Director of Machine Learning Insights [Part 3: Finance Edition]
If you're interested in building ML solutions faster visit hf.co/support today!
👋 Welcome back to our Director of ML Insights Series, Finance Edition! If you missed earlier Editions you can find them here:
- Director of Machine Learning Insights [Part 1]
- Director of Machine Learning Insights [Part 2 : SaaS Edition]
Machine Learning Directors within finance face the unique challenges of navigating legacy systems, deploying interpretable models, and maintaining customer trust, all while being highly regulated (with lots of government oversight). Each of these challenges requires deep industry knowledge and technical expertise to pilot effectively. The following experts from U.S. Bank, the Royal Bank of Canada, Moody's Analytics and ex Research Scientist at Bloomberg AI all help uncover unique gems within the Machine Learning x Finance sector.
You’ll hear from a juniors Greek National Tennis Champion, a published author with over 100+ patents, and a cycle polo player who regularly played at the world’s oldest polo club (the Calcutta Polo Club). All turned financial ML experts.
🚀 Buckle up Goose, here are the top insights from financial ML Mavericks:
Disclaimer: All views are from individuals and not from any past or current employers.
Ioannis Bakagiannis - Director of Machine Learning, Marketing Science at RBC
Background: Passionate Machine Learning Expert with experience in delivering scalable, production-grade, and state-of-the-art Machine Learning solutions. Ioannis is also the Host of Bak Up Podcast and seeks to make an impact on the world through AI.
Fun Fact: Ioannis was a juniors Greek national tennis champion.🏆
RBC: The world’s leading organizations look to RBC Capital Markets as an innovative, trusted partner in capital markets, banking and finance.
We all know that ML is a disrupting force in all industries while continuously creating new business opportunities. Many financial products have been created or altered due to ML such as personalized insurance and targeted marketing.
Disruptions and profit are great but my favorite financial impact has been the ML-initiated conversation around trust in financial decision making.
In the past, financial decisions like loan approval, rate determination, portfolio management, etc. have all been done by humans with relevant expertise. Essentially, people trusted “other people” or “experts” for financial decisions (and often without question).
When ML attempted to automate that decision-making process, people asked, “Why should we trust a model?”. Models appeared to be black boxes of doom coming to replace honest working people. But that argument has initiated the conversation of trust in financial decision-making and ethics, regardless of who or what is involved.
As an industry, we are still defining this conversation but with more transparency, thanks to ML in finance.
I can’t speak for companies but established financial institutions experience one continuous struggle, like all long-lived organizations: Legacy Systems.
Financial organizations have been around for a while and they have evolved over time but today they have found themselves somehow as ‘tech companies’. Such organizations need to be part of cutting-edge technologies so they can compete with newcomer rivals but at the same time maintain the robustness that makes our financial world work.
This internal battle is skewed by the risk appetite of the institutions. Financial risk increases linearly (usually) with the scale of the solution you provide since we are talking about money. But on top of that, there are other forms of risk that a system failure will incur such as Regulatory and Reputational risk. This compounded risk along with the complexity of migrating a huge, mature system to a new tech stack is, at least in my opinion, the biggest challenge in adopting cutting-edge technologies such as ML.
ML, even with all its recent attention, is still a relatively new field in software engineering. The deployment of ML applications is often not a well-defined process. The artist/engineer can deliver an ML application but the world around it is still not familiar with the technical process. At that intersection of technical and non-technical worlds, I have seen the most “mistakes”.
It is hard to optimize for the right Business and ML KPIs and define the right objective function or the desired labels. I have seen applications go to waste due to undesired prediction windows or because they predict the wrong labels.
The worst outcome comes when the misalignment is not uncovered in the development step and makes it into production.
Then applications can create unwanted user behavior or simply measure/predict the wrong thing. Unfortunately, we tend to equip the ML teams with tools and computing but not with solid processes and communication buffers. And mistakes at the beginning of an ill-defined process grow with every step.
It is difficult not to get excited with everything new that comes out of ML. The field changes so frequently that it’s refreshing.
