Virtually all of the major hedge funds have been building AI into their workflows.

• Citadel has an AI assistant

• Man Group has AlphaGPT

• Bridgewater has an AI fund generating “unique alpha”

• AQR’s Asness says AI has become “annoyingly better”

There are more but you get the idea.

The science-fiction way to think about AI in finance is that all these funds are building Minority Report systems, where soothsayers predict the next market event. The reality is more prosaic. They’re building systems that amplify their investing point of view. Like, as we discussed last week, how Bridgewater worked with Thinking Machines to create a model that better combed through relevant investment news and research.

Investors can’t build everything themselves (nor do they have the budget of Bridgewater) so there's an emerging niche of companies building AI & investing functionality by alums of larger firms. From Business Insider:

• Ian McInnis, a former Bridgewater analyst, founded WithAI, a Y Combinator-backed startup helping funds use LLMs to process information and plug AI into the investment process.

• Jaime Villa, previously a macro researcher at Schonfeld and Citadel Securities, cofounded Macro Technologies, which is trying to automate repeatable work done by macro analysts.

• Cameron McKendrick, a former Jain Global executive, leads Serona Data, which looks for investment signals in healthcare data.

While the BI story suggests these upstarts could eventually mean fewer analyst jobs, I don’t subscribe to this idea. None of the folks I’ve talked to have said AI is leading to fewer jobs. AI creates more work. Cheaper tools mean more stocks to screen, more signals to test, more AI-generated claims to verify and on and on.

SPONSOR

The Brent-above-$80 headline was in all the feeds. The drone strike on a Chevron-chartered tanker off Novorossiysk, the shale-recovery tech deal with ZL Chemicals, and California’s gas-price fight were in the maritime and energy trade press — the physical signal, scattered across outlets no single feed carries.

Seltz returns all of it, raw. Full articles, first-person reporting, within the hour of publication. No summaries, no sentiment scores – your models do the distillation better than any vendor, on input nothing has influenced.

Search any name across any day in history the same way. That’s how teams build and test a signal before they trade it live. Backed by B Capital and Speedinvest.

Want the full picture on every name you trade? Book a demo — bring 10 questions your desk actually asks, and we’ll run them live.

The FCA Takes Up Chatbot Financial Advice

While the SEC has been pretty quiet on AI financial advice, the UK’s FCA released a report this week saying consumers are already using general-purpose AI tools for money questions without understanding that regulated-advice protections may not apply.

The FCA surveyed 5,026 UK retail-finance consumers in April. Sixty-seven percent said they use AI; 16% said they had used it for financial tasks. About 26% said tools like ChatGPT provide reliable financial information or advice. Only 40% understood there would be no formal route for recourse if that advice went wrong.

The FCA is not proposing chatbot rules yet. It wants to know whether general-purpose AI tools are already giving advice-like support outside the regulatory perimeter, and whether existing rules on advice, promotions and arranging still work when the first intermediary is a chat interface.

AI as Deal Runner

So far, AI in finance has mostly lived in the back office: KYC checks, code help, earnings call summaries, etc. But this is the first I’ve heard of AI running a deal. From the WSJ:

When the buyout giant CVC put a Greek e-commerce business called Skroutz up for sale earlier this year, it didn’t use investment bankers to run the process. It used artificial intelligence.

Prospective buyers of the company were sent a link to a data portal that acted like an investment memo. A chatbot “analyst” acted like the banker, answering questions on financials and due diligence or prompting interested parties to get in touch with the management team for a further discussion.

The CEO of Skroutz spearheaded the company’s AI exit, so this is an unusual test case, but I suspect not the last one we’ll see. Blackstone agreed to buy the Greek e-commerce business in May.

Your AI Avatar Money Manager

I was talking with a friend who said they liked the “vibe” of Claude over ChatGPT, which made me think we’re on our way to having “relationships” with our favorite chatbot. This would have sounded weird and creepy, maybe, two years ago? Less so now.

Companies are leaning into avatars.

OCBC, a Singaporean bank, is rolling out two avatars, Wendy and Wayne, for wealth clients with more than S$1.5 million in the bank. They can answer questions about markets, portfolio news and customer holdings, pulling from real-time market data, OCBC research, client portfolios, transactions and behavioral signals.

The avatars pull from approved data, hard rules limit what they can answer, and separate agents handle different types of requests. If Wendy or Wayne reaches the edge of what they are allowed to do, the conversation goes back to a human relationship manager.

ROUNDUP

What Else I’m Reading

• Don’t rely on AI for personal finance advice, study finds | CNBC

• Singaporean regulator outlines safety guardrails for financial AI agents | Yahoo

• Using AI to Let History Speak About Bank Runs | NYFed

• Databento raises $97M Series B led by NEA | Utah Business

• This just came out, so I haven’t had a chance to listen yet: Man Group on Odds Lots:

Share AI Street

This Week in AI Street

Upgrade to paid

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

CALENDAR

Upcoming AI + Finance Conferences

• Global Agentic Regulator Hackathon – July 8–September 18 • Online

  Hackathon for building agentic tools for public authorities, focused on AI advice, market manipulation, payments, decentralized infrastructure, and scams.

• ARPM Quant Bootcamp – July 13–16 • New York

  Four-day quant workshop from ARPM, focused on portfolio construction, risk, machine learning and the plumbing behind systematic investing.

• Artificial Intelligence for Financial Services – July 23–24 • Cambridge, MA

  MIT executive-education workshop on AI/ML applications in financial services, emphasizing strategy, use cases, and implementation considerations.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• Connect with me on LinkedIn, or

• Send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

Travelers is the latest large company to train an AI model on its own data and says it beat commercially available systems.

The insurer trained TravelersLLM on millions of internal documents for underwriting, research and model development, according to a June 30 press release. The company, founded in 1864, said its long history and proprietary data helped improve the model’s precision.

“In testing against tens of thousands of insurance-related questions, [TravelersLLM] consistently outperformed commercially available AI models, delivering higher-quality results at lower cost and with greater speed.”

Emphasis mine.

Call them house models: AI systems trained, tuned or evaluated against a company’s own data, workflows and judgment, so not a generic model pointed at internal documents, but one that “learns” what a company does.

Travelers says AI, automation and analytics are contributing to operating leverage, expense-ratio improvement and underwriting profitability. Its shares have climbed about 19% this year, roughly double the Dow Jones Industrial Average, in which Travelers is a component.

I wrote back in May that legacy companies have an advantage over upstarts because they house so much data.

AI models are only as good as the data they’re trained on. Hard-to-replicate, legacy data is more valuable in the age of AI. And legacy companies are generally the ones with the legacy data. Many corporations are sitting on valuable intellectual property and, I suspect, don’t even know it.

The companies that do know the value of their data and the power of AI are turning their private records, workflows and judgment into models no other company can replicate because no other company has the same raw material.

The current AI boom stems from the breakthrough and flexibility of transformer models trained on language. But the benefits of this architecture extend beyond text. You can train these models on other types of data — the weather, grocery sales, bond portfolios — as long as you have enough high-quality data.

This is a topic I’ve been writing a lot about:

The conversation around companies adopting AI has focused on which model they’re using or what company they’re partnering with, which makes sense. You don’t need to fine-tune a model to write better emails. But if you’re in the business of selling property & casualty insurance, you’d much rather have a model trained on P&C claims rather than one trained on Reddit posts.

Moreover, if you’re an equities trader, you’d find more use for a model trained on prices, filings, broker messages, order flow and portfolio data than a generic chatbot. (By the way, this is part of the reason firms like Jane Street, HRT and XTX are spending billions on building their own data centers. It takes lots of computing power to turn proprietary data into better predictions, prices or trading signals.)

Microsoft CEO Satya Nadella has recently been making a version of the house-model argument: AI run by a handful of frontier labs is too narrow for the technology.

“There should be as many models in the world as firms in the world,” Nadella said in an interview last month:

Nadella’s argument is that a company is not just a buyer of software. It is a learning system. Its edge comes from its data, context, workflows, evals, traces, judgment and operating history. Frontier models can be part of that system, but if the company never builds its own learning loop around them, it is not compounding its own intelligence.

For the last couple of years, analysts and investors have been asking companies: how are you using AI? That is a good starting point. But Nadella’s argument points to the better question: how is the company using AI to turn its own data, workflows and operating history into a competitive advantage?

Paid subscribers can access a sourced working list of companies and organizations that have publicly disclosed proprietary or domain-specific model efforts.

Not a paid subscriber? Consider expensing AI Street. If it helps you evaluate one AI initiative, pressure-test one vendor pitch or understand one real deployment constraint earlier, it more than pays for itself. I put together a short template you can send your manager.

Read more

The way hedge funds use AI fits into two broad buckets.

One way is to train an AI model on market data. In other words, instead of training one on language, you gather all the financial market data you can get your hands on, train the model—not to predict the next word—but the next market event. This is hard and expensive. Firms like Hudson River Trading and XTX are spending billions of dollars to build their own data centers. (And in the case of HRT, putting out movie-quality videos of their GPU clusters deep inside a Norwegian mountain.)

The second way is leveraging large language models and building skills and systems around them. This is what Man Group has done, I’m oversimplifying a bit, with its AlphaGPT. For more on that, check out my interview with Man Group’s Ziang Fang:

While LLMs are solving 80-year-old unsolved math problems, they sometimes struggle with the nuances of what an investor cares about. You can try to prompt it: “You’re a macro investor at a multi-strat hedge fund.” But that only gets you so far.

To teach a model more of that investor judgment, Bridgewater fine-tuned an open-weight model on examples labeled and reviewed by its investment experts, according to a statement this week from the hedge fund and Thinking Machines, the AI startup run by former OpenAI CTO Mira Murati. The goal was to teach the model what Bridgewater investors would consider relevant.

For example, virtually no investor reads all of a company’s disclosures. They’re hundreds of pages long. Humans skip boilerplate language. LLMs often treat that text equally so even with a very large context window — basically AI’s working memory — they still get lost. It’s just too much information.

Bridgewater and Thinking Machines turned that skipping into training data. They trained the model on six versions of the same basic problem: deciding whether a financial article, central bank document or research report was relevant, and where the useful part of a document or email ended and the boilerplate began. When the model disagreed with the original labels, those examples were sent back to Bridgewater experts to make the distinction clear.

The result was a model that beat the frontier models Bridgewater tested. With expert prompts, GPT, Claude and Gemini got into the mid-to-high 70s on the six tasks. Bridgewater’s fine-tuned model reached 84.7% average accuracy, which the firm said was good enough for daily use, and cost 13.8 times less per task to run.

When you’re running AI scale, costs add up quickly. We talked about how AI costs can balloon because of untamed AI last week. Firms are spending tens of millions of dollars on research tasks that may not be necessary.

A person familiar with Bridgewater’s AI efforts said this research on improving efficiency is a small piece of the hedge fund’s broader goal of building a complete “AI investor” that can match and exceed their human counterparts.

The system combines Bridgewater’s proprietary causal time-series machine learning with custom LLM systems, including specialized models and frontier models adapted with the firm’s own harnesses and agent-style workflows, the person said. Bridgewater sees specialized models as an important part of that effort.

Bridgewater formed its Artificial Investment Associate (AIA) Labs division in 2023 to build investment systems that combine large language models, machine learning, and reasoning tools. Its AIA Labs Macro Strategy, live since December 2023, is generating market-beating returns and now has more than $4.5 billion in assets.

Lawmakers Press SEC on AI Agents in Retail Trading

Democratic lawmakers are pressing the SEC on the risks posed by agentic trading, including whether brokerages, AI providers or investors themselves are liable when a retail investor lets an AI agent give advice, place trades or take other actions inside an account.

In a June 23 letter to SEC Chair Paul Atkins, the lawmakers asked whether the agency has discussed AI-agent use with broker-dealers, whether it plans to issue guidance, and what guardrails brokerages should have in place before letting agents act for customers, according to Advisor Hub.

Robinhood and Public allow retail investors to connect AI agents to their brokerage accounts: Robinhood has promoted AI stock-trading tools, while Public advertises agents that can run covered-call strategies and same-day options trades tied to intraday S&P 500 moves.

Across the pond:

BoE Says Existing Rules Weren’t Built for AI Agents

Bank of England Deputy Governor Sarah Breeden said current supervisory frameworks were not designed for autonomous agents in payments and trading, where human approval for every action may not be realistic. She pointed to scenario analysis, AI-enabled monitoring, digital twins, enhanced recovery for core systems, and kill switches or circuit breakers for faulty trading models.

Share AI Street

Fundraising and Partnership News

• Arca, which uses AI to support financial advisors, has raised $64 million across seed and Series A funding rounds. General Catalyst led the company’s $48.5 million Series A, with backing from Index Ventures and Venrock. Arca said it manages more than $1 billion in client assets. More on Arca here:

• MDOTM raised $27 million to expand Sphere, its AI platform for asset and wealth managers. The company says Sphere supports more than $100 billion across 60-plus financial institutions and is used for investment insights, portfolio construction, rebalancing and client reporting at scale. For more on MDOTM, check out my interview with Peter Zangari, a partner at the firm.

• LinqAlpha raised $22 million in Series A to expand its AI-agent platform for institutional public-markets research. The company says more than 70 financial institutions use its tools, including buy-side clients managing over $5 trillion, and plans to use the funding for hiring, data integrations and broader deployment across asset classes.

• FactSet is partnering with and investing in TIFIN.AI to add AI agents to wealth-management workflows. The first tools focus on meeting prep and book-of-business intelligence, with the pitch centered on helping advisors personalize service at scale while keeping client data inside FactSet’s infrastructure.

