AI Street : Research

Citadel Securities, Two Sigma and Bridgewater Detail AI Use Cases

Matt Robinson — Tue, 20 Oct 2026 17:30:49 GMT

Hey, I’m Matt. I’m a former Bloomberg News reporter, and you’re reading AI Street, where I report on how Wall Street uses AI. On Tuesdays, I highlight novel research, emerging use cases, and expert interviews.

Citadel Securities, Two Sigma, Bridgewater and other financial firms detailed AI use cases beyond basic summarization and document search at a conference in New York last week.

Examples included flagging internal projects at risk of running into trouble at Citadel Securities, proactively suggesting research questions at Bridgewater and automatically tracking who owns each AI agent and which tools it can access at Two Sigma.

Use Case Tracker

I’ve started a database tracking what financial firms have publicly said about how they’re using AI. So far, it includes 79 entries from 40+ companies, drawn from interviews, conference talks and technical posts, with links and key takeaways.

Paid subscribers can access the tracker by clicking the link below:

We’ve previously covered how firms are tailoring AI to their investment needs:

Below are five examples from the AI Engineer conference. More use cases are in the AI tracker.

PRESENTED BY BRICKROAD

The Best Dataset Isn’t For Sale — Yet

Months before Google agreed to pay $10 million for Spirit Airlines’ data, Brickroad had already flagged it.

The company’s AI agents scan news, earnings reports, app-store rankings and changes to terms of service for signs that a company is producing valuable data — and might be willing to license it. When a source looks promising, Brickroad contacts the company, secures a sample and helps the buyer evaluate the data before committing to a deal.

Citadel Securities Builds Internal ‘World Model’

Citadel Securities combines periodic project reports with Slack messages, meeting notes and code changes to track how projects are progressing.

In the talk, João Wenzel, chief operating officer for AI at Citadel Securities, said the team built up three years of historical records to test whether the system could flag troubled projects before their official status deteriorated.

In a backtest presented, 18 of 100 projects rated green and flagged by the system turned yellow or red within three months, versus 10 of 100 unflagged green projects.

Splitting the material among specialist agents, then having other agents review their findings, more than doubled useful project updates compared with a single-model pass.

Recording · 6:38:33

Why are firms that have historically kept a low profile now more public?

A big reason why firms are now more public than they were, say, 10 years ago, (or maybe even five years ago) is the need to recruit AI/ machine learning talent. Frontier labs are fishing in the same PhD pool of math/ computer science/ engineering folks that were previously mostly sought after by quant firms.

Two Sigma Tracks Who Owns Which Agents

As the number of “AI employees” grows, firms need to know who owns them, what they do and which systems they can access.

John Feminella, senior vice president of architecture at Two Sigma, described automatically registering agents created through approved internal frameworks. The registry records each agent’s owner, purpose, tools and endpoints, making agents visible and reusable across the firm.

Feminella also argued that controls should reflect what an agent does. A tool summarizing support tickets should receive different scrutiny from one deploying production software. Feminella said he was speaking in a personal capacity.

His broader framework included restrictions on agent access, shared evaluations and checks on agents’ actions as they run.

Recording · 0:23:55

Bridgewater’s AI Suggests What to Research Next

Bridgewater is building an investor workflow where AI proactively suggests questions worth investigating.

The system draws on Vault, which contains more than 100 million documents from 135 external sources, alongside internal research and discussions.

In the talk, Sam Green, team lead for applied AI at Bridgewater’s AIA Labs, described agents presenting investors with research topics, summaries and supporting evidence. Investors decide whether to dig deeper, and findings from their work with the agents are saved back into Vault for future research.

The repository also supports AI-generated macroeconomic forecasts and investment signals.

Recording · 7:22

Apollo Uses AI to Prepare Loan Portfolios

At Apollo’s Redding Ridge Capital, preparing a proposed portfolio of 200 to 400 credits previously took two or three days, involving legal documents and Excel models.

Jack McDonald, a principal at Apollo and chief operating officer of Redding Ridge, said the firm now uses Rowspace to combine document libraries, trade and compliance systems, and current pricing to prepare portfolios in minutes.

McDonald said the process gets portfolios roughly 80% to 90% complete, leaving portfolio managers to apply their judgment and refine the result.

He also described automating reports on loan purchases, sales and settlements that previously occupied two or three employees for hours each week. Staff now check the output daily.

Recording · 7:50:03

Coatue Maps Which Research Tasks AI Should Handle

Coatue’s head of AI, Frank Long, described breaking investment research into tasks best handled by software, machine learning, AI agents or people.

His example: processing Korean-language news about semiconductors and AI. That involves collecting, filtering and translating articles, then checking the information against the firm’s research and views.

