayer networks, each layer reflecting a different source of network connection. The proposed model considers different types of connections between borrowers, such as geographical location and choice of mortgage provider, and the evolution of these connections over time.

How It Works

GNNs are a class of deep learning models designed to operate on graphs — structures that represent relationships between entities. These models are adept at capturing the complex patterns inherent in networks of borrowers. On the other hand, RNNs excel at processing sequential data, making them ideal for analyzing the temporal dynamics of these borrower networks.

The model introduced in our study, by our PhD student Sahab Zandi, in collaboration with Prof. Christophe Mues and Kamesh Korangi from the University of Southampton and Prof. María Óskarsdóttir from Reykjavík University, adds a layer of sophistication with an attention mechanism. This mechanism prioritizes certain time points over others, based on their relevance to the borrower’s default risk. Such an approach allows for a more nuanced analysis, distinguishing the model from traditional methods of credit scoring.

Empirical Evidence of Superior Performance

When tested against a dataset provided by U.S. mortgage financier Freddie Mac, the model not only outperformed traditional credit scoring methods but also offered more in-depth insights into the nature of default risk. We found that the model’s ability to account for the dynamic and multilayered nature of borrower connections enhanced its predictive accuracy. This suggests that the future of credit scoring lies in the ability to understand and model the complex web of relationships that influence financial behaviour.

Looking Ahead

The implications of this study are multifaceted. For lenders, adopting such models could help understand how risk propagates and affects both individual borrowers and their portfolios in a high-stakes market. For borrowers, it could translate into more access to credit by empowering so-called ‘Second Look’ models, which provide thin-file borrowers with a more detailed evaluation. Our results can be part of such evaluation. And for the field of operational research and finance at large, this study paves the way for further exploration into the use of machine learning and network science in multilayered, dynamic, environments.

As we move forward, the exploration of even more sophisticated models — incorporating additional layers to capture a broader array of connections or employing different types of GNNs and RNNs — promises to unlock new insights into credit risk and beyond. The journey towards a more interconnected and intelligent approach to credit scoring is just beginning, and its potential benefits for both lenders and borrowers are immense.

Interested in the topic? Read the working paper on ArXiV!

Untitled150x150px4.png?fit=150%2C150&ssl=1

A bit late to the party, but as the first paper on this topic was recently published at EJOR (ArXiV, Journal), I thought it may merit the post.

Last year at the credit scoring conference, I was interviewed by Brendan Le Grange, host of the podcast How to Lend Money to Strangers. We discussed the paper I presented there, part of Sherly’s PhD thesis, on causal models for credit limit settings. The preprint of that paper will appear soon, read the first one meanwhile!

If you want to hear more about my thoughts on credit limit setting and how it relates to causal modelling, listen to the podcast episode in this link.

We had a great time attending the 2023 INFORMS Annual Meeting that took place between October 15^th to October 18^th in Phoenix, AZ. This is one of the largest conferences in the field of OR, with 6,000+ attendees and 1,400+ sessions.

The BAL had a strong presence at the conference with four presentations:

On October 15^th, we had two presentations:
- Cristián Bravo presented the work with Kamesh Korangi and Christophe Mues on “Large-Scale Portfolio Optimization using Graph Attention Networks”.

Daniel Abib presented the work with Cristián Bravo, Raffaella Calabrese, and María Óskarsdóttir on “Optimal Feature Split in Classification Models with Dependency”.

On October 17^th, Sahab Zandi presented the work with Kamesh Korangi, María Óskarsdóttir, Christophe Mues, and Cristián Bravo on “Leveraging Dynamic Multilayer Networks for Modelling Credit Risk Contagion in SMEs”.

On October 18^th, Yuhao (Jet) Zhou presented the work with Collins Ntim, María Óskarsdóttir, Matthew Davison, and Cristián Bravo on “Uncovering the Network Power Gap: A Deep Learning Approach to Investigating Gender Disparities in the Boardrooms of Canadian Public Firms”.

We also had fun trying a few good restaurants in Scottsdale, AZ!

This was a great chance to showcase the preprints that will come out in the next few months. Stay tuned for them!

We had a great time attending the Credit Scoring and Credit Control Conference XVIII that took place between August 30^th to September 1^st in Edinburgh, UK. This conference bridges the academic/practitioner divide and is the world’s premier conference for credit scoring and credit risk related topics.