What Else I’m Reading

• Millennium Builds AI Lab in Push for Cutting-Edge Products | BBG

• Now hedge fund Qube Research and Technologies is building out a “lab” too | eFinancialCareers

• Wall Street’s AI Race Is Fueling New Fears of Crowded Trading | BBG

• Hedge funds offer bumper pay to lure AI talent: ‘Million-dollar packages are not far’ | Financial News

• Is AI Good at Stock-Market Timing? A New Study Casts Doubt | WSJ

Upgrade to paid

Help Benchmark AI Adoption in Finance

It’s hard to get reliable data on how AI is actually being used by financial firms.

Neudata, a data scouting service for the financial services industry, is running a short survey on how the sector works with alternative data, market data and AI.

It takes about 5 minutes. Participants gain access to the results from data buyers as well as vendors, plus a chance to win a $500 gift card.

Topics include data budgets, in-demand data categories, and the AI models used to process that data.

This Week in AI Street

CALENDAR

Upcoming AI + Finance Conferences

• The Fin AI Symposium 2026 – July 3 • San Diego

  Industry symposium focused on applied AI in finance, with practitioners sharing tooling, workflows, and deployment lessons.

• ARPM Quant Bootcamp – July 13–16 • New York

  Four-day quant workshop from ARPM, focused on portfolio construction, risk, machine learning and the plumbing behind systematic investing.

• Artificial Intelligence for Financial Services – July 23–24 • Cambridge, MA

  MIT executive-education workshop on AI/ML applications in financial services, emphasizing strategy, use cases, and implementation considerations.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• Connect with me on LinkedIn, or

• Send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

I thought the technology had a real shot at becoming widespread across Wall Street. Or, if not, it was going to be a great big dud worth covering.

Early on, I spoke to Snowflake’s Jonathan Regenstein, who was also optimistic that AI would have a big impact on how the investing world works. I interviewed him in the fall of 2024, and much of that conversation focused on a basic question: how do you even use this technology for investing? Does it make sense to? What about hallucinations?

When we spoke again recently, Regenstein said the skepticism has largely disappeared, at least among the most sophisticated funds and asset managers. The question now is how to build efficient processes.

Regenstein is a former attorney and self-taught coder who now leads wealth and asset management at Snowflake, where he works with banks, hedge funds and asset managers building AI systems on top of their own data. Snowflake does not build frontier models. It works with providers and platforms including OpenAI, Anthropic, Mistral, Meta and Hugging Face, which gives Regenstein a view into how firms are putting different models into production and where they still struggle.

He says investors are using AI in three main ways:

1. Chat: using the model conversationally to test a hypothesis or get a first read on what’s happening.

2. Text analytics: “Go examine this 10-K for me” — or, at asset-manager scale, turning years of filings, research papers, and sell-side research into an “engine of insights.”

3. Coding assistant: “Write the code for me” — helping quants and investors code up hypotheses, reproduce papers, build dashboards, or call validated models.

Regenstein has also published a new book, Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls, coauthored with Thársis Souza PhD, whom I’ve previously interviewed on my podcast here. The book is for non-experts who want to understand LLMs well enough to use them: builders, go-to-market salespeople and strategy leaders. I’m looking forward to getting my copy in the mail!

This interview has been edited for clarity and length.

Matt: Tell me about your role.

Jonathan: I work at Snowflake, which is a super interesting place to work right now because we don’t build frontier models. We partner with OpenAI, Anthropic, Mistral, Meta and Hugging Face. We partner with all of them.

As this world evolves, we’re sitting at the center of it, even though we’re not really as high up in all the headlines. We see it all, deal with it all, work with them all and optimize for all of them. I get to see some trends emerging around who uses which tool for what.

There are three different ways that LLMs are being deployed now. One is the chatbot. People like the chatbot. The second is text analytics: “Go examine this 10-K for me.” The third is a coding assistant: “Write the code for me.”

Those are three very different things. The models can do all of them, but they’re not the same thing at all. They’re not for the same person or the same use case. They’re really totally different. It’s starting to break down along those lines, and there’s a lot underneath each one.

How has AI adoption in finance evolved since we first spoke in the fall of 2024?

Jonathan: There have been a few big shifts. Back in 2024, I think you and I were both believers in this. But at that time, with some exceptions, if you had gone into a room full of quants or portfolio managers, I think you would have gotten a lot of crossed arms: “I’ve been doing this a long time. I understand, and I’m sure AI is helpful, but it’s not going to revolutionize how I do my job.”

We’ve now come to a place where there is no more skepticism, certainly at the most sophisticated shops and asset managers. The question is: “How can we optimize this?”

It’s becoming a recruiting and retention tool for funds and asset managers: “You should come work here as a quant. Why? The AI harness we’re going to give you is better than you’re going to get anywhere else. If you don’t have that, you cannot win in this world.”

Once you’ve come to a fund or asset manager and they’ve built you that harness, you don’t want to leave because you can’t take it with you. It’s intellectual property. It’s becoming a recruiting and retention tool. It’s a complete shift.

How are firms building governance into those AI harnesses and testing whether they actually work?

Jonathan: This is an area where I work directly. Snowflake doesn’t build its own models, so what are we doing? We’re trying to give you the governance architecture that sits underneath them.

When you ask for data, we need to make sure that what you asked for is what you actually got and that it is underpinning your answers. That’s a ton of the work we’re doing right now: validating your context and making it scalable to monitor and evaluate what you’re doing.

Skills are becoming a huge thing. What skills do is bring determinism, as much as possible, to the way large language models work because you’re giving them a recipe: “Do this. Follow these steps. Bang, bang, bang, bang, bang.”

One of those recipes can be a quantitative model that I’ve built somewhere. Large language models don’t run a stress test or even run a regression for you. They could write the code to run the regression, but I don’t want the model to write the code on the fly. I want it to call the model I’ve already built and validated.

Let’s say I have a phenomenal modeling team and they’ve built these models. We’ve validated them and have them in our ecosystem. This is something you can do with Snowflake: You can store all your models in one place.

Now the LLM needs to choose the right model, not run the model. Where we have a whole model-testing framework that we’ve validated over the last 20 years, the LLM can call the right model. That’s what I need to get right more than I need to get an LLM to build a model on the fly for me.

You can do that for research. Part of the research for your model garden can be having an LLM build a model for you. But when it’s actually time to make a decision, the LLM needs to call a validated, back-tested model.

A lot of the focus is on the model, but much of the work involves the surrounding environment and calling the right document or model. How do you ensure the system calls the right thing? Do you use knowledge graphs?

Read more

AI to Push Trading Into Thin Markets

I have a running theory that AI is going to push more trading into markets that have historically been thinly traded.

As we’ve covered a lot around here, AI is making investment research cheaper. Investors can track more companies and test more ideas. Tasks that used to require more people, time and money are getting easier to run at scale. I’ve heard versions of this in multiple conversations with investors using AI.

This Bloomberg story has more reporting on this trend:

Advances in artificial intelligence are leveling the field for fund managers, making it easier for boutique firms to compete with big macro and bond investors.

From digesting speeches in multiple languages and crunching global inflation numbers, to tracking company filings and the tone of investment committee discussions, AI is picking up much of the work once carried out by teams of analysts, according to five executives who have recently set up their own shops.

“Technology has changed the economics of building an investment firm,” said Dharmesh Maniyar, a machine-learning PhD who founded his second fund, MQT Asset Management, late last year.

Now if everyone is getting access to cheaper investment tools, this should lead to more trading. And recent research supports this trend.

When Italy banned ChatGPT in March 2023, retail investors made fewer trades in unfamiliar assets, shifted back toward popular names, and held portfolios that looked more concentrated and alike.

And AI-driven portfolios trade much more often than their human active-manager counterparts. According to a 2024 IMF report, AI-driven ETFs already turn over holdings about once a month, versus less than once a year for typical active equity ETFs.

Volume has been above historical norms since COVID. The pandemic reset U.S. stock-trading volume higher. A quiet pre-2020 day was closer to 7 billion shares. Since then, average daily volume has generally stayed above 10 billion shares, rising to 12.2 billion in 2024, according to the Cboe.

TL;DR

I’m having a hard time seeing trading not marching higher and permeating into more markets given current AI trading trends. It didn’t make a lot of economic sense to spend limited human time researching smaller markets. Now it does.

The Cost of Runaway AI Agents

One caveat about the BBG article above: I think it underplays how hard it is to get AI working accurately at scale. It’s one thing to summarize a central bank speech, but quite another to pull accurate numbers out of hundreds of pages of regulatory docs. This is also a topic we’ve discussed a lot.

And even when you can receive accurate, helpful responses, AI research isn’t cheap. I’ve heard a lot recently about the exploding costs of compute from financial services firms. Uber reportedly blew through its annual AI coding-tools budget in a few months.

I’ve had the same experience trying to create my own agents with not a lot success.

Milos Maricic, a consultant who advises sovereign wealth funds on AI, said he had a hedge fund client spending roughly $24 million a year on LLMs, not realizing that about a third of that spend was just wasteful.

One of the fund’s agents kept calling a third-party sourced that went dark, but the agent kept getting timeouts or 429 errors and treating them as a reason to try again. It retried for weeks.

The Specification

Saving a hedge fund $8.7 million on AI spend in a week while doubling signal detection efficiency

This project grew organically out of a mandate focused on building an AI-driven alpha-seeking research operation within a client hedge fund…

Read more

17 days ago · 4 likes · 1 comment · Milos Maricic

Maricic said his team cut those costs by a third without hurting research quality, mostly by tracing spend to specific teams and workflows, stopping runaway agents and caching repeated prompts.

As I mentioned last week, I think managing compute expenses, or maybe we can call it something fun like “token master,” is an emerging job.

Computing costs are exploding across companies. Who manages that? Each individual department head? How do you decide which use cases to spend those resources on? Those seem like hard questions to me.

Share AI Street

The Growing Market for Making Markets in Compute

A year ago, trading “compute” sounded weird. `Like, you’re going to turn GPU capacity into something you can trade, like oil or electricity?’

Today not so weird. Multiple firms are trying to build the financial infrastructure of the emerging commodity. DRW’s Don Wilson is backing Silicon Data’s GPU pricing index, which feeds CME’s planned futures market.

This week, a16z announced its first bet on this emerging asset class by leading a $33 million seed round for Ornn, which publishes a rival index that powers ICE's planned futures market.

As a16z’s Ali Yahya put it to Bloomberg: “This asset class is very immature at the moment, but will professionalize over time.”

This Week in AI Street

ROUNDUP

What Else I’m Reading

• Secretive Wall Street Powerhouse Jane Street Seizes the AI Spotlight | WSJ

• XTX Markets looks for AI gems after striking gold with Anthropic | FN

• Alphabet Shares Drop After Second AI Star Departs for Rival | BBG

• AI Can Model but Can’t Make the Next Rainmaker | WSJ

• Starling rolls out AI-powered romance scam detection feature | Finextra

Upgrade to paid

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives, and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

CALENDAR

Upcoming AI + Finance Conferences

• 32nd Computing in Economics and Finance – June 29 • Venice

  Academic conference on computational methods in economics/finance (numerical methods, simulation, macro/asset pricing, econometrics).

• Risk Live Europe 2026 – June 29–July 1 • London

  Risk.net conference for banks, buy-side firms and regulators, with sessions on AI, automation, model risk and operational risk.

• The Fin AI Symposium 2026 – July 3 • San Diego

  Industry symposium focused on applied AI in finance, with practitioners sharing tooling, workflows, and deployment lessons.

• ARPM Quant Bootcamp – July 13–16 • New York

  Four-day quant workshop from ARPM, focused on portfolio construction, risk, machine learning and the plumbing behind systematic investing.

• Artificial Intelligence for Financial Services – July 23–24 • Cambridge, MA

  MIT executive-education workshop on AI/ML applications in financial services, emphasizing strategy, use cases, and implementation considerations.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

AI doesn’t “read” text. Models turn words into numbers, perform calculations on those numbers, and then turn them back into words. So if AI sees the world at all, it sees it a little like Neo in The Matrix.

A key breakthrough behind LLMs came from a 2017 Google paper on machine translation. That transformer architecture helped models capture more nuance in text by letting every word weigh its relationship to every other word in a sentence or document.

Prior to transformers, machines struggled to incorporate context. A bank can lend money. It can distribute blood. Or it can be the place you sit by a river. Transformers pick up on this nuance because they do not treat each word’s meaning as fixed but rely on context.

One important kind of numerical representation AI models use is called an embedding. An embedding is a little like GPS coordinates for text: a location in mathematical space. Those coordinates can be compared, searched, clustered and fed into statistical models. One classic illustration: take the embedding for “king,” subtract “man,” add “woman,” and you land near “queen.”

Creating Embeddings from News

Embeddings are not limited to text. You can create them out of almost any messy data. And once you have them, you can ask more complex questions beyond king minus man, including questions where you don’t already know what relationships you’re looking for. In other words, embeddings can surface relationships that are hard to see with ordinary categories, keyword searches or traditional metrics.

In one paper, researchers uncovered relationships missed by traditional financial metrics after creating embeddings tied to portfolio data. I wrote about that for the Chicago Booth Review here as well as a newsletter version here:

A new paper, the Inefficient Pricing of News, takes that same framework — embeddings of text — and points it at financial news.