Long said Coatue has preserved decades of notes on investment decisions and outcomes, providing internal context for AI systems.

The goal is better work at each step. A simple rule may beat a large model, while assessing portfolio implications may still require a person.

Recording · 6:47:46

If you’ve read this far, I’d like to hear what you think of this format. Reply with “quick-hit” for this fast rundown of firms or “slow down” if you’d prefer a deeper dive into a particular use case.

Thanks for reading!

⁠⁠Telling AI to Act Like Different Investors Barely Changes Its Forecasts: Study

Matt Robinson — Tue, 13 Oct 2026 15:30:32 GMT

If you tell ChatGPT “Take on the persona of Warren Buffett,” you know you’re not getting the real Oracle of Omaha.

But you might think, as I did, that instruction helps get some of the billionaire’s investing judgment to come through in its answers.

New research is calling that assumption into question. Instructing an AI model that it is a “therapist, a rocket scientist or an elite money manager” does little to shape the model’s actual forecasts. The results may sound cogent, but they don’t really reflect actual judgment but “a continuation of the prompt rather than a stored belief.”

That last quote comes from a recent paper from the University of Chicago that tested how AI forecasts compared to those from humans. The short answer: not good.

JFYI: We’ve previously covered other shortcomings of AI in finance. Some examples: positional bias (the same information, presented in a different order, can produce different outputs), framing effects (the same odds stated two ways lead to different decisions), and, of course, biases inherent in the data the models are trained on. For more info see:

NOTE FROM OUR SPONSOR

The Best Dataset Isn’t For Sale — Yet

Months before Google agreed to pay $10 million for Spirit Airlines’ data, Brickroad had already flagged it.

You can’t really call up Warren Buffett and say “hey, Warren, pal, should I be buying banks in this high-rate environment?” But you can test AI predictions against humans who already provide their forecasts to the public: prediction markets and analysts’ earnings forecasts.

Here’s how University of Chicago professors constructed their research. (Full disclosure: I write for Chicago Booth Review, a publication of the business school.)

In prediction markets, they used event contracts, which are binary. Either the event happened or it did not. The authors don’t specify the exact kinds of events, but provide general topics such as economics, politics, sports and weather.

The AI model is asked the same questions dozens of times, but what changes is the persona. The AI is instructed to be younger or older, and more or less willing to take on financial risk. Here’s their prompt:

The brackets could be filled out from fixed choices—White, Black, Hispanic or Asian for race; ‘some college or less’ or ‘college or more’ for education; and low, moderate or high tolerance for financial risk, along with set choices for age, gender, income and politics.

All of these different permutations created 1,296 possible fake AI personas. The researchers didn’t test every single persona but rather 54 profiles for each event on a given day, drawing from the larger pool over time.

The researchers compared the model’s forecast with the current market price of the event, which shows the odds implied by the trades. In other words, how likely humans think an event will happen.

And as for the results: The different personas mostly agreed with one another. The AI assigned the outcome that eventually happened a 60.6% chance on average, compared with 84% implied by market prices—a gap of more than 23 percentage points. The AI stand-in remained much closer to a coin flip than the traders.

Behind the paywall: what happened when the researchers fed the AI a deliberately false market price—and how its forecasts stacked up against Wall Street earnings analysts.

Agents Think 3x Faster on Dedicated GPUs: Study

Matt Robinson — Tue, 06 Oct 2026 15:30:45 GMT

Wall Street’s Compute Arms Race Rolls on

Despite the high cost of Nvidia’s GPUs, with its most advanced chips estimated at about $55,000 apiece, trading firms can’t get enough of them.

Ron Minsky, co-head of Jane Street’s tech group, on a company podcast in May:

“These days we are in something like the range of like tens of thousands of GPUs and we will in not too long be in the range of hundreds of thousands of GPUs. And we think it’s like well justified by the business.”

Iain Dunning, head of AI at Hudson River Trading, on Bloomberg’s Odd Lots podcast in June:

“One of my greatest failures has been… predicting how many GPUs we would need,” Dunning said. “You’re constantly playing catch up.”

Álvaro Cartea, director of the Oxford-Man Institute of Quantitative Finance, said the firms would not be spending billions unless they expected the investment to pay off.

“However many billions these guys spend, it’s because they know that it’s very likely they’ll recover them,” Cartea told me. “These billions are a bet that the odds are on their side.”

A new STAC research report helps explain in part why trading firms want dedicated access to GPUs rather than relying entirely on external APIs: for some workloads, it can dramatically reduce the inference time involved in building and testing quantitative models.

NOTE FROM OUR SPONSOR

The Best Dataset Isn’t For Sale — Yet

Months before Google agreed to pay $10 million for Spirit Airlines’ data, Brickroad had already flagged it.