The BAL had a strong presence at the conference with six presentations:

On August 30^th, Cristián presented the work with our PhD student Mahsa Tavakoli, cosupervised by Rohitash Chandra from UNSW, on “Multi-Modal Deep Learning for Midcap Credit Rating Prediction Using Text and Numerical Data”.
On August 31^st we had two presentations:
- Our collaborator Prof. María Óskarsdóttir from Reykjavík University, Iceland, presented the work by our PhD students Sahab Zandi and Kamesh Korangi, cosupervised by Prof. Christophe Mues from Southampton University and Cristián, titled “Credit Scoring with Dynamic Multilayer Graph Neural Networks”.
- Cristián presented the work led by our PhD student Sherly Alfonso Sánchez, cosupervised by Prof. Kristina Sendova here at Western, called “Causal Learning for Credit Limit Adjustment in Revolving Lending Under Adversarial Goals”.
On September 1^st, we had three:
- Daniel Abib, who joined earlier this year as a postdoc at the Lab, presented the work coauthored with Prof. Raffaella Calabrese for Edinburgh University, Prof. María Óskarsdóttir, and Cristián. The work was called “Optimal Feature Split in Credit Risk Models with Dependency”.
- Our PhD student Kamesh Korangi presented the work from his PhD, coauthored with Christophe Mues and Cristián, on “Deep Temporal Graph Networks for Behavioural Scoring Prediction in Revolving Credit Lines”.
- Our PhD student Sahab Zandi presented the work with coautored with Kamesh, and cosupervised by Prof. María Óskarsdóttir, Prof. Christophe Mues, and Cristián. These last two works are part of the collaboration with one of the largest consumer banks in the world. Sahab’s presentation is titled “Modelling Credit Risk Contagion for SMEs over Supply Chains using Dynamic Multilayer Networks”.

The conference provided a great opportunity to meet and network with people in the field of credit risk from both academia and industry. We were honestly surprised and happy with the reception that we had from the conference attendants. We had many interesting talks and we look forward to what will come out of these chats!

We also had a blast having a reunion with some friends and colleagues after a while in Edinburgh!

We would like to thank the organizers, Professor Galina Andreeva and Professor Jonathan Crook from the Credit Research Centre at the University of Edinburgh, plus of course our collaborator Prof. Christophe Mues for hosting this wonderful conference. We look forward to attending the next one in 2025!

I had a great time presenting a keynote and a paper at the SIGKDD 2023, one of the elite conference in computer science. In my personal opinion, the KDD is the top applied data science conference, as NeurIPS is a bit too theoretical, while the KDD is a bit more integrative. Lots of papers this year in network science and causal learning, which was encouraging when evaluating the current research lines of the group.

I presented two works in the Machine Learning in Finance workshop of the conference:

The keynote “Leveraging Deep Learning for Multimodal Data Analysis in Credit Risk Assessment“, where I summarized the latest work of the lab in using multimodal data for the analysis of credit risk in midcaps and retail lending. I presented results of our latest preprints (paper on midcaps, paper on mortgages) and preliminary results on using LiDAR for mortgage analysis and social network analysis for credit risk modelling. I shared the stage again with three other keynotes from Bloomberg and Blackrock, the same groups that presented in the Columbia-Bloomberg seminar in May we also presented in, and with Srijan Kumar from Georgiatech. Srijan was keynote in the NeurIPS workshop we presented our paper on influencer detection (NeurIPS paper, journal preprint). The machine learning in finance world seems to be pretty small!
The paper “Graph Attention Networks for Portfolio Optimisation: Empirical Evidence for Mid-Caps” by our PhD student Kamesh Korangi. Kamesh couldn’t attend as the US visas are taking forever. In this talk I showed the preliminary results of our work on using GATs for optimizing portfolios. We are really excited about this work! I can’t wait to show the world about in the near future. Stand by for the preprint, it should be available in a few months. The presentation itself will be available in a few weeks, I’ll update this post when it is.

It was a fantastic experience to attend the KDD in person. Sadly, we weren’t able to do so in 2020 due to the pandemic, where we presented the preliminary results of our paper in default propagation across multilayer networks (KDD paper, YouTube video with the presentation at CORS 2021, extended journal paper). It was great to be able to present now and share with so many top researchers.

The conference was in Long Beach, so I also got to have some great weather.

Cristian in front of the sunny Long Beach Convention Center

The acceptance rate of the conference was higher than previous years, around 20%. However, the MLF workshop was even more selective with only around 10% of the papers being selected for spotlight talks! It was great to be in such an exclusive group.

Slide with the acceptance rate of the KDD conference per areas. Finance is 22%.

Also, it was great to see so many Latin Americans there! We had a few meetings and even some went salsa dancing. We were very pleased to also meet with Ricardo Baeza-Yates, a legend in the field.

Latin Americans sharing a lunch at the KDD.

Overall, it was a great experience. Hopefully we’ll be able to attend to KDD 2024 in Barcelona!

Our research focuses on using reinforcement learning (RL) to address the credit limit modification problem for companies offering credit card products. This involves two main challenges: defining the RL problem for this specific task and training the RL agent without conducting online experiments with customers.

