Applying AI: The Present and Future for Financial Services Firms


In this interview, Charles Elkan, Head of Machine Learning at Goldman Sachs, explains how financial services firms benefit from AI, how Goldman Sachs is applying it to their business, and highlights industry breakthroughs that are allowing AI to progress.

He delves into the mechanics of deep learning, explains the difference between superficial and subtle understanding, and talks about how NLP methods have been successful at Goldman Sachs. Finally, he describes how much further AI can realistically develop to meet the needs of financial firms.

Interview Transcript

Nitin Bharti: I’d like to welcome Charles Elkan from Goldman Sachs. He is the head of machine leaning there driven by the firm’s engineering organization which is focused on building machine learning and artificial intelligence strategy and applications, applying leading techniques to commercial opportunities within the firm. Charles, welcome to the podcast.

Charles Elkan: Thank you, Nitin. It’s an exciting time in AI applications specifically as we look at how to apply automation to the law of financial services. AI can help improve the speed, access and reliability across a number of businesses. From a customer applying for a personal loan online to a client receiving automatic notifications when they might want to (divest) from a particular industry or stock.

Nitin Bharti: Charles, how can financial services firm benefit from AI and how (is) Goldman Sachs applying it to their business? Where do you see the biggest advantages?

Charles Elkan: I think there are many opportunities. Making trading decisions, for example, whether to close a position quickly or to be willing to keep it open. Making investment decisions, for example picking stocks for a mutual fund to invest in. Or three, improving conversion for retail customers. For example, increasing the number of visitors to the Marcus Web site who opens a savings account with us.

Nitin Bharti: Can you explain the difference between superficial and subtle understanding? What is achievable now and in the foreseeable future and how will this impact the way banks apply AI?

Charles Elkan: Shallow understanding is the ability to answer questions of just one type such as counting the number of faces in a photograph. Deep understanding is the ability to answer questions using human level background knowledge such as why a person in a photograph might be feeling happy or embarrassed. Currently AI methods are only capable of shallow understanding. So, applications that need human level deep understanding are really not realistic. There is no law of nature that says that deep understanding is intrinsically impossible but new advances in basic scientific research would be needed and these advances are not predictable.

Nitin Bharti: Take me through the mechanics of deep learning. How is deep learning affecting the way information is processed?

Charles Elkan:Deep learning is the latest generation of neural networks. These were invented in the 1950’s. Research breakthroughs in the 1980’s enabled high value applications such as automatic processing of handwritten checks and real time fraud analysis for credit card transactions. New breakthroughs in the 2010’s have enabled even more impressive applications such as natural language translation and voice recognition. Financial firms are using language translation and voice recognition for business applications that are similar to those in other types of companies such as customer service.

Nitin Bharti: How have NLP methods have been successful at Goldman Sachs? Where do you see gains being made?

Charles Elkan: We’re still in the early stages of applications of deep learning at Goldman Sachs and in the world as a whole. We have a lot of opportunity to abstract structured data from text. For example, analyzing millions of loan documents to be systematic about identifying the collateral for each loan. Another example is to understand the relationships between companies such as who is a supplier for whom.
Recently a company in Michigan had a destructive fire. The company is a major supplier to the auto industry so any disruption can have an effect on production and in turn on auto sales. By applying natural language processing to relevant documents these relationships can be identified and stored in a data base which human analysts can then use to understand the implications of a news event quickly.

Nitin Bharti: Let’s discuss the mechanics of AI. Python is the most leveraged programming language being used to create AI applications. Additionally, hardware and mathematical advances are the next frontier of deep learning. Can you take us through industry breakthroughs that are allowing AI to progress?

Charles Elkan: At this point the choice of programming language is not a major factor affecting the success of AI applications. As you just noted a lot of the innovations nowadays are in hardware because deep learning is too intensive computationally for conventional computer architectures. There is a lot of innovation in algorithms and in mathematics.

For example, at Goldman Sachs we’re exploring something called quantile regression. Most regression methods predict the average of something but often it is useful to know the entire distribution of possible outcomes. Concretely, it’s useful to estimate how many shares of a given stock will be traded in the rest of today but really, we want to estimate a lengthy lower bound. (For instance,) what is the number of shares such that with 90% probability, the volume will be at least this number. Quantile regression) is a way to do this.

Nitin Bharti: Charles, are there any unsolved issues that arise repeatedly across applications? What methods should developers focus on going forward?

Charles Elkan: The fundamental unsolved problems have to be addressed by researchers and you can’t make research breakthroughs happen on demand or on a schedule. One example of a research area where we need breakthroughs is transfer learning -- which is how to modify machine learning models trained to solve one task, to solve a similar but different, new task.

The main benefit of transfer learning is that then less training data is needed for the new task. With that in mind, developers have focused on taking advantage of all the great software that is newly available and are understanding carefully what is really important for their users.

In particular, there are several excellent, free open source libraries for deep learning such as TensorFlow, MXNet and PyTorch.

Nitin Bharti: Looking to the future, how much further can AI realistically develop to meet the needs of financial firms?

Charles Elkan: There’s enormous room for new and improved applications using the current generation of AI methods. There are pretrained models for recognizing adjacent images, non-linear forecasting methods, and much more. For example, forecasting of volume quantiles for trading in many different instruments is an application that is feasible today but that we haven’t yet built everywhere that it can be leveraged commercially.

However, because current AI methods don’t achieve genuine multifaceted deep understanding of language of the real world, foreseeably, humans will still be in the loop in some way. Humans are using AI today for translation of scanning of enormous quantities of data and text. And that is allowing us to consume data at a rate the human mind just isn’t capable of. However, we cannot rely on the computers to tell us what this data means. Humans are still an essential part of the process.

Nitin Bharti: What are your thoughts on the future of AI? Do you see AI evolving into artificial super intelligence? What are the potential application advances and dangers of achieving this?

Charles Elkan: Artificial intelligence that is at the level of human intelligence isn’t impossible but it will require research breakthroughs that are not on the horizon now. We don’t know where they will come from or whether they will come. In 60 years since AI research just began we have seen repeatedly that software can perform tasks that require intelligence in humans such as playing chess and translating language but the methods that do these tasks don’t have general purpose intelligence and don’t have deep understanding. We can speculate about artificial super intelligence and obviously it can present great opportunities and great dangers but it is rather like speculating about life on other planets. We just don’t know.

Nitin Bharti: Charles, on behalf of InfoQ, thank you so much for your time today.

Charles Elkan: Thank you.

Login to InfoQ to interact with what matters most to you.

Recover your password...


Follow your favorite topics and editors

Quick overview of most important highlights in the industry and on the site.


More signal, less noise

Build your own feed by choosing topics you want to read about and editors you want to hear from.


Stay up-to-date

Set up your notifications and don't miss out on content that matters to you