Benchmarking Agents in Quant Workflows

STAC, an industry group that develops and runs technology benchmarks for financial firms, tested a workload based on a real-world quant research session supplied by a large US market maker.

A Real-World Quant Research Workload

The original session involved an AI coding agent developing and evaluating short-term trading models using Bitcoin market data.

“We worked with one of the big quant trading firms to use a real-life example from their agentic quant research sessions,” James Corcoran, STAC’s head of AI, said in an email. “So we know this is representative of what some in the industry are doing.”

It was the first time STAC had benchmarked inference infrastructure for agentic workflows, Corcoran said.

STAC replayed 300 requests from the session using DeepSeek-V4-Pro, the same model used in the original session.
It ran the requests on a dedicated Lambda cluster containing 16 Nvidia B200 GPUs, then compared the results with DeepSeek’s public API.

Faster on Dedicated GPUs, but More Expensive

When STAC ran a single 300-request research session, the dedicated cluster completed the replay in about 25 minutes. The same requests took roughly 72 to 75 minutes through DeepSeek’s API.

“With this use-case, speed is the name of the game: how quickly can an agent build and test a quant trading model,” Corcoran said. “Time to task completion is extremely important.”

STAC then ran 8, 32 and 64 sessions simultaneously. The speed advantage narrowed and reversed as more sessions shared the fixed cluster.
At 64 simultaneous sessions, the dedicated cluster took about 93 minutes per session, while the API remained around 72 to 75 minutes.
At off-peak rates, the API was also cheaper per task. The dedicated setup offered greater speed for a single session, but at a higher cost.
One caveat: The test focused on how quickly the systems generated the agent’s responses. It excluded the time needed to execute code and train trading models.

This reminded me of what Matthew Dixon, co-author of Machine Learning in Finance, told me in May:

Brute force gets you a long way in finance. Most applications aren’t Rolls-Royce engineering problems. They’re bread-and-butter: get me a number, good enough, ballpark. AI fits that very well.

More speed means firms can discard weak ideas and pursue promising ones while competitors are still waiting for the same research cycle to finish.

Coming up for paid subscribers: More of my conversation with Oxford-Man’s Cartea on why compute is becoming the new latency race, why firms want models carrying their own “genetic code,” and whether billion-dollar infrastructure costs will leave only a handful of firms able to compete.

How Agents Learn to Manipulate Markets: Study

Matt Robinson — Tue, 29 Sep 2026 15:31:13 GMT

OpenAI, Anthropic, Meta, and Google have all recently reported their AI agents broke into real websites—in what were supposed to be simulations—to complete assigned tasks.

Unlike traditional software where engineers code black-and-white rules, AI agents have broad latitude to pursue their goals and can act in ways—as the examples above show—their designers didn’t expect.

The companies are now asking governments, the U.S. in particular, to protect the public from ~~themselves~~ further fallout.

This got me wondering: what does rogue AI look like in markets?

But this question turns out to be imprecise.

If I tell you to trade profitably in a basket of stocks, it’s (hopefully) understood that you will do so within the bounds of the law. A human knows, or should know, they’re crossing the line. Machines struggle to know where the line is.

In simulated trading tests, researchers have found AI agents pursue a number of illegal and questionable tactics. Agents learned to spoof orders, collude without communicating, and create and exploit price bubbles without being instructed to break the rules, according to academic research.

This goal-seeking behavior is driven by an AI training method called reinforcement learning, which teaches AI through trial and error, rewarding it for actions that achieve a goal.

Agents discover these manipulative tactics independent of each other. They’re not explicitly working together. Reinforcement learning “teaches them” to pursue more profitable trading strategies.

NOTE FROM OUR SPONSOR

The Best Dataset Isn’t For Sale — Yet

Months before Google agreed to pay $10 million for Spirit Airlines’ data, Brickroad had already flagged it.

I recently spoke with Wharton professors Itay Goldstein and Winston Dou, who, along with Yan Ji at the Hong Kong University of Science and Technology, have been studying how common AI training methods impact markets, specifically reinforcement learning (RL).

Here’s how Itay describes their recent research Financial Market Fragility in the Era of AI Planning:

“We are looking at pretty simple agents that are just doing reinforcement learning. They’re not communicating. They’re not writing messages to each other. They’re not explicitly saying, ‘Let’s join forces and do this and do that.’”
…
“They just find, over time, through repeated interactions, that collaborating, acting not competitively and creating bubbles are in their best interest.”

What they tested

The researchers built a simulated stock market where two AI traders competed alongside three groups of investors following preset trading rules. Some chased rising prices, some bought when prices looked cheap and sold when they looked expensive, and others were slower to react.

The two AI traders got information about changes in the stock’s value before everyone else, giving them an advantage similar to traders who process news faster.