Currently, we are good at solving individual problems: computer vision, the next word prediction, data point generation, etc, but we haven’t been able to address multiple problems at the same time. I’m excited to see how we can model such behaviors in mathematical expressions that currently seem to contradict each other. Hope we get there soon!
Debanjan Mahata - Director of AI & ML at Moody's Analytics / Ex Research Scientist @ Bloomberg AI
Background: Debanjan is Director of Machine Learning in the AI Team at Moody's Analytics and also serves as an Adjunct Faculty at IIIT-Delhi, India. He is an active researcher and is currently interested in various information extraction problems and domain adaptation techniques in NLP. He has a track record of formulating and applying machine learning to various use cases. He actively participates in the program committee of different top tier conference venues in machine learning.
Fun Fact: Debanjan played cycle polo at the world's oldest polo club (the Calcutta Polo Club) when he was a kid.
Moody's Analytics: Provides financial intelligence and analytical tools supporting our clients’ growth, efficiency and risk management objectives.
Machine learning (ML) has made a significant positive impact in the finance industry in many ways. For example, it has helped in combating financial crimes and identifying fraudulent transactions. Machine learning has been a crucial tool in applications such as Know Your Customer (KYC) screening and Anti Money Laundering (AML). With an increase in AML fines by financial institutions worldwide, ever changing realm of sanctions, and greater complexity in money laundering, banks are increasing their investments in KYC and AML technologies, many of which are powered by ML. ML is revolutionizing multiple facets of this sector, especially bringing huge efficiency gains by automating various processes and assisting analysts to do their jobs more efficiently and accurately.
One of the key useful traits of ML is that it can learn from and find hidden patterns in large volumes of data. With a focus on digitization, the financial sector is producing digital data more than ever, which makes it challenging for humans to comprehend, process and make decisions. ML is enabling humans in making sense of the data, glean information from them, and make well-informed decisions. At Moody's Analytics, we are using ML and helping our clients to better manage risk and meet business and industry demands.
Reducing the False Positives without impacting the True Positives - A number of applications using ML in the regtech space rely on alerts. With strict regulatory measures and big financial implications of a wrong decision, human investigations can be time consuming and demanding. ML certainly helps in these scenarios in assisting human analysts to arrive at the right decisions. But if a ML system results in a lot of False Positives, it makes an analysts' job harder. Coming up with the right balance is an important challenge for ML in finance.
Gap between ML in basic research and education and ML in finance - Due to the regulated nature of the finance industry, we see limited exchange of ideas, data, and resources between the basic research and the finance sector, in the area of ML. There are few exceptions of course. This has led to scarcity of developing ML research that cater to the needs of the finance industry. I think more efforts must be made to decrease this gap. Otherwise, it will be increasingly challenging for the finance industry to leverage the latest ML advances.
Legacy infrastructure and databases - Many financial institutions still carry legacy infrastructure with them which makes it challenging for applying modern ML technologies and especially to integrate them. The finance industry would benefit from borrowing key ideas, culture and best practices from the tech industry when it comes to developing new infrastructure and enabling the ML professionals to innovate and make more impact. There are certainly challenges related to operationalizing ML across the industry.
Data and model governance - More data and model governance efforts need to be made in this sector. As we collect more and more data there should be more increase in the efforts to collect high quality data and the right data. Extra precautions need to be taken when ML models are involved in decisioning. Proper model governance measures and frameworks needs to be developed for different financial applications. A big challenge in this space is the lack of tools and technologies to operationalize data and model governance that are often needed for ML systems operating in this sector. More efforts should also be made in understanding bias in the data that train the models and how to make it a common practice to mitigate them in the overall process. Ensuring auditability, model and data lineage has been challenging for ML teams.
Explainability and Interpretability - Developing models which are highly accurate as well as interpretable and explainable is a big challenge. Modern deep learning models often outperform more traditional models; however, they lack explainability and interpretability. Most of the applications in finance demands explainability. Adopting the latest developments in this area and ensuring the development of interpretable models with explainable predictions have been a challenge.
- Not understanding the data well and the raw predictions made by the ML models trained on them.
- Not analyzing failed efforts and learning from them.
- Not understanding the end application and how it will be used.
