As AI and Machine Learning Progress at Scotiabank, Challenges Include Making the “Black Box” Models More “White Boxy”
Vishal Gossain, Vice President of Global Risk Management for Scotiabank, headquartered in Toronto, uses AI and machine learning to build predictive models that help the bank customize retail products for customers. With Scotiabank’s new focus on AI/ML, he has seen the company organization chart evolve to include a new senior EVP position of Data, Business Insights and Analytics. As decision models become more complex, they become more effective in predicting customer behavior, but also harder to explain. Seasoned managers who have relied on their own judgement to make decisions, can have difficulty with having an AI model tell them what to do. Gossain has experience managing large teams and managing retail portfolios in North America, Latin America, Europe and Asia. He recently spent a few minutes being interviewed by John P. Desmond, Editor of AI Trends, about the impact of AI on his business.
AI Trends: Thank you for taking some time with us today. From your perspective of years of experience in bank risk management and consumer credit cards and loans, what has been the impact of AI and machine learning on your business?
Vishal Gossain: Our bank has been using AI and machine learning in three different areas. The first area is underwriting, where we are trying to predict a particular behavior (risk or revenue) of the customer and accordingly tailor our products. The more accurately you can predict the behavior of a customer using artificial intelligence and machine learning, the better product you can offer to the customer. In addition, we have begun to use AI and machine learning in new areas such as cyber risk, Fraud and AML.
Second, we use AI and machine learning to create a better customer experience. For example, we use chatbots through our contact and call centers, and automated voice analysis in our contact center phone systems to enable sentiment analysis. When we are onboarding a customer, we need to process a lot of documents. If we can automate the document processing, it becomes easier for the customer and easier for us.
The last category is bank process automation (such as for payroll processing) where efficiencies can be gained.
What does the future hold for AI and machine learning in the banking business?
Machine learning is definitely the future for the banking business. The more sophisticated the model, the better we can predict the customer’s behavior and future actions which, in turn, better enables the bank to deliver the right products. For example, if a young person wants to buy a house, we can offer instantaneous mortgage and/or a credit card pre-approval to him/her, and have it seamlessly delivered with the right limits. If we can reduce the paperwork required, we have a happy customer who wants to bank with us for a long time. So predicting a customer’s behavior in a seamless way will become more important.
Second, we are using AI more to predict abnormal patterns as they relate to money laundering and credit fraud. The AI can do a better job because it is able to ingest much more data and see patterns that humans may not be able to.
Also, the AI can self-learn. That is important because most of the models we have built are static, which we have to update over time. Sometimes, it can take three to five years to update them. By that time, the consumer behavior has shifted. The AI allows us to stay abreast with shifting consumer behavior with the self-learning models. The AI will bring a foundational change to banking, just as the internet has brought profound change for knowledge sharing and storage. Of course, enhancing the customer experience and process automation will continue to be important areas.
You have operated around the world, managing portfolios in North America, Latin America, Europe and Asia. You must be tuned into geopolitics. On the topic of immigration, what lessons do you see the Canadian immigration system having for the US?
The Canadian immigration system has some unique features which the US might benefit from. I have gone through the immigration systems in both countries, so I feel I have relevant experience. The Canadian immigration system is more immigrant-friendly, for example, in that they have a permanent residence process that is faster than that in the US..
Most applications nowadays I hear are getting done in almost six months, end to end. Second, they have a good experience with a points-based-system that is attracting professionals with the right mix of education and experience. They settle in Canada and help to fill the skills gap. It’s a very good strategy to get Canada the right skill sets for the future.
Third, going though the US process myself, when I landed I was given F1 as a student. After graduation, I transferred to the work visa, H 1-B. The problem was when my wife came, I asked for a dependent visa. She was equally educated compared to me, but from the UK. That became a challenge. She had to take an F1 visa. When I applied for my permanent residence, we had to go through some audits, which took three of four years to come through. So by the time things turned around, I was in the US for almost a decade and still waiting for my permanent residence.