Each time through the simulation, the AI traders could trade twice. That gave them a chance to learn how an initial purchase could move the price and make a later sale more profitable. They couldn’t exchange messages, but they could see price changes.

The researchers let each pair practice billions of times, then started over with a fresh pair, repeating the process 1,000 times for each market setup. Each pair needed roughly 10 billion to 200 billion trips through the simulation before its trading strategy stopped changing.

Results

Each AI trader earned an average of $1.48 million per simulated earnings cycle, 66% more than the $890,000 in the researchers’ benchmark where traders acted independently.

Again, this was a simulation.

The agents learned to buy aggressively, pushing prices above the stock’s fundamental value. That drew in investors programmed to chase rising prices. The AI traders then slowly sold into that demand in order to limit too much downward pressure on the stock.

Who lost? Trend followers bought at inflated prices before the stock returned to its fundamental value. The agents also learned the reverse strategy: push prices down, induce others to sell, then buy at depressed prices.

The trading made the market more volatile. In the first trading round, return volatility was 81% higher than in the benchmark where traders acted independently.

The outcome depended on who else was trading. When the researchers removed the trend followers, the AI traders still coordinated, but their coordination reduced price swings.

“[Collusion without communication] makes regulation, enforcement, even detection much more challenging.”

—Winston Dou

Enforcement

Again, these were market simulations, not evidence of shenanigans going on in live markets today. A recent ECB paper, which echoed many of the concerns of RL in trading markets, cited a 2022 Dutch regulatory study suggesting that reinforcement learning was not widely used by proprietary trading firms. The authors cautioned that adoption may have changed since. (For what it’s worth, I’ve not heard of any firms using reinforcement learning for trade execution.)

Bank of England Gov. Andrew Bailey raised a related point in July parliamentary testimony, describing agentic trading systems as “more talk than actuality.” But he pointed to behavior reported by developers of advanced AI systems as a reason to resolve questions about responsibility before adoption spreads. Such systems “learn to cheat and they learn to lie,” he said, adding that they “cover things up.”

So who is responsible when an AI trading agent goes off the rails?

While collusion without a paper trail makes it harder for regulators to detect, it doesn’t necessarily leave them powerless, according to Alessio Azzutti, a lecturer and legal researcher at the University of Glasgow.

Existing rules already require firms deploying trading algorithms to test, monitor and control them, he said. Some market-abuse sanctions can also apply without proof of direct intent. The harder cases would be those requiring investigators to establish human knowledge or intention when the trading strategy originated inside the system.

“The human specifies the end, but the agent discovers the problematic means,” Azzutti told me in an email. He argues that accountability should cover decisions about how systems are designed, trained, tested and supervised, including whether firms can intervene when something goes wrong.

As I’ve said often around here, we’re in early days in this new AI world and this topic is one I’m sure I will be covering in future posts on AI Street.

Behind the Paywall

Paid subscribers can read more of my conversation with Goldstein and Dou below. We discuss:

Which markets could be most vulnerable
How prices let agents coordinate without exchanging messages
What regulators could do to disrupt that behavior

Balyasny’s AI Beat Market Odds on Merger Deals

Matt Robinson — Wed, 16 Sep 2026 15:31:29 GMT

Hey, I’m Matt. I’m a former Bloomberg News reporter, and you’re reading AI Street, where I report on how Wall Street uses AI. On ~~Tuesdays~~ Wednesday this week, I highlight novel research, emerging use cases, and expert interviews.

Over the past few weeks, we’ve looked at how hedge funds are putting their own expertise into AI systems, rather than relying on off-the-shelf models:

Balyasny Asset Management joins this list. The $38 billion hedge fund fine-tuned a model on merger arbitrage deals, where investors bet on whether announced corporate deals will close, attract a higher bid, or collapse.

In M&A deals, the target company’s shares typically trade below the offer price, reflecting the risk that the deal might fall apart.

Merger-arbitrage traders buy the target’s stock after a deal is announced, sometimes while shorting the acquirer, to capture that gap. They earn it if the deal closes and take losses if it fails.

To judge those odds, analysts read through hundred-page merger agreements, assess regulatory regimes across different jurisdictions, and scrutinize shareholder bases and voting histories.

Balyasny researchers built an AI system that they say can help do that work. In fact, in the July paper outlining their research, they say they deployed it as a decision-support tool for analysts and portfolio managers.

A company spokesperson didn’t respond to requests for comment on whether the hedge fund is still using the model.

Subscribe now

How Balyasny Trained the Model

Since there is no readily available dataset of historical M&A deals, Balyasny built one.

The researchers assembled 1,648 public-company deals announced between January 2022 and December 2025. They excluded minority-stake purchases, asset sales, SPAC and smaller deals, such as those under $1 billion. (Where certain deal details were missing, they used AI agents to fill them in and reviewed incomplete information by hand.)