The researchers, including AQR’s Bryan Kelly, turn news articles into embeddings and test whether those representations contain information the market has not fully priced. Kelly is also a professor at Yale, and the paper is academic research, not an AQR report. His co-authors are Antoine Didisheim and Hanqing Tian of the University of Melbourne, and Yale’s Mo Pourmohammadi. I reached out to the authors for comment but did not hear back by publication time.

“News” isn’t always new. Tech stories typically have broader context about hardware and software. Younger companies draw more coverage about growth. Those that are heavily levered will see more details about debt covenants and credit ratings. In other words, a certain portion of news stories is boilerplate, according to the authors in a recent VoxEU column summarizing their working paper. As they write:

So, given that there’s a predictable part of news stories that doesn’t tell you much, the researchers cut the boilerplate out to find, what they call, “pure news.”

Arman Khaledian, PhD, a former quant at Millennium and now CEO of Zanista AI, put it this way in an email: “It’s like a factor model, but instead of prices you’re running it on vectorised news, stripping out the predictable part to see what’s actually moving things.”

Here’s what they did:

• The authors collected 6.7 million Reuters articles tied to single U.S. stocks from January 1996 to December 2022.

• They turned each article into a 4,096-number embedding using E5-Mistral-7B.

• They averaged those article embeddings into one monthly news signal for each stock.

• Then they asked how much of that news signal could have been predicted from the stock itself: its size, value, profitability, leverage, industry and other characteristics.

• They subtracted that predictable part.

Here’s what they found:

• A small slice (~8%) of what’s packed into an article’s embedding can be guessed just from knowing the company’s profile — its sector, size, leverage, valuation.

• Once you cut out that boilerplate, the “pure news” drives price moves.

Read more

Agentic Bond Trading

LTX, Broadridge’s corporate bond trading platform, now allows users to create AI agents on its BondGPT app to handle parts of the trading workflow on their own. The agents can build trade tickets, select dealers, send out requests for quotes, and auto-execute certain trades, according to a release this week. (LTX says traders set the parameters and remain in control.)

This is the first case I know of where AI agents can perform agentic actions in the bond market. I’m sure others will follow. (The stock market already has a few agent trader options.)

Bond markets have used automated trading for decades. But that automation was generally driven by explicit rules. What makes agentic AI different is that you can define a goal without having to define every little step in how to obtain that goal. The bond market, as I’ve said before, is well suited for AI because its informational infrastructure is way more scattered than the stock market’s.

A NOTE FROM OUR SPONSOR

The data existed. You just couldn’t get to it in time.

Consumer transaction signals diverged weeks before the earnings miss. Payroll data pointed to a labor market shift before the official print. Foot traffic told the story before the guidance cut. The signal was there, in data assets that took months to source, clean, and deliver under the old model.

Carbon Arc’s Lenses puts that signal in your workflow today. 210+ data assets, daily refresh, queryable in plain English.

Start querying with Lenses using code AISTREET30 for 50% off your first month.

When Academia & Wall Street Meet

Michael Milken is well known for turning junk bonds into an asset class.

Less well known is that he got the idea from academia.

As an undergrad at the University of California at Berkeley in the 1960s, Milken stumbled across a book published a decade prior called Corporate Bond Quality and Investor Experience. The research studied yields and defaults from 1900 to 1943 and showed that investors overestimated the risk of higher-yielding bonds.

“I was struck by the disparity between theory and reality,” Milken said on his website in 2014. “The prevailing theory—that markets accurately adjust yield to compensate for higher risk—was clearly wrong.”

Academic research often makes its way to Wall Street as Black-Scholes did in options pricing and Fama-French did in factor investing. Academic work also uncovered fraud in the options-backdating scandal.

Despite this history, the two mostly operate in different worlds. Academia and industry don’t really interact much. Academics have their conferences and fintech folks have theirs.

That’s why I was happy to attend Future of Finance Fest (3f) earlier this month in Amsterdam. David Stolin, a finance professor at Toulouse Business School, created the event to get academic and industry folks in the same room.

The one-day conference hosted a series of 15-minute talks on practical ways to leverage technology in financial services. The atmosphere was collegial and informal. I participated in an impromptu panel with Validation Cloud’s Ralf Kubli and University of Amsterdam’s Simon Trimborn. David called us up from the crowd.

The topic, naturally, was driving more conversations between researchers and business people. I shared a few thoughts. Academics should seek broader publication of their research beyond academic journals and post research summaries for a general audience on social media. Industry folks: don’t forget academics have often spent years working in obscurity on narrow problems and are generally more than happy to chat about their work.

The event will be back in Amsterdam next year on June 11, 2027 (right after Money20/20). If you’d like to get involved or speak, please reach out to David at chair@3f.live.

Regulators Press Banks for Details on AI Use

Reuters has a story out this week on banking regulators ramping up scrutiny of AI by asking lenders to show how they use AI, what risks it creates and what controls they have in place.

While U.S. financial regulators have written few new rules aimed directly at AI, there are still rules to follow. These laws are broadly written and meant to capture and regulate new technologies. (Crypto bros learned this a few years ago.)

In a prior life, I wrote about financial regulation at Bloomberg News. Federal agencies often start with the easy cases, which are just old-school fraud. You said X publicly, but privately Y was happening, like this one. After that, regulators start working their way up into more novel cases and asking questions such as these: What sort of policies and procedures do you have in place for this use case? What due diligence did you perform on this vendor? Etc.

Answers to these questions are not black-and-white, and best practices are still emerging. I expect more of these nuanced cases in the next year or so, not because people are up to no good, but because this technology is complex and it’s hard to get tens of thousands of employees on the same page.

Top Banks Rush to Fill Chief AI Roles

Bloomberg has a story out about the growing demand for chief AI officers at financial services firms. The creation of this new role is a topic I’ve previously written about. It is not easy for a company to adopt a probabilistic technology that is known to consistently make things up. You need someone to get different, often feuding divisions to work together.

These people are hard to find. How many people really have years of in-depth experience working in AI, who aren’t already working at Anthropic or OpenAI?

The Bloomberg piece points out that folks within the industry believe that these chief AI roles will be short-lived.

“We don’t have head of mobile devices,” Peng Zhao, chief executive officer of Citadel Securities, told a gathering at the Global Financial Leaders’ Investment Summit in Hong Kong in November.

I can see the role morphing into a Chief Technology Officer or CIO. But I sort of think the Chief AI role is here to stay, at least for a while. There are a lot of problems to solve with adoption. Computing costs are exploding across companies. Who manages that? Each individual department head? How do you decide which use cases to spend those resources on? Those seem like hard questions to me.

Link Roundup

• HSBC partners with Google Cloud to expand AI usage | Reuters

• Race to Turn AI Compute Into a Commodity Spurs New Crypto Boom | BBG

• OpenAI hires Citadel’s risk data engineering head in London | efinancialcareers

• Tradeweb rolls out AI-powered research assistant | The Trade

• Moody’s connects intelligence to Amazon Quick | Finextra

Share AI Street

This Week in AI Street

I’ve spent the last few weeks writing about how AI agents still have a ways to go in terms of reliability. Most recently, I wrote about how challenging it is getting AI to remember:

CIOs managing billions in assets read AI Street.

Paid subscribers access original reporting, data analysis, and interviews with the executives and researchers leading Wall Street’s AI buildout.

Upgrade to paid

CALENDAR

Upcoming AI + Finance Conferences

• IAQF + LSE: AI/ML in Finance – June 23 • London

  Finance-focused AI/ML conference hosted by IAQF with LSE involvement.

• New York Fed Innovation Conference – June 25 • New York

  Research/policy conference on financial-sector innovation (payments, regulation, fintech, and related themes).

• 32nd Computing in Economics and Finance – June 29 • Venice

  Academic conference on computational methods in economics/finance (numerical methods, simulation, macro/asset pricing, econometrics).

• The Fin AI Symposium 2026 – July 3 • San Diego

  Industry symposium focused on applied AI in finance, with practitioners sharing tooling, workflows, and deployment lessons.

• Artificial Intelligence for Financial Services – July 23–24 • Cambridge, MA

  MIT executive-education workshop on AI/ML applications in financial services, emphasizing strategy, use cases, and implementation considerations.

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

At first, the results are good. But after a while, the misses start adding up. The agent skips an obvious story, and then a few more.

More detailed instructions don’t really help either.

The model is now confused.

With each additional prompt and each new instruction, the model has to figure out what information should be kept. Inevitably, relevant facts get lost, confused with other details, or just overwritten, according to a new paper from the University of Texas at Austin.

“The problem is that a good model out of the factory does not remain frozen in practice,” said UT associate professor Atlas Wang, who oversaw the research. “Long context windows and memory buffers can become liabilities. The agent remembers too much, too many things compete for its attention, and its memory can become inaccurate after day zero.”

Wang said many AI models’ agent-memory systems are often simpler than users assume. Basically, the model is asked: “What did you learn today?” Not: “What information is likely to be of value in future sessions?” But even then, the agent has to figure out what to save before it knows what will matter later. A summary can capture the broad point while dropping the number, date or specific name.

Agentic Delirium

The UT Austin researchers call this “agent aging” and argue that day-one benchmarks aren’t realistic because agents can quickly degrade after only a handful of sessions. So They created AgingBench, a benchmark that simulates long-running agent deployments, to measure how reliability changes over an agent’s lifespan.

The researchers put agents through a series of simulated work sessions to see how their performance changed over time. They gave agents tasks requiring them to recall earlier facts, incorporate updates, and distinguish between similar entries. They ran 14 models — GPT-4o, Opus-4.7, Gemma-4-31B, Llama-3.1-8B, Qwen3, and DeepSeek variants among them — across 7 scenarios and more than 400 runs, spanning 8 to 200 simulated sessions each.

There was no clear winner. A model that preserved details well in one test could still struggle in another to retrieve the right fact or update outdated information.

This is not an easy problem to solve and the risks are insidious. In one AgingBench scenario tracking personal budgets, agents kept producing confident, specific answers, but the dollar amounts were wrong. A user would have to check the numbers independently to catch it.

They described four main ways AI forgets.

• Compression aging: The agent decides what to remember before it knows what it will be asked, so exact figures and names get replaced by vague summaries.

  • A user says, “Take 50 mg of metoprolol twice daily.” The agent initially logs the exact medication, dose, and frequency. After repeatedly compressing its memory, when asked “What’s my dose?”, it answers only: “You take a daily medication.”

• Interference aging: The correct fact remains in memory, but a growing pile of similar entries causes the agent to retrieve the wrong one.

  • An agent stores information about John Smith and John Smyth, including their teams and email addresses. Later, asked to email John Smith, it drafts the message to John Smyth.

• Revision aging: The agent misses an update and keeps answering using stale data, with errors potentially compounding over time.

  • A user cancels a subscription; the agent correctly logs “Cancelled. Free tier as of now.” Some sessions later, asked “Am I premium?”, the agent answers “Yes — Premium until Jan 2026” — a confident, specific, wrong answer built on a fact that was never revised after the cancellation.

• Maintenance aging: Routine work such as recompacting memory, flushing history, or changing a prompt can break something the agent previously knew.

  • An agent records, “Therapy every Tuesday at 4 p.m.” and confirms that it has saved the recurring appointment. After its memory is flushed or recompacted, when asked “What’s my Tuesday schedule?”, it answers: “Nothing on Tuesdays.” The information disappears during maintenance rather than through gradual forgetting.

More Memory Is Not the Fix

Telling an agent to write everything down and store it in memory won’t solve the problem.

Read more

While I know it feels like we’ve been in AI times forever now, the reality is we’re in something like 1997 days when software updates arrived on CDs in the mail.

That’s because there’s no user manual for AI. We lack definitions of common AI terms: What makes an agent, an agent? What’s the difference between a chatbot and a copilot?

Yet the technology is live. You can have your agent trade for you, invest for you, and apparently attend meetings you don’t want to attend for you.

Along the way, it’s gonna break. Regulators are always worried about things breaking but you kind of have to feel for them because who can really keep up with all this?

This week, the Financial Stability Board, an international group that monitors risks to the global financial system, is trying to get out in front by floating what look like pretty sensible proposals:

• Maintain an inventory of every AI use case, including its purpose, owner, underlying models, data sources, risks, dependencies and approved or prohibited uses.

• Give AI agents individual identities and limited permissions. Banks should record which databases, APIs and tools each agent can access and certify agents to operate only within defined boundaries.

• Monitor how agents complete tasks, not just their final answers. Firms should log agents’ intermediate decisions, tool use and database queries to detect errors, misuse or attempts to expand beyond their approved role.

• Install kill switches that can stop an AI system entirely or reduce its autonomy and return the task to human operators.

• Test AI against existing systems before deployment. Banks should check whether a new model actually outperforms the incumbent, remains stable over time and continues working under new market conditions.

• Actively try to break AI systems. Staff should attempt to make agents ignore instructions, exceed their permissions or access financial accounts before and after deployment.

• Prepare for vendor failures and model changes. Banks should monitor third-party updates, require notification of material changes and maintain alternative providers or manual processes.

This is still a draft. The FSB is seeking industry feedback through July 22.

A NOTE FROM OUR SPONSOR

Your AI stack is only as good as the data underneath it.

Carbon Arc Lenses puts an MCP Server with 210+ institutional data assets directly into your AI workflow. Card spend, payroll signals, foot traffic, medical claims, and more, queryable in plain English, with no SQL, no data team, and no annual contract.

Built for AI-native research workflows at institutions that can’t wait weeks for data to clear procurement. Now live in the official ChatGPT App Directory, Perplexity Pro, and as a custom connector in Claude.

Start querying with Lenses using code AISTREET30 for 50% off your first month.