And that creates an environment of instability because I had to renew the visa yearly which became a challenge to kind of settle down, buy a house and make investments. I had to renew my work visa every year, and I was never sure it would be renewed.
On the other hand, when I applied to Canada from outside, my wife and I could apply together. Our application was processed very smoothly and efficiently within six months. We were able to move to Canada, settle down, buy a house and start raising a family. So we found the Canadian system to be very fast and practical, and more tuned to what families need than the system in the US.
I am sure you look at a lot of data. Is the bank taking any steps to protect against data bias and data sets, especially for consumer credit worthy?
Bias is an important issue for all banks. Scotiabank is taking a leadership position in Canada by for example, helping to organize conferences on ethics and bias. We were one of the first banks to release principles on ethics and bias. As we move towards sophisticated AI, some criticize it as becoming more black box [unexplainable AI].These principles become very important in practical applications of AI/ML.
We tackle bias in multiple ways. At a high level, we make sure that the data we use to train our models is unbiased. We remove variables which can tend to bias the model by sex, gender and religion, for example. We train our models over a longer period of time, to make sure the model is not biased towards certain temporal events. We ensure that the final model outputs are not biased towards any characteristic of a consumer that we don’t consider as “fair” to measure for credit worthiness. We also have independent reviews of the models.
However one important factor here, which is that every decision you make in life is generally biased. One very good study in the US showed that some judges have longer jail sentences after their favorite football team had lost a match. [See, Football Team Losses Can Affect Prison Sentences, in the PacificStandard.] You cannot be completely unbiased.
Our job is to establish the acceptable bias levels and to conform to that as our models become more accurate in their predictions and we become more “black boxy.” That’s the approach we are taking.
Can you tell us what key AI related tools, technology and software is important to your work?
I will answer that question in different pillars. We use supervised techniques, unsupervised techniques and reinforcement learning; we use all three techniques. We are also exploring new fields such as Bayesian analytics. We are also heavily investing in data encryption techniques such as FHE, Secure Enclaves and MPC.
It’s important we have the right software languages to be able to use these tools; we are language agnostic. We use SAS, R and Python heavily. Which ones get used depends on the skill set of the involved data scientist.
For the data component, no model can be built which cannot have the right amount of data to feed into it. And the data needs to be stored; it’s important to have the right storage. We use Hadoop, and we are now moving towards the cloud. We have the Google Cloud Platform as our primary cloud strategy. We also leverage Microsoft Azure in some areas.
Once we get the data, we use software to build our models, then the hardware becomes important. We are increasingly going to the use of GPUs. We have been using NVIDIA DGX boxes [purpose-built system for data scientists], and have leveraged cloud based GPUs as well.
Once the model is built, then we have to deploy it. For deployment, the bank uses some very robust systems. You can imagine the amount of data that we have. The systems we use are tailor made for very specific purposes. For example, we use FICO systems for the retail products and ORACLE for Money Laundering. We are increasingly leveraging new and fast APIs to deploy our models.
To develop the models, we are increasingly moving towards AGILE approaches instead of the traditional waterfall approach. For example, my team is currently running 6 AGILE Labs with cross functional teams to deliver our new AML [anti-money laundering]models.
Lastly, we need to be able to leverage tools for the purpose of effective governance. We need to ensure our models are well-documented. We are investigating automated documentation procedures (Natural Language Processing)
What would you say are the challenges using and implementing AI and machine learning in your work?
I would say we have four major challenges in our implementation. The first is data. The bank has about 25 million customers. The challenge is the data for these customers is very scattered in multiple systems. It’s very difficult to unify the data and get a customer-centric view. Also, we need a unified view across all transaction and all products, so we can use that to help a customer. We have issues of data privacy. I am sure you have heard about the Capital One breach [Ed. Note: See The Capital One breach is more complicated than it looks The Verge] We have to move forward without putting our customer data or customers at risk. That’s a big challenge. The second challenge is technology. Banks face a conundrum in which they have relied on very robust technological systems to service customers. But these systems have not evolved as fast as AI and machine learning. So, I can build a very sophisticated AI models, but today I am not able to implement that in my retail systems very efficiently at all. You can get the best person in the world, but if your deployment system cannot ingest that model, your model is pretty much useless.