- Trying complex techniques when simpler solutions might suffice.
I am really blown away by how modern ML models have been learning rich representations of text, audio, images, videos, code and so on using self-supervised learning on large amounts of data. The future is certainly multi-modal and there has been consistent progress in understanding multi-modal content through the lens of ML. I think this is going to play a crucial role in the near future and I am excited by it and looking forward to being a part of these advances.
Soumitri Kolavennu - Artificial Intelligence Leader - Enterprise Analytics & AI at U.S. Bank
Background: Soumitri Kolavennu is a SVP and head of AI research in U.S. Bank’s enterprise analytics and AI organization. He is currently focused on deep learning based NLP, vision & audio analytics, graph neural networks, sensor/knowledge fusion, time-series data with application to automation, information extraction, fraud detection and anti-money laundering in financial systems.
Previously, he held the position of Fellows Leader & Senior Fellow, while working at Honeywell International Inc. where he had worked on IoT and control systems applied to smart home, smart cities, industrial and automotive systems.
Fun Fact: Soumitri is a prolific inventor with 100+ issued U.S. patents in varied fields including control systems, Internet of Things, wireless networking, optimization, turbocharging, speech recognition, machine learning and AI. He also has around 30 publications, authored a book, book chapters and was elected member of NIST’s smart grid committee.
U.S. Bank: The largest regional bank in the United States, U.S. Bank blends its relationship teams, branches and ATM networks with digital tools that allow customers to bank when, where and how they prefer.
Machine learning and artificial intelligence have made a profound and positive impact on finance in general and banking in particular. There are many applications in banking where many factors (features) are to be considered when making a decision and ML has traditionally helped in this respect. For example, the credit score we all universally rely on is derived from a machine learning algorithm.
Over the years ML has interestingly also helped remove human bias from decisions and provided a consistent algorithmic approach to decisions. For example, in credit card/loan underwriting and mortgages, modern AI techniques can take more factors (free form text, behavioral trends, social and financial interactions) into account for decisions while also detecting fraud.
The finance and banking industry brings a lot of challenges due to the nature of the industry. First of all, it is a highly regulated industry with government oversight in many aspects. The data that is often used is very personal and identifiable data (social security numbers, bank statements, tax records, etc). Hence there is a lot of care taken to create machine learning and AI models that are private and unbiased. Many government regulations require any models to be explainable. For example, if a loan is denied, there is a fundamental need to explain why it is denied.
The data on the other hand, which may be scarce in other industries is abundant in the financial industry. (Mortgage records have to be kept for 30 years for example). The current trend for digitization of data and the explosion of more sophisticated AI/ML techniques has created a unique opportunity for the application of these advances.
One of the most common mistakes people make is to use a model or a technique without understanding the underlying working principles, advantages, and shortcomings of the model. People tend to think of AI/ML models as a ‘black box’. In finance, it is especially important to understand the model and to be able to explain its’ output. Another mistake is not comprehensively testing the model on a representative input space. Model performance, validation, inference capacities, and model monitoring (retraining intervals) are all important to consider when choosing a model.
Now is a great time to be in applied ML and AI. The techniques in AI/ML are certainly refining if not redefining many scientific disciplines. I am very excited about how all the developments that are currently underway will reshape the future.
When I first started working in NLP, I was in awe of the ability of neural networks/language models to generate a number or vector (which we now call embeddings) that represents a word, a sentence with the associated grammar, or even a paragraph. We are constantly in search of more and more appropriate and contextual embeddings.
We have advanced far beyond a “simple” embedding for a text to “multimodal” embeddings that are even more awe-inspiring to me. I am most excited and look forward to generating and playing with these new embeddings enabling more exciting applications in the future.
🤗 Thank you for joining us in this third installment of ML Director Insights. Stay tuned for more insights from ML Directors.
Big thanks to Soumitri Kolavennu, Debanjan Mahata, and Ioannis Bakagiannis for their brilliant insights and participation in this piece. We look forward to watching your continued success and will be cheering you on each step of the way. 🎉
If you're' interested in accelerating your ML roadmap with Hugging Face Experts please visit hf.co/support to learn more.