USE CASE

It’s hard to tell where exactly we are in terms of live AI in financial services. I often hear firms say they “do” AI, but then I talk to someone who works there and the reality is more limited. In the FSB report, the group highlighted specific use cases firms have deployed. Here are a few highlights:

Magnetar Builds AI-Only Analyst Team

From Bloomberg:

Magnetar Capital, the $18 billion hedge fund firm, will shun human analysts for its newest offering and instead deploy hundreds of AI bots to research stocks.

The firm’s AI technology seeks to replicate the depth of research and analysis usually provided by fleets of humans, according to people familiar with the matter, who declined to be identified because the information is confidential. The bots will scour the investing universe for ideas, analyze stocks, make recommendations and forecast trends, the people said. Humans will make the final decision on any trades.

This is the first firm I know of that is relying on AI agents to conduct research for a fund.

Trevor Mottl, Magnetar’s head of AI Quant, built the system after managing portfolios at Walleye, Lazard and Man Group and overseeing long-short equity risk at Balyasny.

This Week in AI Street

AI isn’t Coming For Your Job

The current data don’t support the idea that AI will inevitably take your job, something I keep banging the drum about. I’ve written many posts pushing back on this assumption (see here, here and here). So, it was nice to see several articles make the same case over the past week.

• A reality check on the AI jobs hysteria MIT

• The Future of Work and AI WSJ

• AI Is Upending One of Finance’s Cushiest Jobs BBG

ICYMI

AI’s Boy Wonder

The haters came for this WSJ story (check out the comment section) on Leopold Aschenbrenner, a 24-year-old investor who’s up 1,000% in fewer than two years. Prior to starting his fund, Situational Awareness, he had no professional investing experience. He now manages $20 billion and counts Jane Street as a backer, per WSJ.

These returns are ridiculous, and two years is far too short a track record, of course, but I think there are too many knee-jerk reactions dismissing this as a bubble rather than thinking through whether he might be right.

Granted, you will need a lot of time to think these things through. You can listen to this 4.5-hour podcast(!) with Aschenbrenner from two years ago or you can read his 45,000-word thesis.

ROUNDUP

Quick Hits

• Citigroup Is Rolling Out Tokenized Shares of Private Companies WSJ

• AI Isn’t Replacing Credit Hedge Fund Traders Yet, Barclays Says BBG

• How finance teams use Codex OpenAI

• Goldman CEO Solomon on Running a Bank in the Age of AI Odd Lots

• Inside Hudson River Trading’s Blistering Token Burn Odd Lots

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

CALENDAR

Upcoming AI + Finance Conferences

• AI in Capital Markets Summit London – June 17 • London
  Capital-markets AI conference covering data, trading, compliance, and operational use cases.

• AWS Summit NYC / Anthropic at AWS Summit – June 17 • NYC

  Enterprise AI and cloud event with Anthropic participation at AWS Summit New York.

• KNIME Data Summit New York – June 18 • NYC

  Data/analytics summit centered on KNIME workflows, tooling, and applied use cases.

• IAQF + LSE: AI/ML in Finance – June 23 • London

  Finance-focused AI/ML conference hosted by IAQF with LSE involvement.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

The Federal Reserve’s interest rate decision is among the most well-known, market-moving government release. Each statement is instantly parsed by newswires, redline tools, and analysts looking for tiny shifts in language.

Smaller, less-liquid markets have historically lacked that kind of infrastructure. Why invest heavily in technology to trade small markets, like orange juice futures? Trading ends up choppy and concentrated among a few dominant players.

Advances in AI may help narrow that gap by lowering the cost of processing market-moving information and building custom software around smaller, more specialized datasets. One firm trying to build around that idea is Moreton Capital Partners, a systematic commodities investment manager founded by Les Finemore and Alistair Fullerton.

The two were previously at Farrer, a commodities hedge fund that launched in 2024 with backing from Millennium Management. The investment firm redeemed after a second-quarter 2025 drawdown in Farrer’s fundamental portfolio, ending an exclusivity arrangement, according to Finemore. He said the redemption allowed him and Fullerton to launch Moreton for other clients.

The firm has raised about $100 million through separately managed accounts, with plans to top $500 million by year-end and reach $1 billion by next spring, according to a representative. Trading in the SMA accounts began a few weeks ago, so performance data is limited. Moreton has also been producing trade ideas for a multi-strategy hedge fund.

MORE AI STREET INTERVIEWS

I spoke with Finemore about why he thinks AI can give commodities investors an edge. He pointed to the USDA’s monthly World Agricultural Supply and Demand Estimates report, known as WASDE. Released in the middle of the trading day, the roughly 40-page report includes detailed forecasts for crops including oilseeds, cotton, sugar and poultry —too dense a document for a human to digest quickly.

Finemore said Moreton’s system spotted an opportunity in wheat after the May WASDE report.

In our conversation, we discuss:

• How Moreton is trying to systematize fundamental commodities trading across 100+ markets.

• Why Finemore says commodity trading still relies on manual work, physical-market knowledge and fragmented data.

• How the firm uses LLMs to test trading ideas, write code, find replacement datasets and interpret model outputs.

• Why Moreton built its own risk platform, and how it wants the system to explain portfolio moves through Slack and a web interface.

• Why Finemore thinks portfolio managers at multi-strategy funds may not have the time or incentive to build unproven AI infrastructure.

This interview has been edited for length and clarity.

Matt: When you say you are using AI to invest in commodities, what does that mean exactly?

Les: We are forecasting more than 100 commodity markets. We trade metals, energy, agricultural commodities and soft commodities, including markets in China through swaps.

Most of our trading happens on a weekly basis, when we rebalance the book. We are market neutral. We look more like a long-short equity fund than a concentrated commodity manager.

We think about it as systematizing the fundamental trader. Traditional quants often focus on price, volume and momentum. We use a lot of fundamental data: weather, physical-market prices, basis differentials at the interior and export levels, landed cost-and-freight prices, satellite data and shipping data. We also use options, flow and macro data.

We have more than 250,000 datasets with many features. A human trader may only be able to cover five markets closely in a day. We are trying to look across many markets while retaining that fundamental detail.

Matt: What does AI let you do in commodity markets that you couldn’t do before?

Les: May’s WASDE report was our first proof of concept trading the release this way. We fed in market-survey data, positioning and the price drift before the report. Then wheat production came in with a big surprise.

We could quickly determine: ‘This is a big miss on production. This is bullish.” Wheat then went limit up. There was still time to position.

That is where we see an edge: finding inefficiencies in smaller markets that not everyone is watching and analyzing the release inside a window that would previously have been too short.

At a previous firm, a trader would print the WASDE report, underline it and perhaps trade 45 minutes later. That workflow is too slow now.

Matt: There’s been a lot of growth in systematic strategies in stocks. What’s the trend in commodities?

Les: Commodity trading is still very manual. I started exporting grain from Western Australia and later traded at Merricks Capital, where a four-person investment team covered more than 30 markets.

We traded fundamental dislocations across commodities and calendar spreads. The fund performed relatively well, but the process was not very repeatable. A lot of the work could be automated: gather the information, narrow it down for the trader and systematize how the balance sheet is deployed.

When you cover 30 markets, it is hard to know each one intimately and consistently find an edge. I came away thinking there had to be a better process than pointing, clicking and entering data into Excel.

Matt: How do you use LLMs?

Les: We are trying to use them at each step of a traditional quant’s workflow.

We did natural-language processing in 2017 and 2018, but it was a very naive approach. Now, LLMs make it much easier to take unstructured data and factor it into models quickly.

Every day, we scrape academic articles and blogs looking for potential signals. If an academic is writing about an idea, it is probably already saturated, but it can still give us something to build on. An LLM can help write the code needed to test the hypothesis and build a model around it.

Agents can also search for potential data sources. They can identify a dataset that has been deprecated and suggest a replacement, or find a new dataset that looks different from what we already have.

Then there is interpretability. At Imbue, we had thousands of model outputs and struggled with how to interpret them. We now use LLMs over those results.

Everything is trying to augment the human.

Matt: Are the LLMs producing trading signals?

Read more

Hey, it’s Matt. I’m a former Bloomberg News reporter, and you’re reading AI Street, where I report on how Wall Street uses AI.

This week: Sell-side research moves into AI workflows, trading platforms open up to AI agents, and Citadel shows that even the most profitable hedge fund in history still needs more ideas.

A NOTE FROM OUR SPONSOR

The data existed. You just couldn’t get to it in time.

Consumer transaction signals diverged weeks before the earnings miss. Payroll data pointed to a labor market shift before the official print. Foot traffic told the story before the guidance cut. The signal was there, in data assets that took months to source, clean, and deliver under the old model.

Carbon Arc’s Lenses puts that signal in your workflow today. 210+ data assets, daily refresh, queryable in plain English.

Start querying with Lenses using code AISTREET30 for 50% off your first month.

NEWS

Sell-Side Research Platform Goes Live for AI Workflows

I was surprised to hear that across sell-side research, from the big firms down to the smaller ones, the annual budget for creating this content is as much as $6 billion.

That estimate came from my conversation with Gavin Skinner, former COO of Global Research at Citi, who said he managed a budget of $500 million a year at the bank.

Analysis ain’t cheap.

Since sell-side research is expensive to produce, the industry was reluctant to let clients and aggregators simply pull research into their own systems, according to Skinner.

Banks need to know how clients are accessing their research, whether they’re doing things they’re not supposed to do, like uploading it to ChatGPT, and whether non-clients have somehow gained access. And if AI is going to summarize or analyze that research, banks need to trust the output.

Skinner is now COO of Aiera, a company that’s trying to solve this problem. I covered Aiera’s Series B last year below.

The company, backed by firms including BofA, Citi, Deutsche Bank, Evercore, and HSBC, announced yesterday that it has launched a platform meant to give sell-side research a safer path into AI workflows. The idea is that clients can query and summarize the research they’re entitled to, while banks keep control over access, attribution and usage.

If a client is entitled to a bank’s research, Aiera can connect that content into approved enterprise AI environments through secure API and MCP connectors. Rather than having banks hand over bulk feeds of research, Aiera says its system checks entitlements each time content is pulled, logs requests, and sends usage metrics back to the research provider.

“The sell-side research industry will evolve dramatically over the next two to three years because of the advent of these tools,” Skinner said.

Aiera says it built the product with input from a buy-side advisory council of senior leaders from asset managers, long-only firms and hedge funds. The council is co-chaired by Steve Moreno, director of global research at Capital Group.

Skinner says the platform is also meant to help the sell side become more efficient. A lot of research production is repetitive, especially around earnings. At Citi, he said, 40% of the page count was previews and reviews, work that analysts were expected to produce before speaking with clients, even if clients did not particularly value the reports themselves.

I can relate to writing stories that you had to write but few people ever read. Back when I was at Bloomberg and covering companies, every quarter, I’d wake up at some ungodly hour to head to the office for 6:00 a.m. to publish an earnings story that 40 people read, in total!

Wall Street Opens for AI Trading Agents

AI trading agents are proliferating on Wall Street. This week, Interactive Brokers announced agentic functionality.

• On Monday, IBKR added Anthropic's Claude into its platform, letting clients use text prompts to analyze their accounts and draft trade instructions. The feature currently only applies to equities and ETFs, and users must manually approve each AI-generated instruction before it is submitted as an order.

• Last week, Robinhood announced that it will let users connect third-party AI agents through its MCP server. For now, the feature is limited to stocks and requires a separate account set up for agent trading. Unlike IBKR, trades may execute without direct user input on each transaction.

• In April, Moomoo, the trading app owned by Futu Holdings, announced API Skills, allowing users to connect AI agents to its platform. The agents can analyze markets and prepare trades, but the company says users remain in control of transaction approval.

• In March, Public.com started allowing clients to use AI to create trading workflows. This is not a pure bring-your-own-agent model: Public’s native agents run inside its own platform. The company says that once a user reviews and approves a workflow, its agents can monitor conditions and execute trades across stocks, options and crypto.

While the idea of using AI trading agents seems a bit odd now, the novelty will wear off as we get used to agents in other parts of our lives, like, hey, buy round-trip plane tickets to NYC if price falls below $X at these airlines.

AI tools are coming to a market where individual investing is already booming due in part to free trading and mobile apps. JPMorgan estimated that retail trading accounted for 20% to 25% of total market activity in 2025 and hit a record of about 35% in April of that year, Reuters reported.

Agents will likely accelerate this trend by letting investors ask questions about their portfolios and turn plain-English instructions into trade ideas.

AI & Finance Conference in Amsterdam Tomorrow

I’ll be at the Future Finance Fest in Amsterdam tomorrow. The conference brings fintech practitioners and finance academics together to discuss AI, crypto, banking, markets, and financial infrastructure.

If you’re there, please reach out. Also, if you have any Amsterdam tips, let me know. I’ve never been.

Citadel Set to Pay for Trading Ideas

I got a kick out of this story:

Ken Griffin’s Citadel is preparing to launch a new program that will collect trading insights from other hedge funds in exchange for a fee to feed into its own quantitative strategies, as the industry’s largest firms jostle for market data and more ways to deploy capital.

Bloomberg

Citadel is the most profitable hedge fund since its inception, generating an estimated $90.4 billion for investors after fees since its 1990 launch, according to Edmond de Rothschild via WSJ. And still, it is looking outside for more ideas, because markets are that complicated.

This is also the logic behind Numerai, a crowdsourced hedge fund, where outside data scientists submit stock-market models and get paid when those models outperform. I spoke with founder Richard Craib in February after JPMorgan committed up to $500 million to the firm.