The third challenge is people. We have a new breed of data scientists that have come about, and we have seasoned folks who use a lot of judgement to run their business. Some of them struggle with a black box model telling them what to do. It may not be as intuitive for them as consumer behavior. What they are used to seeing has shifted. So we have a heavy debate between some more conservative folks and the new data scientists. The culture has to evolve to become more data-focused and to make decisions based on data rather than on judgement.
The fourth challenge is explainability as we go towards AI and machine learning. The models are becoming more accurate and are also increasingly becoming a black box. Therefore, it becomes very difficult to understand why the model made a decision, compared to more vanilla techniques of logistic regression, which has been used for 30 years to build models. The use of new sophisticated models such as deep learning is a more recent phenomenon. In the regression model, we know exactly what variable contributed to the model. You know exactly why a customer scored worse or better. With the new techniques, it’s a challenge to be “open boxy.” So explainability is important; it’s a difficult, but not an insurmountable challenge to mathematics.
Once you overcome the challenges, the rewards are gratifying. For example, Scotiabank was one of the first bank to put a deep learning model in production for collections, which has generated $10+ MM in incremental annualized income for the bank since launch. The bank has also used explainable AI for adjudication for automotive and mortgage portfolios with expected returns exceeding $20 MM in annualized risk adjusted returns.
What has been the effect of AI on your organization charts? What new job titles have you seen introduced in the past few years?
We have seen significant changes in our organization charts. We now have a new EVP position called Data, Business Insights and Analytics, which did not exist before. We have a new analytics community of practice, which did not exist before. So, analytics in general and AI and machine learning in particular is leading to the creation of new positions which are more data-science oriented. Quants are getting a seat at the table. [Ed. Note: Quants use computers to decide what to buy or sell.]
The models are taking on more decision-making that a human was doing earlier. That has really changed the organization chart.
What advice would you have for young, early career people interested in a career in AI? What should they study? What job experience should they go for?
While studying is very important, that’s a stepping stone for them to go through their careers. Many institutions offer some really good courses on AI and machine learning, and quantitative analysis. Some programs include MIT, UC Berkeley, Queens, Western, UBC etc. They should take electives in areas of programming like Python, where it’s helpful to have some hands-on experience. Certifications in cloud computing from cloud providers can be helpful. Any quantitative discipline, with a bachelor’s or post-graduate degree, a master’s or a PHD, which allows them to apply AI and machine learning techniques in any area, is definitely an asset.
Those who are more seasoned professionals can definitely learn AI on the job. There are many self-learning courses on the market right now. I strongly believe that a strong programmer and a strong scientist, a person with strong analytical skills, can pick it up. Education always helps.
Thank you so much for your time. Anything you would like to add?
Thank you for the opportunity. I am very passionate about AI and machine learning in general. This field has been underutilized in the past because of the lack of data storage and computing power. But over the last decade or so, as computing power has become very cheap and as data storage has significantly expanded, AI and machine learning is becoming much more realistic. Banks like Scotiabank have been applying it very successfully in the past two or three years to practical problems. For example, we built the collections model using these new networks, which has already lead to about $10 million a year in savings and increased income. That was not possible before.
Also, our auto lending and mortgage lending have recently utilized models which are expected to lead to multimillion dollar returns incrementally. More recently in anti-money laundering, we have been working on a consortium project. The idea is to get data from different banks together, to leverage data encryption and subsequently create an interbank AI model. That protects the customers’ data and increases the bank’s probability of detecting money laundering. So the sky’s the limit as we take the potential to a new level.
I also have to mention that Canada has invested a lot in the AI field. Toronto is the fourth largest city in North America and has a lot of tech talent. It’s a great city to live in.
Learn more at Scotiabank.