To define the RL problem, we consider the financial history of credit card holders and the expected losses due to defaults when deciding whether to increase or maintain their credit limits. The actions available are increasing the limit or keeping it the same. We calculate the reward function based on the expected profit, considering the revolving aspect of credit card usage. This differs from previous studies that overlooked this aspect in profit calculations.

To train the RL agent offline, we use a two-stage model to simulate the balance after taking an action. This involves selecting the balance type and predicting the balance amount using a regressor model. Through our experiments, we found that our trained Double-Q learning agent outperformed other strategies, including the one used by Rappi, a Latin American fintech company known for its delivery and commerce services that has also ventured into banking with its RappiCard credit card, and that was our collaborator in this research.

Our research contributes by providing a conceptual framework for applying RL to credit limit adjustments and emphasizes data-driven decision-making rather than relying solely on expert judgments. Furthermore, we discovered that incorporating additional predictors did not improve the performance of our simulator. This implies that fintech companies do not necessarily have an advantage over traditional banking institutions in this specific task. Figure 1 provides an overview of the proposed methodology’s general workflow.

Figure 1: Methodology’s general workflow.

Link to the working paper: https://arxiv.org/abs/2306.15585

By Mahsa Tavakoli @Bal:

Our research study was undertaken with the aim of enhancing the accuracy of predicting company credit ratings, a critical factor in evaluating their financial stability. Unlike previous studies that solely focused on structured data, such as numbers and figures, we recognized the significance of incorporating other, non-structured information. Thus, our primary objective was to evaluate the effectiveness of employing advanced deep learning models to merge both structured and unstructured data, particularly textual information. Through this approach, we sought to provide a more comprehensive analysis and improve the overall predictive capabilities of the models. In our quest for the optimal approach, we conducted thorough testing of various fusion strategies and deep learning models, including CNN, LSTM, GRU, and BERT. To our surprise, we discovered that a CNN-based model (Figure 2), which effectively amalgamated data from diverse sources, outperformed more intricate models. Leveraging this model enabled us to achieve highly precise credit rating predictions.

Furthermore, we delved into the contribution of different data types to these predictions. Textual data, such as insights shared by company managers, played a pivotal role, particularly during challenging periods like the COVID-19 pandemic. This underscored the significance of contextual information and managerial perspectives in accurately predicting credit ratings.

Additionally, our research encompassed a comparative analysis of ratings provided by various agencies. Moody’s credit ratings emerged as the frontrunner, surpassing those of other agencies like Standard & Poor’s and Fitch Ratings, especially in medium-term predictions.

Collectively, our research provides a comprehensive framework that empowers rating agencies and financial institutions to make well-informed decisions when assigning credit ratings. By incorporating a combination of structured and unstructured data and leveraging the most effective deep learning models, they can significantly enhance the precision of credit rating predictions, thereby augmenting their overall decision-making process.

Fig1: Blending Textual Managers’ Insights and Companies Numerical Data for Precise Credit Rating Predictions

Fig2: Architecture of the CNN ensemble for the best model, showing
the convolution and dropout layers with two streams of data that includes
text and numerical data (N1, N2, N3, N4).

Link to the working paper: https://arxiv.org/abs/2304.10740

I had the true honour of being invited by Prof. Ali Hirsa to present at this excellent workshop. The organizers review 1400 abstracts published during the year and select the ones that according to them reflect the most significant research at the intersection of ML and Finance. Our work with Kamesh Korangi and Christophe Mues, “A transformer-based model for default prediction in mid-cap corporate markets” [paper, ArXiV], was one of them!

Excellent roster of speakers, including people from MIT, Stanford, Blackrock, Morgan Stanley, and of course, the hosts Columbia and Bloomberg. Everything went amazing except that the IT team lost a few of my slides, but I think the audience did not mind very much. I look forward to keep engaged with such an excellent team moving forward. Here are a few pictures of the event.

Cristián on the stage before his presentation at the Alfred Lerner Hall

The presentation was at the amazing Alfred Lerner hall at Columbia University. 400 attendants from industry and academia were present.

The stage at the Alfred Lerner hall

Research

Award Winners at the AAAI Conference!

The Power of Dynamic Multilayer Graph Neural Networks in Mortgage Credit Scoring

The BAL at How to Lend Money to Strangers – and the paper is out!

The BAL at INFORMS 2023

The BAL at the Credit Scoring Conference 2023

The BAL at the SIGKDD 2023

Optimizing Credit Limit Adjustments Under Adversarial Goals Using Reinforcement Learning

Multi-Modal Deep Learning for Credit Rating Prediction Using Text and Numerical Data Streams

9th Columbia – Bloomberg Machine Learning in Finance Workshop

The Banking Analytics Lab