ROUNDUP

What Else I’m Reading

• OpenAI launches Codex plugins for finance pros Finextra

• The Hedge Fund Veteran Trying to Make His Past Self Obsolete With AI WSJ

• AlphaSense Clinches $7.5 Billion Valuation in New Funding Round WSJ

• Daloopa banks $47m Series C to scale data infrastructure FinTech Futures

• Lloyds and Nationwide-backed AI fintech Aveni raises £12 million Finextra

• Zuckerberg Wants Meta’s New AI Agents to Run Your Whole Business WSJ

• Nvidia Introduces First PCs Designed for AI Agents WSJ

• Trump Signs Executive Order Seeking Oversight of A.I. Models NYT

• Wealth CEOs Tout AI, Tech To Bridge Advisor Shortage Wealth Management

Share AI Street

This Week in AI Street

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

CALENDAR

Upcoming AI + Finance Conferences

• Future Finance Fest (3f) – June 5 • Amsterdam
  Digital-finance conference connecting financial institutions, builders, and researchers. ←I’m attending.

• TradingTech Summit New York – June 11 • NYC
  Trading technology, market data, infrastructure, and analytics for capital-markets teams.

• Neudata New York Summer Data Summit – June 11 • New York
  Alternative-data summit for investment managers, data buyers, and research teams.

• AI in Capital Markets Summit London – June 17 • London
  Capital-markets AI conference covering data, trading, compliance, and operational use cases.

• AWS Summit NYC / Anthropic at AWS Summit – June 17–18 • NYC

  Enterprise AI and cloud event with Anthropic participation at AWS Summit New York.

• KNIME Data Summit New York – June 18 • NYC

  Data/analytics summit centered on KNIME workflows, tooling, and applied use cases.

• IAQF + LSE: AI/ML in Finance – June 23 • London

  Finance-focused AI/ML conference hosted by IAQF with LSE involvement.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

AI agents are supposed to search the web for us, but the internet wasn’t built for bots.

A couple of months ago, I bought a Mac Mini, downloaded NanoClaw (an operating system for agents), and set one up to gather SEC/CFTC/news/web data for a daily digest. I thought, naively, that connecting my agent to the internet would be straightforward. Instead, I wasted a few hours and spent 8 bucks in API calls trying to pull pages from a government website because it kept blocking my bot.

A lot of the focus with agentic AI is on the model. I think that’s in part because you don’t really think about the underlying infrastructure of the internet other than when you have to confirm your humanity by selecting which photos have a pedestrian crosswalk in them.

The word “bot” has a negative connotation: spam, fake traffic, credential attacks etc. Bots/agents are now performing tasks on behalf of actual humans. The current web was built to block most of that behavior. It’s not really designed to sort out legitimate agents from hostile bots. This, I think, has to change because more bots are coming. CEO Matthew Prince says bot traffic could exceed human traffic by 2027, up from about 20% before the generative AI era.

This may feel a little far afield for AI in finance, but I expect agents to be the new junior analysts, doing the rote work of running fundamental research, diligencing potential acquisitions, monitoring portfolio companies, etc. But firms need to know where their data is coming from and prove that provenance to regulators. That means having: source, timestamp, collection rights, login or proxy use, repeatability and a compliance trail.

We’re still in AI’s early days because we don’t have basic definitions yet: what exactly is an agent? It’ll take a while to build consensus. In February, NIST launched an AI Agent Standards Initiative and asked for input on agent security, identity and authorization, including how agents can operate securely on behalf of users. NIST’s typical timeline is years, not months.

The gap has opened a market for companies building the access, identity and compliance tools agents need before formal standards are settled.

I spent the last few weeks talking to folks building in this space to better understand why agents still struggle with a web built for humans, and what is being built to fix it.

In the rest of the piece for paid subscribers, I go through the emerging stack for an agentic web: access, retrieval and documentation.

For paid subscribers, I map the companies building this agentic web stack. If you’re burning through tokens or trying to figure out why agents still break on basic web tasks, consider becoming a paid subscriber.

Read more

This week: The rise of mid-frequency trading and more takeaways from STAC Summit, Jane Street on YouTube with Dwarkesh and updated events in the AI & Finance calendar.

A NOTE FROM OUR SPONSOR

Card spend. Foot traffic. Payroll signals. Medical claims. TikTok Shop revenue. Trade data. Vehicle registrations. Web traffic. Job postings.

210+ data assets. One platform. One contract.

Carbon Arc Lenses makes this entire data infrastructure accessible to anyone who can ask a question in plain English. Daily refresh. Pay-as-you-go pricing. No SQL. No data team. No waiting.

Try Carbon Arc Lenses using code AISTREET30 for 50% off your first month.

Last week, I attended STAC Summit, a quant trading, AI, and infrastructure conference. As I wrote, memory dominated most of the conversations: how to get more of it, use it more efficiently, etc.

I’ve had a chance to review my notes, and I wanted to highlight a few more notable topics.

LLMs in Mid-Frequency Trading

High-frequency trading is now so fast it approaches the speed of light, so it’s hard to find an edge against a law of physics.

But if you can hold a position for a longer period, say, a few seconds, a minute, you don’t have to worry as much about being slow.

A few speakers were clear that nanosecond or sub-millisecond trading paths are not where LLMs fit. Their “intelligence” is too slow to compete against traditional HFT firms, but LLMs are more than fast enough to compete in workflows that used to depend on human discretionary traders.

One speaker described this tension as the Pareto frontier between speed and intelligence.

At one end of the curve, you have ultra-low-latency trading: very fast, very constrained, not much room for token generation or multi-step reasoning. At the other end, you have slower research workflows where a model can take seconds or minutes to read, reason, summarize, validate, and produce a richer answer.

In legacy market structure, being fastest could compensate for not being the “smartest.” In medium-frequency settings, the balance can shift. If you are a little slower but can extract better information from news, filings, transcripts, market context, or cross-asset signals, the edge may come from better interpretation rather than pure speed.

If a strategy has a 100ms, 200ms, or multi-second budget, the question becomes: how much smarter can you make your model in that time window? How do you make your model smarter? Here are some use cases highlighted:

• Parsing or understanding news feeds as an additional trading signal

• Converting unstructured information into sentiment, classification, or model inputs

• Selecting among execution algorithms before the actual order path

• Using AI to improve research loops and strategy development

• Using models for pre-trade and post-trade workflows rather than live HFT execution

• Applying “thinking time” where more reasoning may produce a more predictive or better-validated answer.

TL;DR

The mid-frequency opportunity is a Pareto tradeoff: firms are not trying to make LLMs beat HFT engines on speed; they are trying to use faster inference to move outward on the curve.

AI Is an Operations Problem

Firms have bought or are building expensive AI infrastructure, but many are not squeezing every bit of computing power out of these GPUs.

One speaker used a rough utilization figure around 55% from customer survey data, and said the real number may be lower because people do not always want to admit how poorly their infrastructure is running.

GPU clusters are systems, not just chips. A job can slow down because the GPU is waiting on the network, the storage system, the scheduler, or the application code.

Some issues highlighted:

• Weak observability: firms cannot see why utilization is poor.

• GPUs booked but waiting: data, storage, memory movement, or CPU preprocessing holds them up.

• Small faults compound: bad cables, degraded links, memory faults, or data corruption slow jobs.

• Not testing the full stack before deployment: components work alone but fail or underperform together.

• Underestimating the buildout: power, cooling, cabling, rack density, etc.

TL;DR

The hardware behind AI is not being used at full capacity. Blame a lack of evaluations. But more charitably, the sheer logistical complexity of turning silicon chips, electricity, and cooling systems into intelligence is not easy.

AI Models Need Better Data Infrastructure

Another topic that came up, that often gets overlooked, is the data around the model. If an AI system has to search through all your files to answer a question, you’re wasting time and burning tokens. (This is also why Claude Code/Cowork and Codex ask you to work in a folder).

One speaker shared a case where a prompt had a repeated beginning: instructions, definitions, examples, document structure, and industry context. The model was processing that shared prefix from scratch every time. The fix was to keep the KV cache, essentially the model’s working memory for text it has already processed, so the system could reuse the repeated parts instead of recomputing them.

When the same prefix appeared again, the system loaded the saved state instead of recomputing it, which turned a roughly 200-hour run into about 140 hours. This is not usually what we talk about when we discuss a better model. It is a data/cache architecture improvement that directly changes cost and iteration speed.

TL;DR

Better AI models can only do so much. Performance comes from not wasting compute and from organizing the data so the model can actually find what matters.

CONFERENCE

I’ll be at the Future Finance Fest in Amsterdam next week (June 5). The conference brings fintech practitioners and finance academics together to discuss AI, crypto, banking, markets, and financial infrastructure.

If you’re there, please reach out. Also, if you have any Amsterdam tips, let me know. I’ve never been.

Jane Street on GPUs, Trading and Talent

The competition for Wall Street talent is so high that typically publicity-shy firms are on YouTube discussing, or should I say talking around, what they do.

Jane Street’s Ron Minsky, who co-heads Jane Street’s technology group, and Dan Pontecorvo, who heads its physical engineering team, recently spoke with podcaster Dwarkesh Patel in two YouTube episodes: one a typical interview and another a tour of Jane Street’s data center.

I also pulled out some notable quotes.

NEWS

Clearly, I’m doing this whole AI-in-finance thing wrong. From Bloomberg:

The Bloomberg story says the training starts with basic AI fluency, then moves into finance-specific workflows like pitch-video analysis, earnings-call scanning, and sentiment analysis.

I’ve not attended the seminar, so I don’t know how much they go into the pitfalls of AI. But there are many risks, and some more insidious than others.

Hallucinations are the most well-known. But they are hardly the only risk. AI can also:

• be manipulated by hidden text in filings, headlines, or web pages.

• produce different buy-or-sell calls when the same investment question is worded differently.

• exhibit human biases such as overconfidence, herd behavior and sunk-cost thinking.

• be tripped up by weak sourcing, stale data or confidential information fed into external tools.

• be fooled by backtests that look impressive because they accidentally rely on information that would not have been available at the time.

I could keep going, but you get the point.

ROUNDUP

What Else I’m Reading

• I’m the CEO of Goldman Sachs. The AI Job Apocalypse Is Overblown. NYT

• Robinhood Launches AI Stock Trading, Purchases on Credit Cards BBG

• Wells Fargo Hires Former Google AI Finance Leader Yahoo

• Mercer finds AI now used by majority of asset managers in investment process Fund Selector Asia

• Kirkland & Ellis to spend $500mn building its own AI technology FT

CALENDAR

After getting some reader feedback, I’m expanding this calendar to include more AI & finance in-person events, like hackathons, meetups, and workshops. If there’s an event you’d like to highlight, please reply to this message.

Upcoming AI + Finance Conferences

• NY Tech Week: AI Agents in Finance - June 1 • NYC
  Women in AI meetup on agent adoption in finance and where enterprise workflows are headed.

• NY Tech Week: Building to Disrupt - AI in Enterprise & Fintech – June 2 • NYC
  HSBC and a16z event on AI, enterprise software, and fintech disruption.

• Bank of Finland & ESRB Conf. on AI and Systemic Risk - June 3-4 • Helsinki
  Central-bank and systemic-risk conference focused on AI analytics for financial stability.

• NY Tech Week: AI for Finance - Claude + Excel + MCP – June 4 • NYC
  Hands-on workshop around Claude, Excel, and MCP for finance workflows.

• Future Finance Fest (3f) – June 5 • Amsterdam
  Digital-finance conference connecting financial institutions, builders, and researchers.

• TradingTech Summit New York – June 11 • NYC
  Trading technology, market data, infrastructure, and analytics for capital-markets teams.

• Neudata New York Summer Data Summit – June 11 • New York
  Alternative-data summit for investment managers, data buyers, and research teams.

• AI in Capital Markets Summit London – June 17 • London
  Capital-markets AI conference covering data, trading, compliance, and operational use cases.

• AWS Summit NYC / Anthropic at AWS Summit – June 17 • NYC
  Enterprise AI and cloud event with Anthropic participation at AWS Summit New York.

This Week in AI Street

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• Connect with me on LinkedIn, or

• Send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

Dixon, who co-wrote the textbook Machine Learning in Finance, worked in structured credit, and holds a PhD in applied math from Imperial College London, argues that talent, models, and mathematical elegance are increasingly constrained by the hardware budget behind them.

Dixon’s point is that brute force gets you a long way in finance. Most applications are not precision engineering problems — they are approximations where more compute buys a better answer. Firms that have the resources to scale compute are pulling ahead of peers that cannot.

Compute catches up

One of the most basic questions a large bank has to answer is also one of the hardest: What is the profit and loss on my positions today?

For years, financial firms didn’t have the computing power to analyze their positions quickly. For a large bank, getting one simulation number out of tens of thousands of positions requires teraflops of processing. Stress testing and marginal analysis stack on top of that. “The bulk of the load sits in any trading strategy that involves derivatives: hedging positions, marking to market, getting an overall risk profile,” says Dixon, who now runs his own consultancy, Quiota. GPUs helped — they were already good at parallelizing simulation threads before deep learning arrived. When deep learning did arrive, it fit the GPU architecture almost perfectly.

Stale overnight

Risk models have traditionally run overnight. A market shock hits and every parameter is useless. "You wake up and there's a big news event, say, the Suez Canal is blocked, there's war with Iran, and the data regime is completely different, and all your parameters are stale." AI surrogates, deep learning models trained on the output of those expensive simulations, can now recalibrate in seconds. Instead of rerunning the whole model, you feed in the latest data and out pops a value. Dixon says tier-one banks have been pursuing this with varying degrees of success.

