INTERVIEW
AI Orders Wall Street’s Messy Data

A lot of business data is organized a bit like sticky notes – All over the place. But large language models can help turn this disparate data into organized index cards. As AI reshapes the financial landscape, firms are grappling with how to leverage the technology effectively. 

To learn about the current state of AI adoption, I spoke with Jonathan Regenstein, head of financial services AI at Snowflake. The AI data cloud company has become central in helping financial institutions manage and analyze their data. 

Regenstein works with financial institutions across banking, asset management, insurance, and payments to develop and implement AI strategies. In this interview, Regenstein shares insights on:

• The shift from structured to unstructured data in financial services

• How AI is changing decision-making processes in the industry

• The importance of robust data strategies in successful AI implementation

• The potential for AI to unlock value in previously untapped data sources

Regenstein is a former attorney and self-taught coder who has co-authored books on working with macroeconomic data using R and Python. His latest book is “A Practical Guide to Macroeconomic Data with Python" co-authored with Georgia Tech Prof. Sudheer Chava and Texas State Prof. Emmanuel Alanis. 

Hope you enjoy our chat!

This interview has been edited for clarity and length.

What’s your role at Snowflake? 

I'm the head of financial services artificial intelligence at Snowflake, which bleeds into machine learning quite a bit these days. I get to touch a lot of different things across the whole financial services industry – banking, asset management, insurance, payments, and what we call our ecosystem, which is really data providers. It's a big part of the financial world.

In this capacity, I work with what you can think of as the chief AI officer, sometimes that's the chief data officer, on how to start preparing for this world of AI and machine learning.

How are LLMs changing the financial services landscape? 

LLMs are broadening the aperture of data architecture. We're now thinking very hard about unstructured data - how to ingest it, tag it, do good metadata cataloging, and give it enough structure that it's actually usable by LLMs. We're also starting to think about how to append our unstructured data to our structured data strategy.

There's a growing recognition that there are incredible insights sitting in the world of unstructured data. Financial services firms have a long expertise in structured data, and they've been at the forefront of adopting modern data science and modern machine learning. But now, with LLMs and generative AI, we're broadening the scope to include unstructured data, which requires a pretty big rethink of data strategy.

Can you talk about data strategy and implementing AI? 

You can't just take AI and bolt it onto an organization that doesn't have a good data strategy underneath. It requires a grand vision around data strategy. For large organizations that are really dialed in on their data, this is going to be an accelerant and they're going to move very fast. Others will have to start by building their data strategy.

This involves not just collecting data, but also ensuring its quality, accessibility, and usability across the organization. It's a big undertaking, but it's necessary to fully leverage the power of AI and LLMs.

What advantage do LLMs have over traditional natural language processing? 

LLMs can give you an explanation of why they made certain decisions, which classical NLP couldn't do. For example, you can ask an LLM, "Why did you flag this transaction as potentially fraudulent?" or "Why did you give this a positive 2 score on ESG risk?" It gives us a different way of thinking about data.

This explanatory capability is really powerful, especially in areas like fraud detection or risk assessment where understanding the reasoning is crucial.

What's a key use case for LLMs in finance? 

One powerful use is analyzing changes in text over time. You could take all the 10-Ks and Qs of a public company, along with their earnings transcripts, and ask an LLM what's changing over time. What's the tone? Are they getting more optimistic or pessimistic?

You can do this at a very granular level, like focusing on their discussion of ESG risk factors across the entire Russell 3000. You can ask the LLM to create a numeric scale from negative 3 to positive 3 on whether the tone is becoming more optimistic or pessimistic. This allows for a very granular analysis that can be done at scale across many companies and documents.

How are quants using LLMs? 

Quants are finding big value in using LLMs to analyze and create numeric factors from textual data at scale. Once you've created those numeric features, they become an input to your quant model. Now you have something that's back-testable.

It's not a question of whether the model hallucinated, but whether it added anything to your model. If it did, then it's probably valuable. This ability to create numeric factors from unstructured data is really where quants are seeing the most value right now.

What's your advice on coding and understanding data in the AI age? 

I'm a big proponent of coding things up yourself. The way to really know something is to write it yourself. As you struggle with that code, you notice things in the data you maybe didn't notice before. It's a great way to really engage with the data.

Even in the AI and generative AI world, I still say code up your own stuff, create your own charts, and then bring AI in where you need to. This hands-on approach helps you understand the nuances of your data and the limitations of your models, which is crucial for effective AI implementation.

How has the conversation around AI implementation evolved? 

In 2023, which I called the year of the Cambrian explosion of these models, it was a very technologically led conversation. Now, we're starting to see much more focus on the commercial outcome we're solving for, and then working backwards into the technology we need to solve that.

It's no longer about chasing the latest technology, but about understanding what business problems we're trying to solve and how AI can help us do that more effectively.

Do you think AI will create performance disparities between funds? 

I think for large organizations that are really dialed in on their data, this is going to be an accelerant, and they're going to move very, very fast. There will be a move to emulate what those firms have done, but the emulators will have to start at the data. It's not necessarily about size, but more about having a strong data strategy.

The firms that are really succeeding right now had their data estate completely in order. They had been modernizing their data for maybe 3-5 years, so they're kind of ready for this wave and able to take advantage of these models much faster.

What do you see as the future potential of AI in financial decision-making? 

I think what we're starting to see with these multi-agent architectures is that they're going to be much more like decision-making aids going forward. I think that's like 5 years down the road. It's a long way away until we can trust the decisions that are coming out of this, but that's the direction we're heading.

This could fundamentally change how financial decisions are made, providing more data-driven insights and potentially reducing human bias in decision-making processes.

Thanks for reading!

Drop me a line if you have story ideas, research, or upcoming conferences to share. [email protected]