They split the deals by announcement date, using 1,244 deals announced through January 2025 for training and validation. The held-out test set contained 404 deals announced from February to December 2025. Across the full dataset, the system made 4,155 forecasts at different points in a deal’s life, including 1,115 in the test set.

Next, they built 12 AI research agents, each covering one part of a deal. They run in stages:

First

Ticker Resolution: Matches the buyer and target to their identifiers across financial databases.
Deal Card: Pulls the core deal terms, including structure, key dates, regulatory requirements, termination fees and voting thresholds.

Then, the detailed research

Filings Analysis: Reads merger agreements, proxy statements and other SEC filings for red flags, such as litigation risks, and for provisions that protect the deal.
Stakeholder Ownership: Identifies major shareholders on both sides, activist involvement and likely voting dynamics.
Market View: Reviews expert commentary and market sentiment about the deal.
Current Climate: Assesses the wider backdrop, including antitrust climate, interest rates, sector trends and geopolitics.
Regulatory Risk: Evaluates antitrust and other approval risks across countries.
Financing Risk: Checks whether the buyer can pay, including debt availability, credit ratings and balance-sheet health.
Precedent Mergers: Finds similar past deals and pulls their completion rates, timelines and risks.
Timeline Events: Lays out past and upcoming milestones, such as filings, shareholder votes and court dates.

Then, a check

Gap Analysis: Reviews everything collected so far for missing or overlooked information.

Finally, what’s next

Catalyst Tracker: Flags upcoming events that could shift the deal’s odds, such as regulatory decisions, shareholder votes and financing deadlines.

Merger-Arb Specialists Review Agent Work

For the training deals, the researchers built example reports for the model to learn from. First, an AI agent wrote a post-mortem of each deal, covering how it unfolded and which early signals turned out to matter. It was the only part of the system allowed on the open web, since the outcome was already known. Then GPT-5.1 combined the agents’ research with the post-mortem to write a report as of the forecast date. It could cite only information public by that date, but it knew which facts deserved the most weight.

Veteran merger-arb specialists spent months reviewing the agents’ output on historical deals they had covered and flagging gaps. (The paper doesn’t say how many specialists were involved or give a more precise timeline, only that their feedback helped refine the agents’ responsibilities, prompts and output formats.)

The researchers also set the probabilities the model would learn from. Training on final outcomes alone makes forecasts overconfident, especially early in a deal, according to the paper.

They then fine-tuned GPT-4o on those reports and probabilities.

Finally, they tested it. For each test-set forecast date, the fine-tuned model read the agents’ research and produced five forecasts. The system used the median probabilities, while GPT-5 checked the five written reports against source documents and combined them into one.

Results

Most deals end well for shareholders. In the sample, about 14% collapsed without a better offer. So a model could look accurate just by predicting every deal would go through. To prevent that, the researchers gave each collapsed deal more weight in the score than each completed one.

The fine-tuned model beat every method they tested.

Cloning CFOs

Matt Robinson — Tue, 08 Sep 2026 15:32:08 GMT

A big challenge in measuring how well AI predicts outcomes is look-ahead bias: quizzing a model on a test it's already taken.

AI providers generally don’t disclose their training data, so researchers are in the dark on whether a model has already seen the outcome they are asking it to predict.

But if the data has never been made public, you can be more confident that the model is actually making a prediction.

Duke has one such dataset. Its Fuqua School of Business has surveyed CFOs quarterly since 1996, with the Atlanta and Richmond Feds joining the project in 2020. The survey’s CFO Optimism Index has historically been a leading indicator of hiring and GDP growth, while research using the responses has also found that CFO expectations predict corporate investment.

The aggregate results are published. The individual executives’ answers remain confidential.

In a new paper, CFOs Meet LLMs, Duke finance professors John Graham and Campbell Harvey teamed up with Georgia State University professor Manish Jha to test whether an LLM could act as a digital twin for an individual executive.

The test: Could an LLM predict how a particular CFO answered in a particular quarter?

The researchers analyzed 6,075 survey responses from public-company CFOs between 2002 and 2025. In each response, the executive rated their optimism about the overall US economy on a scale from zero to 100.

GPT-5.4 was given the executive’s real name, company and ticker, along with details such as revenue, employment, foreign exposure and credit rating. It could search for company information, analyst reports and economic data available at the time, but was told to ignore anything published after the survey date. If a CFO had taken the survey before, the model also got up to 12 of their earlier confidential optimism scores.

What is a digital twin?

A digital twin is a computer model of something real, loaded with data so it can mimic how the original behaves. A digital twin of a jet engine might use temperature and vibration readings to predict when a component will fail. Bologna is building a twin of the entire city to test how changes in transportation, energy use and flood protection might play out.