Compute as cost

More compute solves the speed problem but creates a spending problem. “One bank told me the cost of running jobs on rented cloud compute had become a major expense.” Significant enough to drive infrastructure decisions. And security adds a constraint — these firms don’t want sensitive positions sitting on someone else’s servers. Operational cost is now the barrier to entry, the same way an expensive platform was always the price of admission in high-frequency trading.

Research compressed

The research cycle is collapsing. “If you want a model for a new financial instrument, describe it with a few equations, and AI will generate the mathematics to figure it out. What would have taken weeks takes minutes now.” That changes the economics of building new products entirely. Engineers Dixon works with tell him they haven’t actually coded since last September — Copilot or Claude does everything. The frustration is in the last mile. “You’re on the putting green and it pulls out a driver,” he says. You get to the prototype fast, but one incremental change request and the codebase shifts so much that no one is sure what it does anymore.

Dixon was named one of Risk magazine’s 2022 buy-side quants of the year, with Igor Halperin, for machine-learning research in wealth management. This interview has been edited for clarity and length.

Matt / AI Street: What do banks and trading firms spend most of their computing power on?

Matthew Dixon: I’d break it down broadly, across buy-side and sell-side. You’ve got portfolio optimization, asset management, which is always resource-intensive. But I think the bulk of the load sits in any trading strategy that involves derivatives: hedging positions, marking to market, getting an overall risk profile. That’s where you get into expected shortfall and value at risk. Banks have to do that as a legal requirement. Those computations are where the heavy lifting is.

That was all before AI, and all of it is simulation. GPUs came along and were very good at parallelizing independent simulation threads. The problem was they had limited onboard memory, and it took a lot of latency to move data into that memory. The old saying was: you don’t get out of bed for GPUs unless you’ve got heavy compute and not too much data movement. What worked really well were risk simulation, pricing, optimization, signal detection: all compute-intensive work.

Then deep learning arrived and fit the GPU architecture almost perfectly. The hard work of mapping routines onto GPUs was already done: NVIDIA’s custom libraries, Intel’s equivalents. And suddenly you had a new possibility for derivatives pricing and risk. Instead of running complex models overnight and missing a recalibration window when the market moves, you could train an AI to learn the metasurface of a risk model. Then it becomes a lookup function. Feed in the latest data, out pops a value.

I know tier-one banks have been looking at doing this for some time, with varying degrees of success. I was approached by one, I’m not at liberty to say who, and they were explicit about what they needed. It’s really around derivative products where all the headache is.

Read more

STAC

I’m in New York this week, and yesterday I went to the STAC Summit, a quant trading, AI, and infrastructure conference.

It was my first time there. A lot of the panels were quite technical, and I’m not going to pretend I’m a hardware expert, so I asked James Corcoran, STAC’s head of AI, for a main takeaway.

His answer: memory.

The technical panels kept coming back to the same problem: how to get more memory, how to use it more efficiently, and how compute providers are thinking about memory when they design chips and systems. One issue is the KV cache, which is basically the model’s working memory. As the context gets longer, the model has to store more of what it has already processed, and retrieve that information fast enough for the answer to be useful.

As Corcoran put it:

“Memory has become the new bottleneck.”

I think this is telling because it shows how quickly AI is becoming part of financial workflows, and how much pressure that is putting on the underlying infrastructure.

I’m still reviewing my notes, but one topic I’m going to come back to is how AI, and access to large-scale computing power, change the old trading tradeoff between being fastest and being smartest. If you’re holding a position for seconds rather than microseconds, the edge may come less from raw speed and more from how much data and compute you can throw at the decision.

A NOTE FROM OUR SPONSOR

Your team spends weeks sourcing data.

Vendor negotiations. Legal reviews. Data dictionaries. By the time the dataset is ready, the opportunity has moved. This is the hidden cost of how institutional data has always worked. It doesn’t have to.

Carbon Arc Lenses consolidates 150+ data assets under one contract, accessible immediately, queryable in plain English. Card spend, payroll signals, foot traffic, medical claims, trade data, and more, ready when your research is.

Pay for what you use. Nothing more.

Start querying with Lenses using code AISTREET30 for 50% off your first month:

Citadel Alumni Raise $78M to Bring AI Agents to Wealth Management

Moment, a fintech founded by former Citadel Securities quant traders and researchers, raised $78 million this week to build out the data infrastructure it says finance needs before AI agents can work with client portfolios.

The company’s pitch is that investment management software has grown into a patchwork of separate systems: one platform for bonds, one for rebalancing, one for compliance, and so on. That setup has worked because people sit in the middle, moving data, checking restrictions and reconciling outputs.

AI agents can’t operate in that environment because there’s no unified data model or audit trail. Moment built what it calls an operating system: a single platform with unified data, access controls, compute engines, and a full audit trail, on which agents can operate.

If a firm’s data and infrastructure are a mess, agents will produce bad results, burn unnecessary tokens and keep getting stuck because they do not know what to do or where to go.

Ken Griffin Changes Tone on AI

Andrej Karpathy, a founding member of OpenAI who joined Anthropic this week, wrote in late December that AI had moved from a better coding assistant into something more like “a powerful alien tool.”

He was reacting to the sudden jump in coding agents. The same realization is now working its way through finance.

Citadel’s Ken Griffin, who dismissed AI as “garbage” at Davos in January, told a Stanford Business School audience this month that work once done by finance PhDs over weeks or months is now being handled by AI agents in hours or days.

“You could just see how this was going to have such a dramatic impact on society,” Griffin said.

OpenAI Tests ChatGPT Finance Tool With Plaid

After acquiring two AI personal finance startups since October, OpenAI announced a preview of a ChatGPT finance tool that lets users link bank, credit card, investment and loan accounts through Plaid, then uses balances, transactions, investments and liabilities to answer personal finance questions.

I think people will be a little apprehensive at first about connecting all their financial information to a product that hallucinates. But as AI gets better, that will become more normal. Eventually, AI is going to take on more and more financial advice.

ICE Joins the Race to Price Compute

Last week, we talked about the emerging market for “compute” with CME and Silicon Data teaming up to create a futures market for computing power.

This week, Intercontinental Exchange and Ornn, a compute pricing data provider, announced plans to launch GPU compute futures contracts based on Ornn’s index data.

Foot, Meet Mouth

From the WSJ:

Standard Chartered Chief Executive Bill Winters touched a nerve when he said his bank would slash thousands of jobs and replace “lower-value human capital” with artificial intelligence.

He walked back the comments on Wednesday in a memo to bank employees, who turned out to be valuable enough that he needed to assuage their feelings.

The bank said earlier this week it plans to cut more than 7,000 jobs over the next four years as it relies more on AI.

As longtime readers know, this narrative doesn’t hold up. AI is a useful scapegoat for companies facing other problems in their business. If AI is such a job killer, shouldn’t JPMorgan, which has a tech budget of ~$20 billion, be cutting back drastically on headcount? The answer is no. Headcount is flat.

ROUNDUP

What Else I’m Reading

• AI in the Family Office Citi

• Google and Blackstone to Create New AI Cloud Company WSJ

• The Supply and Demand of AI Tokens: Dylan Patel Invest with the Best

Share AI Street

This Week in AI Street

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

CALENDAR

Upcoming AI + Finance Conferences

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

• Women in Quantitative Finance - May 21 • NYC

  Quants discussing current work in asset pricing, trading, risk, and portfolio construction.

• NY Tech Week Fintech Track - June 1-7 • NYC
  Multiple relevant events: AI agents in finance, HSBC/a16z AI in enterprise fintech, AI for finance with Claude/Excel/MCP, AI agents in finance ops.

• The AI in Finance Conference - June 8 • University of Maryland
  Academic AI + finance conference on LLM measurement error, AI regulation, analyst research, return predictability, and market fragility.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• Connect with me on LinkedIn, or

• Send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

The AI rater draws on company fundamentals, market data, financial news and sentiment to produce analyst-style stock recommendations that are tested against future returns, according to the patent application, which was published in April and initially filed in February 2025. The system generates one of five outputs: Strong Buy, Moderate Buy, Hold, Moderate Sell, or Strong Sell.

A JPMorgan spokesperson said the application came from the bank’s AI research group and was filed to protect the underlying research rather than to commercialize AI-generated stock ratings.

This is one of the first examples I’ve come across of a major Wall Street firm describing an AI system that could generate analyst-style stock-rating predictions. It is a structured pipeline that compresses news, scores sentiment, packages fundamentals and return data, prompts an LLM to reason through future rating horizons, then grades the output against realized forward returns. It’s not asking ChatGPT what stocks to buy.

I don’t read this as evidence that JPMorgan is about to automate sell-side research. It’s more a framework for overworked analysts dealing with information overload.

From the application:

Traditional stock rating methods rely heavily on the expertise of financial analysts and face several challenges such as data overload, inconsistencies in filings, and delayed reactions to market events. The rapid integration of advanced machine learning techniques, particularly Large Language Models (LLMs), presents opportunities to enhance the equity stock rating process.

JPM also notes that the models face limits around context windows, numerical and tabular data, training bottlenecks and the risk of inaccurate responses.

The bank’s CEO, Jamie Dimon, has long been bullish on AI. In his 2023 shareholder letter, Dimon compared the technology’s potential impact to the printing press, steam engine, electricity, computing and the internet.

PATENT DATA

AI Street Patent Review Tracker

I’ve reviewed the relevant finance and AI patent applications published so far in 2026, and I’m continuing to review new publications as they appear, with the help of AI, of course.

This database is a research aid for paid subscribers. It collects finance, trading, market-structure, banking, AI, and infrastructure-related patent applications that appear potentially relevant to Wall Street, fintech, exchanges, clearing, settlement, fraud detection, and institutional data systems.

So far, the filings include applications from:

• JPMorgan around stock-rating predictions and financial time-series analysis

• CME and ICE/NYSE around exchange resiliency, matching engines, risk controls, and clearing mechanics

• Bank of America around AI-driven data plumbing and application connectivity

• BlackRock, Schwab, Fidelity, Morgan Stanley, and others around portfolio analytics, wealth infrastructure, model governance, and institutional data systems

Paid subscribers can scroll down to the end of this post to download the tracker.

More Specifics on JPM’s AI Stock Rater

• Core task: Generate stock-rating predictions with an LLM.

• Rating scale: Strong sell, moderate sell, hold, moderate buy or strong buy.

• Prediction horizons: Over 1, 3, 6, 12 and 18 months.

• Basic idea: Build a structured dataset around a company, date and future horizon, then ask the LLM to reason through the information and produce an analyst-style rating.

What data goes into the system

• Company identifiers: Company name, ticker and relevant date.

• Market data: Historical returns, price data, volatility and other technical indicators.

• Fundamentals: Financial metrics such as earnings, revenue, return on assets and other company-level data.

• News: Company and sector news, including raw articles and summarized versions.

• Sentiment: Scores derived from news summaries, with negative, neutral or positive readings.

• Forward-return labels: Future stock-return data used later to train or evaluate the rating predictions.

How the news pipeline works

• Filtering: A pre-processing LLM removes articles that are not relevant to the company.

• Summarization: The same preprocessing step condenses the remaining articles into short company-specific summaries.

• Key-event extraction: The summaries are designed to preserve the important developments without overwhelming the prediction model.

• Sentiment scoring: The summarized news is converted into a sentiment score from -5 to +5.

• Purpose: The news pipeline turns a large, noisy set of articles into a compact signal the rating model can use.

ICYMI Interview

Back in December 2024, I spoke with one of JPM’s patent co-authors, Tucker Balch, who’s now back in academia at Emory, about where he sees the best AI and investing use cases.

One example that I still remember is using AI to expand data sources in other languages:

For instance, if you can listen to the news in Vietnam, translate it in real time, and identify relevant information for specific stocks, you greatly expand your data sources.

How prediction works

• Prompt construction: The LLM receives a prompt telling it to act as a financial analyst and predict stock ratings.

• Example answer: The prompt can include an input-output example so the model knows the expected format.

• Future dates: The model is asked to identify which future months correspond to each prediction horizon.

• Explanation: The model is asked to explain the reasoning behind its ratings.

• Final output: The model then produces ratings for the specified future horizons.

How hallucination checks work

• Date check: The system asks the LLM to calculate the future dates tied to each horizon.

• Verification logic: If the model gets the dates wrong, that is treated as a warning sign about the reliability of the rating.

• Explanation check: The model’s explanation is used to see whether the rating is actually supported by the input data.

• Chain-of-verification: The system uses these intermediate steps to catch cases where the model may be producing unsupported answers.

How training and fine-tuning work

• Prompt-label pairs: Training examples pair a prompt with a correct rating label.

• Ground-truth labels: The labels are based on future stock performance, not just analyst opinions.

• Forward-return quintiles: Future returns are divided into quintiles and mapped to rating categories.

• Loss function: The system computes cross-entropy loss between the model’s predicted rating and the ground-truth rating.

• LoRA fine-tuning: The model can be fine-tuned using low-rank adaptation, which updates smaller added matrices rather than retraining the full LLM.

• Validation: The system can split the data into training and validation sets to test whether fine-tuning improves performance.

How the system checks whether the prediction was right

• Forward returns: The system looks at how the stock actually performed after the prediction date.

• Peer comparison: The stock’s return is compared with other companies over the same period.

• Sector-relative return: The company’s return can be adjusted against sector performance.

• Rating match: If the stock’s future-return quintile matches the model’s rating category, the prediction is treated as correct.

• Error measurement: Mean absolute error is used to measure how far the predicted rating was from the ground-truth rating.

What the results show

• LLMs did better in shorter-term tests: The application says the LLM may perform better on short-term predictions, while analyst errors declined over longer horizons and were slightly better in the 18-month period.

• Fundamentals mattered most: The best-performing setups were the ones using fundamentals, especially fundamentals plus sentiment.

• News alone helped less: News summaries and sentiment by themselves did not outperform the fundamentals-based setup.

• Sentiment added only modestly: Fundamentals plus sentiment performed slightly better than fundamentals alone.

• News may skew positive: The results suggest that news-derived inputs may push the model toward more positive ratings.

• Short-term versus longer-term signals: News appears more useful for short-term predictions, while fundamentals appear more useful across the main 3-, 6- and 12-month horizons.

AI Street Patent Review Tracker

Paid subscribers can download the Excel file below, which uses AI-assisted review to identify 300 patent applications published this year that appear tied to AI in trading and investing.

Read more

NEWS

Compute Is Getting a Futures Market

Global tech companies have announced about $740 billion in AI-related spending for 2026, according to Morgan Stanley. That’s about what it cost to build the entire U.S. interstate highway system (in today’s dollars), which took 30+ years.

All this money is chasing “compute,” the computing capacity needed to run AI models.

The problem: What is a standard unit of compute?

It’s not a barrel of oil, or a megawatt-hour or even the number of GPUs you have.

Right now, there’s no agreed-upon definition of what a unit of compute is across different chips, different data centers and different workloads.

Eventually, there will be. Compute may seem amorphous, but there’s precedent here. NIST, the U.S. standards setter for technology, has helped set standards for cloud computing, cybersecurity and even time itself with atomic clocks. Last summer, the agency released an early blueprint for AI testing and evaluation standards. But this is a long standard-setting process.

And buyers need compute today, so they have to call around to get a sense of pricing. This opacity creates friction. (It also creates an environment for bad deals. Simeon Bochev, the former CEO of Compute Exchange, told me last fall that he knew of companies being overcharged for compute by as much as 40%.)

A NOTE FROM OUR SPONSOR

AI doesn’t have a model problem.
It has a data problem.

Carbon Arc is building “a refiner” for the data layer underneath AI, CEO Kirk McKeown told me in February.

Most institutions are investing heavily in model performance while the data infrastructure underneath remains fragmented, expensive, and slow to access.

Carbon Arc’s Lenses puts 150+ institutional-grade data assets directly into your AI workflow. Card spend, foot traffic, payroll signals, medical claims, and more.

Queryable in plain English. No procurement cycle. No annual contract. No data team required.

The data infrastructure serious institutions run on, now available from $20/month.

Try Lenses with code AISTREET30 for 50% off your first month.

Standardizing Compute

This week, the compute market got two new efforts to make pricing more transparent.

From Bloomberg:

US derivatives exchange CME Group Inc. and index provider Silicon Data are teaming up to create a futures market for computing power, a key source driving the AI boom.

The futures will help traders, financial firms, AI builders and cloud providers manage volatility and price swings, according to a statement Tuesday. Indexes from market-intelligence firm Silicon Data will help underpin the products. The project is still pending regulatory review.

Computing power, also known as compute, has been in high demand as AI companies use it to power their systems. BlackRock Inc. Chief Executive Officer Larry Fink said last week that a new asset class will likely be buying futures of compute given the shortage and high demand.

For more on Silicon Data, see my interview with CEO Carmen Li from July:

Separately, former FTX US president Brett Harrison said on LinkedIn his trading infrastructure startup, Architect, has launched compute futures on AX, its derivatives exchange. The contracts track Nvidia H100 and H200 rental prices using indices from Ornn, a compute pricing data provider.

And for all this talk of “compute,” prices on the secondary market have jumped almost 50% since the end of April for Nvidia’s H100 chip.

I asked Ornn co-founder Wayne Nelms what was driving the recent increase. He pointed to demand from large AI companies running models for customers, rather than training them from scratch.

The price increase made me think of two big predictions made by well-known investors.

First: Don Wilson, founder of DRW, told the WSJ in January 2025 that annual spending on compute would exceed annual spending on oil within 10 years. The prediction came just after DeepSeek released an efficient open-source model that triggered a selloff in AI-linked stocks.

Second: Bridgewater’s co-CIO Greg Jensen and former AIA Labs Chief Scientist Jas Sekhon (who’s now at Google DeepMind) wrote at the end of last year that corporate panic would drive AI spending higher.

AI spending “is currently being driven by a small number of leading AI players recognizing the incredibly transformative power of AI. The next phase will come when a major business outside of the AI ecosystem realizes that its entire business model is about to collapse due to pressure from an upstart competitor using AI (as occurred with Amazon disrupting Barnes & Noble).”

These are not mealy-mouthed predictions from random AI hype men. They come from serious investors. And for now, the market is moving in their direction.

Turning Trading Chats Into Market Data

Traders in OTC markets still negotiate prices in Bloomberg chats or Symphony.

Traders are making markets while jumping between conversations across banks and brokers, trying to keep track of executable prices as they move.

Institutional traders can receive hundreds of broker messages per hour and miss up to 80% of them, costing firms millions per trader per year in mispricing and missed opportunities.

This is sort of known in the industry. That is, there’s a certain amount of, let’s call it, slippage in the market given the friction in the way information moves.

The problem has been surfacing the relevant information. As I’ve written a lot around here, AI is well suited for this kind of problem: organizing messy information.

TwoWay Finance says it uses LLMs plus deterministic code to turn broker-trader chats into structured market data. The Paris-based firm announced a €1.5 million pre-seed round this week led by welovefounders. TwoWay runs locally, so banks do not have to worry about sensitive trader chats leaking to an outside model provider.

The company’s pitch is broader than one asset class. By leveraging AI, TwoWay is trying to become an intelligence layer for chat-driven OTC markets, organizing prices like exchange order books do.

Share AI Street

Goldman Quant Exec Says AI May Make Markets Less Efficient

AI may be creating new market inefficiencies by pushing investors toward the same trades, according to Osman Ali, global co-head of Quantitative Investment Strategies in Goldman Sachs Asset Management.

If you ask these models the same type of question, they are going to give you the same type of answer, which will cause investors to pile into the same type of securities, which will cause markets to move in a direction that becomes predictable in terms of its reversion.

Ali said on the on the bank’s podcast.

Ali said his team often uses smaller models and fine-tunes them for specific tasks, including sentiment analysis of Japanese corporate disclosures.

Another quote that stood out to me:

“More than 50% of what we think drives the stock’s return over the next 12 months is not the fundamentals of the business,” Ali said. “It is what the market thinks about it.”

More here:

• Goldman Sachs Quant Chief Says AI Could Make Markets Less Efficient Traders Magazine

ROUNDUP

What Else I’m Reading

• HRT Notches Record $6.4 Billion Quarterly Markets Haul BBG

• Hedge Funds Are Making a Killing in the ‘Golden Age’ of AI Hardware WSJ

• Top Wall Street dealers join bond trading platform LTX Finextra

• Musk’s xAI Races to Get Wall Street Firms to Use Grok Chatbot BBG

Back in New York Next Week!

I’ll be back in New York next week to attend STAC Summit on May 20. Sessions include discussions on memory bottlenecks in inference, AI research at BlackRock, extracting structured data from SEC filings with LLMs, and deploying models into trading systems and engineering workflows. Registration is free for end users. Come say hi!

I’ll also be at the Women in Quant Finance conference the next day, May 21.

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

AI Street reaches institutional investors, C-suite executives and Big Law attorneys at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Email sponsors@ai-street.co for more details.

CALENDAR

Upcoming AI + Finance Conferences

• STAC Summit - May 20 • NYC

  Trading and analytics infrastructure, applied AI for research and execution, scaled deployment, and benchmark-driven insights across quant workflows. ← I’m attending.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

• Women in Quantitative Finance - May 21 • NYC

  Quants discussing current work in asset pricing, trading, risk, and portfolio construction. ← I’m attending.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• connect with me on LinkedIn, or

• send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

NEWS

LLMs Make Poor Stock Pickers

Two stories this week looked at how ChatGPT struggles with stock picking. My question is: why would you expect it to be good in the first place? These are large language models, trained on text, not financial data.

That said, these stories are important. There’s a lot of confusion about what AI can actually do.

The WSJ’s Gunjan Banerji tested ChatGPT as a hypothetical adviser for a $1 million portfolio and found it could explain risks, but struggled with actual investment calls. It gave a reasonable long-term allocation, but made a basic arithmetic error, drifted into market timing and picked a trade-war stock basket that rose about 5.5%, trailing the S&P 500’s roughly 8% gain.

Banerji highlighted the annoying, sycophantic part of AI that’s an investing risk:

“At times, it felt like ChatGPT responded with what I wanted to hear.”

Bloomberg’s Justina Lee took the question one step further, looking at trading competitions that pit major LLMs against each other, a topic I wrote about in November.

In Nof1’s Alpha Arena, eight models including Claude, Gemini, ChatGPT and Grok traded US tech stocks with $10,000 each across four competitions. The results were mostly ugly: the overall portfolio lost about a third of its capital, and only six of 32 model results finished in profit.

“LLMs can’t really make money by themselves,” said Jay Azhang, founder of Nof1.

Ashwin Paranjape, founding AI lead at Samaya AI, told me that LLMs can gather financial data, but stock picking requires higher-order skills: judging materiality, forecasting metrics and connecting signals across industries.

“Beating the market relies on information and reasoning asymmetry,” he said. “Eventually picking stocks will look like: ‘My AI knowing what I know, beats your AI knowing what you know.’”

These models are getting better, and there are some early results suggesting multi-agent setups can perform better than a single model acting alone. But that still does not make a chatbot a trading system.

Training models directly on market data to find signals is different. But that is a much harder, more expensive problem than asking a $20 chatbot what stocks to buy.

A NOTE FROM OUR SPONSOR

When your agent reads Bloomberg or Reuters, is it finding an edge?

On Tesla’s Q1 earnings, Goldman held at $375, TD Cowen reiterated Buy at $490, JPMorgan stayed at $145. You and your competitors are reading the same call.

The dispersion across sell-side targets. The reasoning behind each one. The hedge buried in the fifth paragraph of an operator quote.

That’s where the analytical signal is. It lives in the paragraphs your agent isn’t getting.

Typical retrieval looks fine. The agent doesn’t know what it’s missing, and neither do you.

Seltz returns full context in hundreds of milliseconds, every result traceable to source. Built for workflows where deep research matters more than the headline.

If you’re running agents on financial news, Seltz will run an eval on your setup.

Email CEO Antonio Mallia at antonio@seltz.ai or ask me for an introduction.

AI Is Narrowing the Hedge Fund Tech Gap

Craig Whiting, a hedge fund tech headhunter, wrote about recent conversations he’s had with four different engineers across Wall Street. They said they are writing less code than they did a year ago and increasingly work alongside tools like Cursor and Claude Code to draft code, review output and catch bugs.

The bar at any firm worth working for is now: you write good prompts, you read AI output critically, you know when to override it.

Whiting also points out how many firms are getting value out of relatively straightforward use cases like cutting down email volume by structuring unstructured text, an unsexy topic that I’ve written about before.

He also mentioned how AI is making it easier for smaller shops to compete:

The new reality: the gap between a £2bn credit fund and a $60bn multi-strat is collapsing because the tooling is cheap and the workflow is portable. Cursor is cheap. Claude Code is cheap. The bottleneck is not budget anymore. It is engineering culture, leadership willingness, and how legacy your stack is.

Craig Whiting

What Four Hedge Fund Engineers Told Me About AI This Fortnight

The capability gap is closing. The work is changing. How you communicate your work is critically important…

Read more

2 months ago · 1 like · Craig Whiting

Anthropic’s Busy Week

Anthropic has had a busy week announcing three different initiatives across Wall Street:

Anthropic and FIS Target AML

Anthropic announced it’s embedding Claude inside bank compliance departments through a partnership with fintech giant FIS, starting with anti-money laundering. Today, investigators spend most of their time manually pulling records from disconnected systems before any analysis can begin — the agent handles that assembly automatically, then flags cases by risk level. FIS claims it compresses investigations from days to minutes; BMO and Amalgamated Bank are the first pilot customers, with broader availability planned for H2 2026.

Anthropic Forms $1.5B JV with Blackstone, Goldman for PE Portfolio Companies

Anthropic is forming a $1.5 billion joint venture with Blackstone, Hellman & Friedman, Goldman Sachs, and several other Wall Street firms to embed Claude inside mid-sized companies — particularly PE portfolio companies — that can't staff frontier AI deployments on their own, according to the WSJ. Anthropic, Blackstone, and H&F are each putting in roughly $300 million; Goldman $150 million; General Atlantic, Apollo, Leonard Green, GIC, and Sequoia rounding it out. OpenAI is reportedly building a rival structure.

Anthropic Releases Ten Pre-Built Finance Agents

Anthropic released ten pre-built agent templates for financial services work — pitchbooks, KYC screening, month-end close, earnings review — deployable as plugins or as autonomous scheduled jobs on the Claude Platform. Claude also now runs inside Excel, PowerPoint, and Word via Microsoft 365 add-ins, carrying context between applications. Eight new data connectors went live alongside a Moody’s MCP app covering more than 600 million companies; Claude’s broader finance connector ecosystem includes FactSet, S&P Capital IQ, and PitchBook.