The idea dates to NASA’s Apollo-era simulators. More data and computing power now make it possible to build twins of systems as complicated as cities, supply chains and, in this paper, a CFO.

Insurance companies are beginning to explore digital twins, according to Samarth Swarup, a University of Virginia researcher who builds digital twins of societies.

“You want to know: How many claims can we expect? Which households are more at risk depending upon how evacuation happens? And what can we do proactively to reduce that?” Swarup told me.

Results

The researchers ran each prompt three times and averaged the results. The model gave very similar scores each time, with correlations of about 0.96 across the runs.

Compared with the real CFOs, the model was slightly more pessimistic and made the executives look more alike than they really were, producing fewer very high or very low scores.

Across all responses, the synthetic scores explained 27% of the differences in CFO optimism. The relationship held after the researchers accounted for each executive’s usual level of optimism, conditions during the quarter and the CFO’s most recent answer.

“It is not asking whether the model can separate persistent optimists from persistent pessimists,” Jha said in an email. “It is asking whether it can tell when a particular CFO turns more or less optimistic than their own usual level.”

Related Research

The researchers accounted for persistent differences among firms, each CFO’s most recent answer and whatever was moving everyone’s mood that quarter. The model’s score remained statistically significant.

But it worked much better when more history was available. Among responses with no previous answers, the model explained 10% of the variation. That rose to 33% when one to three previous answers were available and 49% when more than three were available. The model was much better at extending a known executive’s past views than building one from public information alone.

“The surprising part for me was how much the conditioning did, more than the raw capability of the model,” Jha said in an email.

The model isn’t a stand-in for a human CFO. It does not have access to the private information that makes an executive’s view valuable: current orders, internal costs, financing discussions and changes in corporate strategy. The authors described the synthetic scores as a possible supplement to human surveys, not a replacement.

The study also reconstructed past responses from 2002 through 2025 rather than making new predictions in real time. The researchers plan to test the system against future survey rounds.

For now, the system looks more useful for extending an existing panel of known respondents than for creating thousands of synthetic CFOs from scratch.

TL;DR

GPT-5.4 could predict how individual CFOs rated the economy, but it worked far better when given their previous survey scores.

Behind the paywall

Further commentary from Manish Jha on where digital twins are useful in finance — research and policy — and why he wouldn't trade on one yet.

Inside 142,000 H-1B Filings From Wall Street Firms

Matt Robinson — Tue, 01 Sep 2026 15:30:46 GMT

As I’ve said a few times, AI is great at organizing messy data— the more tedious, the better. I’ve used local models on my Mac Mini to database ADV disclosures and patent applications and wrote about what I found:

You can also use AI to turn that cleaned-up data into an interactive dashboard—which is what I did with the H-1B filings.

These required Labor Department disclosures contain job titles, locations and wages for positions tied to H-1B sponsorship. I downloaded the data here and had OpenAI’s Codex standardize employer names, like “JPMorgan” and “J.P. Morgan,” identify duplicate case records and turn the data into a dashboard.

AI Street Dashboard

Paid subscribers can access the interactive dashboard using the password provided at the bottom of this email.

What the Filings Show

The dataset covers 141,941 unique filings from 91 finance firms, drawn from 26 quarterly releases spanning fiscal 2020 through fiscal 2026’s third quarter.

In the four quarters through June, H-1B filings from the 91 financial firms reviewed were roughly flat, declining about 0.9%. Nationwide filings fell 8.8% year over year in the first three quarters of fiscal 2026. (Nationwide H-1B-related applications peaked in fiscal 2022 and have remained below that mark since.)

Some important nuances: these filings are part of the H-1B sponsorship process, not evidence that someone was ultimately hired. One filing can cover multiple positions, and the salary figures show only the minimum base pay offered—not bonuses, equity or other compensation.

Three Firms Accounted for the Largest Increases

Fidelity, Citigroup and PayPal saw the biggest jumps in filings, adding 1,397 together. Excluding those firms, filings at the remaining firms declined about 8.2%.

The Banks Split

Citigroup and Morgan Stanley reported more filings. Bank of America, UBS and Deutsche Bank reported fewer.

(These are examples, not an exhaustive sector ranking. The dashboard and its download include every firm in the sample, including those with small changes or few filings.)

Payments Firms Moved in Opposite Directions

PayPal filed 23.1% more applications, while Visa and Mastercard filed 29.5% and 35.1% fewer, respectively.

AI Titles Only Make up 2%

Applications with explicit AI titles grew, but remained a sliver of the market, accounting for just 2% of filings, up from about 1% a year earlier.