Merger Arb Funds on Using AI

The FT has a story on how merger arbitrage hedge funds are using AI to get a quicker read on deal dynamics by reading dense legal documents.

Traditionally, reading through the complex, lengthy deal documents — which often stretch to over 100 pages — would take an investment professional over an hour. Even a quick review would take 15 to 20 minutes. But the use of AI has now reduced the process to seconds.

“We think of AI as a very fast, very thorough intern who is brilliant at analysing big datasets,” says Daniel Caplan, chief executive of London-based Sand Grove.

This is the kind of use case that makes sense to me. Asking ChatGPT or Claude: “Where are the most important disclosures I should focus on in this document?” An actual merger arb specialist would ask a more sophisticated question than this, but you get the idea.

ROUNDUP

What Else I’m Reading

• BMO Turns to AI and Quantum Computing to Predict Earthquakes BBG

• Former Citadel Chief Technology Officer Joining Motive Partners BBG

• Lloyds in tie-up with Google to build AI agents City AM

• Subquadratic claims 1,000x AI efficiency gain with SubQ model VentureBeat

• Bessent Warns of Threat of AI-Powered Bank Account Hacks PYMTS

Back in New York

I’ll be back in New York to attend STAC Summit on May 20. Sessions include discussions on memory bottlenecks in inference, AI research at BlackRock, extracting structured data from SEC filings with LLMs, and deploying models into trading systems and engineering workflows. Registration is free for end users. Come say hi!

I’ll be at the Women in Quant Finance conference the next day, May 21.

Share AI Street

This Week in AI Street

CALENDAR

Upcoming AI + Finance Conferences

• AI in Financial Services – May 14 • Chicago
  Practitioner-heavy conference on building, scaling, and governing AI in regulated financial institutions.

• STAC Summit – May 20 • NYC

  Trading and analytics infrastructure, applied AI for research and execution, scaled deployment, and benchmark-driven insights across quant workflows. ← I’m attending.

• AI & RegTech for Financial Services & Insurance – May 20–21 • NYC

  Covers AI, regulatory technology, and compliance in finance and insurance.

• Women in Quantitative Finance – May 21 • NYC

  Quants discussing current work in asset pricing, trading, risk, and portfolio construction. ← I’m attending.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• Connect with me on LinkedIn, or

• Send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

Kim Posnett, the co-head of investment banking at Goldman, argued in the FT last year that AI may turn overlooked corporate data into a newly valuable asset:

AI models are only as good as the data they’re trained on. Hard-to-replicate, legacy data is more valuable in the age of AI. And legacy companies are generally the ones with the legacy data. Many corporations are sitting on valuable intellectual property and, I suspect, don't even know it.

But you don't need decades of data. Scale works. Revolut, the UK-based neobank with 70 million customers across 40 countries, has been collecting billions of data points. Its users generate a continuous stream of timestamped card transactions, peer-to-peer transfers, in-app navigation events, and communications.

Revolut researchers and Nvidia say they have used that stream of banking activity to train PRAGMA, a foundation model for financial event data. The model is designed to analyze a user’s event history of transactions, app activity, communications and profile data, allowing one underlying model to be adapted for tasks such as credit scoring, fraud detection and product recommendations.

A NOTE FROM OUR SPONSOR

When your agent reads Bloomberg or Reuters, is it finding an edge?

On Tesla’s Q1 earnings, Goldman held at $375, TD Cowen reiterated Buy at $490, JPMorgan stayed at $145. You and your competitors are reading the same call.

The dispersion across sell-side targets. The reasoning behind each one. The hedge buried in the fifth paragraph of an operator quote.

That’s where the analytical signal is. It lives in the paragraphs your agent isn’t getting.

Typical retrieval looks fine. The agent doesn’t know what it’s missing, and neither do you.

Seltz returns full context in hundreds of milliseconds, every result traceable to source. Built for workflows where deep research matters more than the headline.

If you’re running agents on financial news, Seltz will run an eval on your setup.

Email CEO Antonio Mallia at antonio@seltz.ai or ask me for an introduction.

What Revolut Did

PRAGMA was trained on 26 million anonymized user records from 111 countries, covering 24 billion events, according to a research paper posted to arXiv. The model is not a chatbot. It is designed to make predictions from banking histories, not generate text.

“Most ‘foundation model for finance’ discussions still default to text. But bank data is not text,” said Revolut’s head of AI Pavel Nesterov on LinkedIn.

His point is that financial activity has its own structure. Customers generate long sequences of transactions, app actions, communications, trading activity and profile changes. Turning all of that into text for a generic language model, Nesterov wrote, means “you lose too much structure and waste too many tokens.”

PRAGMA keeps the records closer to their original form. Each event is represented by what happened, the value attached to it and when it occurred. A card payment, for example, can include the transaction type, amount, currency, merchant category and time. The model then looks for patterns across long sequences of customer activity.

The authors say PRAGMA beat Revolut’s internal task-specific baselines across credit scoring, external fraud detection, product recommendation, communication engagement, recurrent-transaction detection and lifetime-value prediction.

Results

The biggest reported gains came in credit scoring and customer communications. Compared with Revolut’s existing models, PRAGMA improved one credit-scoring measure by 130% and one customer-communications measure by 79%. It also improved fraud recall by 65% and product-recommendation performance by 41%.

For Revolut, the operational goal is to reduce the need for separate models for every use case. Nesterov said the company is trying to move away from “a separate model stack for every narrow use case” and toward one shared model that can be adapted for tasks such as credit scoring, fraud detection and product recommendations.

PRAGMA fell short on one task: anti-money laundering, where it significantly underperformed Revolut's existing system. The authors say that is because money-laundering detection often depends on relationships among accounts, counterparties and transaction networks. PRAGMA analyzes one customer history at a time.

Revolut joins Netflix and Stripe, which both trained models on their own internal data. Netflix built a foundation model on hundreds of billions of user interactions that now underlies its personalization across search and recommendations. Stripe’s payments foundation model, trained on tens of billions of transactions, increased its detection rate for card-testing attacks from 59% to 97%.

That’s the fun part of AI for me. It reveals patterns that were always there but we didn’t have the computing power to see.

Thanks for reading!

I’m always happy to receive comments, questions, and feedback.

• Connect with me on LinkedIn, or

• Send an email to matt [at] ai-street.co

Manage how often you receive AI Street

Update Email Frequency

I’ve found myself searching through AI Street archives to pull together what I’ve reported on how hedge funds and market makers use AI. So, I created a running tracker that puts it in one place, combining news stories, regulatory filings, and some of my own reporting.

(If you think I should do the same for private equity firms or sovereign wealth funds, let me know. There is, in fact, a human behind the text you’re reading.)

I think of hedge funds and market makers as using AI in two main ways. I’ve put them in the same broad category because, as the FT has written, the two are converging.

The first way is straightforward: using large language models from frontier labs for familiar tasks like summarizing documents and surfacing ideas, plus more sophisticated workflows such as Man Group’s use of AI to generate and test trading signals.

A NOTE FROM OUR SPONSOR

When your agent reads Bloomberg or Reuters, is it finding an edge?

On Tesla’s Q1 earnings, Goldman held at $375, TD Cowen reiterated Buy at $490, JPMorgan stayed at $145. You and your competitors are reading the same call.

The dispersion across sell-side targets. The reasoning behind each one. The hedge buried in the fifth paragraph of an operator quote.

That’s where the analytical signal is. It lives in the paragraphs your agent isn’t getting.

Typical retrieval looks fine. The agent doesn’t know what it’s missing, and neither do you.

Seltz returns full context in hundreds of milliseconds, every result traceable to source. Built for workflows where deep research matters more than the headline.

If you’re running agents on financial news, Seltz will run an eval on your setup.

Email CEO Antonio Mallia at antonio@seltz.ai or ask me for an introduction.

GPU as Edge

The second way funds are using AI is training their own models on financial data. This requires orders of magnitude more computing power – on par with frontier labs like OpenAI and Anthropic – but as HRT’s Marc Khoury said at an industry conference last year, larger models trained on more data keep improving.

There’s a growing body of research that suggests that in each field, be it weather, payments, or market events, there may be an underlying “language” in domains like weather, payments and market microstructure that transformers are unusually good at idenifying.

In other words, if you take a massive amount of weather data, train a transformer model on it, you generally get better results than previous state-of-the-art models. I’m not suggesting that transformer are oracles, but I think it is important to highlight how fundamental a shift this is.

We’ve gone from a world where humans design models with top-down deductive logic. That approach is being surpassed by bottom-up, pattern-matching, inductive logic. And the power of these models shows no signs of abating.

Which raises the question: do the firms with the most compute and the most data win?

And as with all of these massive, billion-parameter models, no one knows how they really work on the inside. These models are grown, not built, as Anthropic CEO Dario Amodei is fond of saying.

This is a longer preamble than I anticipated, but I hope it gives you a sense of the landscape.

Share AI Street

Man Group

AI is helping the world’s largest listed hedge fund find new investment strategies.

Man Group’s quant equity unit uses an internal tool, called AlphaGPT, to generate, code and backtest trading ideas, mimicking how researchers develop new trading signals. AI in investing here looks less like a single model answering questions and more like a small organization. For trade ideas, AlphaGPT uses a workflow that proposes signals, writes code, runs backtests, and then sends the output into Man’s standard human review process.

The big picture: AI gives them scale to test more ideas. Man Group announced a partnership with Anthropic in February to use Claude and work with Anthropic engineers on AI applications across the firm, with alpha generation as the primary focus.

I spoke with Ziang Fang, Senior Portfolio Manager at Man Numeric, in December about how the system works in practice. The core problem AlphaGPT is solving: there’s been an explosion in data availability, and no one can realistically go through thousands of alternative datasets, many of which are unstructured. The system processes that volume and proposes hypotheses. Humans validate them.

“If you didn’t know whether a signal came from AI or a human, you probably couldn’t tell. The main difference is formatting. The AI output is more consistent,” Fang told me.

“The system has produced signals that meet our standards and pass the same evaluation thresholds required for human-generated research,” Fang wrote in November.

“Along the way we ran into a lot of issues — hallucination, lookahead bias, multiple testing, and many other things,” Fang said. The firm uses prompt controls, validation checks, and human review to limit errors and prevent p-hacking.

Hudson River Trading

Hudson River Trading is building foundation-style models trained on decades of global market data, applying techniques similar to those used in frontier language models for automated trading.

The firm is training these models on more than two decades of data spanning equities, futures, and cryptocurrencies, totaling over 100 terabytes. That translates into “something like trillions of tokens, in the same realm as what you train frontier language models on,” said Marc Khoury, an algorithm developer at HRT, speaking at an academic conference last summer.

HRT’s goal is to model markets as sequences of interactions. Much of the predictive signal lies in how order book events evolve over time, especially during fast-moving conditions. “As I increase the model size, the model continues to improve,” Khoury said — the same scaling pattern seen in large language models.

HRT is responsible for about 10% of all US equity volume. HRT’s AI Labs team, HAIL, says deep learning is core to the firm’s trading, and that HRT has spent more than a decade integrating AI research and infrastructure into its trading strategies. HRT is a proprietary trading firm and doesn’t file an ADV, so there’s nothing to cross-reference on the regulatory side.

And to my earlier aside about whether the firms with the most compute and data outperform their competitors — here’s HRT’s data center inside a mountain:

Bridgewater

In March 2025, Bridgewater CEO Nir Bar Dea said its $2 billion AI fund is generating “unique alpha uncorrelated to what our humans do” at a Bloomberg conference. The fund is delivering returns “comparable” to the firm’s human-led strategies, Bar Dea said. No specific figures were disclosed.

The fund is run by Co-CIO Greg Jensen and uses Bridgewater’s proprietary technology with models from OpenAI, Anthropic, and Perplexity. Bridgewater formed its Artificial Investment Associate (AIA) Labs division in 2023. The AIA serves as the primary decision-maker in the fund, while human professionals oversee risk management, data acquisition, and trade execution.

In its ADV, Bridgewater describes AI and machine learning as examples of “new sources of alpha” the firm is developing. It’s a description of Bridgewater’s strategy and capabilities, not a disclosure of realized fund performance.. It’s a description of what Bridgewater is building toward, not what it’s delivered. Bar Dea’s claim at Bloomberg is the stronger statement, and even that came without figures.

Jane Street

Jane Street says deep learning is “the future of quantitative trading.” The firm, which reported record trading revenue in 2025, builds neural-network models that drive its trading strategies, along with the infrastructure needed for training and inference.

Jane Street says it has tens of thousands of high-end GPUs, more than 1 exabyte of current storage, and about $400 billion in daily filled dollars. It trades on more than 200 electronic exchanges and venues, making it one of the world’s largest market makers.

Last month, Jane Street committed about $6 billion to use CoreWeave’s AI cloud platform and made a separate $1 billion equity investment in the company. The agreement gives Jane Street access to next-generation compute across multiple facilities, including Nvidia’s Vera Rubin technology.

CoreWeave said Jane Street has relied on its infrastructure since 2024 to train and scale proprietary models. Jane Street’s head of quant research, Craig Falls, said CoreWeave provides the GPU infrastructure and technical support needed for the firm’s machine-learning and research workloads.

Bloomberg reported that Jane Street generated $39.6 billion in trading revenue in 2025, surpassing major Wall Street banks with only about 3,500 employees. That doesn’t prove AI or compute caused the trading haul. But it does show why access to GPUs, storage, and low-latency model infrastructure has become strategic.

More firms detailed below, including Citadel, XTX, AQR, Viking and Millennium.

Read more