JPMorgan and PayPal accounted for many of the explicit AI-title filings, with positions such as applied AI leads and machine-learning engineers. Bloomberg’s filings included a “Senior NLP Research Engineer.”

Meanwhile, 11 firms with at least 500 qualifying filings had no explicit AI titles in the archive. They include Goldman Sachs, Citigroup, Bank of America and Morgan Stanley. That does not mean those firms do no AI work: the data only captures job titles. Explore those firms.

What Firms Offer for AI-Titled Roles

Within employer-and-occupation groups containing both kinds of title, the raw median annualized offered wage floor was $170,000 for explicit AI titles and $145,000 for the comparison group—a $25,000 difference.

But much of that gap reflected differences in occupation, location and seniority. After accounting for those factors, the estimated difference narrowed considerably and should not be interpreted as a universal “AI premium.” See the fixed wage comparison and its limitations.

Data & Methodology

The analysis covers 141,941 H-1B filings linked to 91 selected finance firms. I standardized employer names and removed duplicate case records. Filings are not confirmed hires, and one filing can cover multiple positions. You can read the full methodology—including the employer-matching, AI-title and wage-analysis rules—on the dashboard.

Dashboard Access

D. E. Shaw, Point72 and Ares Hire to Put Investment Know-How Into AI

Matt Robinson — Tue, 25 Aug 2026 15:31:24 GMT

It’s still hard to get AI to approach a problem the way a particular firm would.

Last month, we discussed how Bridgewater taught a model to handle research tasks more like one of its own investors:

…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.

In other words, Bridgewater trained the model on its internal processes to give it institutional knowledge. And it’s not the only firm. A review of job postings at hedge funds, asset managers and private-equity firms shows others on Wall Street are trying to build similar firm-specific knowledge.

Firms are hiring AI specialists to codify how their investors work while retaining control of the data, workflows and feedback loops used to teach the AI. The firms still rely on frontier labs for broader use cases, as many of the partnerships we’ve covered show. But this goes beyond giving engineers access to coding tools.

Alex Gannes, founder of the investment management focused search firm, Banyak Group, described outside models and internal customization as complementary.

“A hedge fund is always positioning itself for an edge,” Gannes said. “Leveraging frontier labs and customizing AI solutions should ultimately support that edge.”

Investors are rightfully protective of that edge. And more broadly, companies are protective of their intellectual property and processes. This is a point Alex Karp, co-founder and CEO of Palantir, has been making in recent months.

“[Customers want] control over their compute, their models, their data stack and their alpha,” Karp told CNBC last month. “They want to know they own the means of production. It’s not being transferred to someone else.”

Karp is talking his book, but I tend to agree. Wall Street folks are a pretty paranoid bunch. I don’t see how firms ever get comfortable codifying how they work unless they know exactly where that knowledge lives and who controls it. (I’m not suggesting that frontier labs train—or would train—their models on customer data without permission. This is more of a cultural issue than a technical one.)

An AI-assisted review of more than a dozen current job postings at major hedge funds, asset managers and private-market firms shows firms pursuing three related goals: capture how their investment teams work, connect that knowledge to proprietary data and turn both into AI systems they control.

Related Coverage

The titles range from Fundamental Equities AI Engineer and AI Strategy Analyst to AI Credit Product Lead and Forward-Deployed AI Engineer. The use cases include idea sourcing, thesis monitoring, trade write-ups, portfolio construction, credit underwriting and investment-committee memos.

Behind those titles is a translation job: turning the judgment of portfolio managers, analysts and deal teams into prompts, retrieval systems, scoring frameworks and agents—and then determining whether the output is any good. The postings mention hallucination checks, audit trails, evaluating outputs without clean labels and getting direct feedback from investment teams.

Banyak Group’s Gannes thinks these teams could eventually support something broader than AI tools built for individual desks.

“A realistic goal in the hedge fund AI roadmap is to architect and scale an enterprise-level ‘hive mind’—linking autonomous AI agents into a single, coordinated cognitive network,” Gannes said. “Hiring a core team of applied AI engineers who can support every function within a firm would be a logical foundation for that effort.”

Behind the Paywall

Below, I break down about a dozen current postings from firms including Point72, D. E. Shaw, PIMCO, BlackRock, Ares and Carlyle. For each, I identified the target investment workflow and proprietary data involved. I then looked at how the firm’s knowledge would be encoded and the outputs tested. I included required investment experience and disclosed compensation where available.

How Bridgewater Is Building an Artificial Investor

Matt Robinson — Tue, 04 Aug 2026 15:31:10 GMT

For more than a decade, Bridgewater’s co-CIO Greg Jensen has been thinking about when machines might make for better investors than humans.

“When will machines write the rules, not just support the rules?” Jensen recalled asking himself in the early 2010s. “When will their intuition actually be better than human intuition?”

At the time, the technology could do “great pattern matching,” Jensen said, but “there was nothing like intuition.” The pieces he believed an artificial investor would need—reasoning, ways to diagnose its conclusions, and methods for working with limited data—had not yet come together.

So Jensen went looking for them.

That search led him to OpenAI, then a little-known nonprofit that would become the face of the AI boom, where he became an early investor around 2016. Later, after a group of researchers he knew left OpenAI to start a new company called Anthropic, he wrote the company’s first outside check.

By late 2022, the year ChatGPT went mainstream, Jensen thought the missing pieces were finally coming together. The following year, Bridgewater created AIA Labs, short for Artificial Investment Associate, a separate research and investment unit tasked with building what the firm calls an “artificial investor.” It started with six people.

Bridgewater kept the group apart from Pure Alpha, its flagship macro strategy. The separation was intended to give the team room to rebuild the investment process around AI—and to keep the new systems from simply reproducing the views Bridgewater’s investors had already developed.

“I don’t want a copy of us,” Jensen said. “We’re so flawed. We’re so bad at so many things. I want something better than us.”

The assignment was broader than finding a new way to analyze data. AIA was supposed to take on more of the work an investor performs: gathering information, developing an explanation for what was happening in an economy or market, testing that explanation and turning it into an investment view.

AI, as it evolves has been, and will be more than ever, an amazing thought partner and, in many ways, thought leader in investment decision making. Decision making solely in one’s head, the way most investors still do it, is obsolete.

Ray Dalio October 2024

AIA was not built around a single model or trading strategy, according to a person familiar with the effort. Instead, it combines Bridgewater’s causal time-series models with specialized and frontier language models, along with systems designed to test and diagnose their outputs. The broader goal is to generate insights across markets, economies and trading strategies.

Bridgewater tested the strategy with its own money first, beginning with roughly $100 million inside Pure Alpha in late 2023. The firm opened it to outside investors in July 2024, attracting almost $2 billion.

In March 2025, Bridgewater CEO Nir Bar Dea said the fund was producing “unique alpha that is uncorrelated to what our humans do.”

Today, AIA is no longer a six-person experiment. The unit has grown to more than 75 scientists, engineers and investors, according to the person familiar with the effort. The strategies now manage about $4.5 billion. AIA Macro has returned 11.3% annualized since inception, with returns that have shown low correlation with markets and other managers, including Bridgewater’s Pure Alpha.

Bridgewater has started using AIA’s systems inside Pure Alpha to accelerate research and systematize investment ideas, the person said.

A Bridgewater spokesperson declined to comment on the hedge fund’s AI efforts.

Read on for how Bridgewater went from writing investment rules by hand to letting an artificial investor manage billions.

The Rise of the House Model

Matt Robinson — Tue, 07 Jul 2026 15:31:46 GMT

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

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.

AI Finds What Markets Miss in News: Study

Matt Robinson — Wed, 24 Jun 2026 15:31:05 GMT

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

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:

“True news is the part that could not have been written before the article appeared: the unexpected guidance cut, the precise earnings surprise, the segment whose growth slowed. The same logic applies to Boeing aircraft-order stories, biotech clinical-trial reports, cybersecurity disclosures, and regulatory announcements. Each article has a predictable layer that follows from the firm’s profile, and a residual layer that contains the genuine surprise.”

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.

Why AI Agents ‘Age’

Matt Robinson — Tue, 16 Jun 2026 15:31:08 GMT

Each morning, AI agents comb through my inbox to read a bunch of news feeds in order to write a daily digest. It’s a good first stop and a step up from Google Alerts.

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.

Agent memory starts degrading after only a few sessions.

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.

Why AI Agents Struggle to Use the Web

Matt Robinson — Tue, 02 Jun 2026 15:32:31 GMT

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.

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.

JPMorgan Seeks Patent for AI-Generated Stock Ratings

Matt Robinson — Tue, 19 May 2026 13:09:24 GMT

JPMorgan is seeking patent protection for an AI system that generates stock-rating predictions, applying AI to one of Wall Street’s most familiar research formats: buy, hold and sell calls.

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.

Revolut Trains AI Model on Its Own Data

Matt Robinson — Wed, 06 May 2026 15:31:14 GMT

Hey, it’s Matt. You’re reading AI Street, where I report on how Wall Street uses AI.

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:

Imagine how a textbook company might use its archives of technical manuals and coursework to train an AI system to do complex scientific processes.

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

How Hedge Funds and Market Makers Are Using AI

Matt Robinson — Tue, 05 May 2026 15:31:17 GMT

Hey, it’s Matt. You’re reading AI Street, where I report on how Wall Street uses AI.

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.