MODELS

AI Trained for Financial Markets

ChatGPT was trained on language, so it struggles with forecasting. Train the same transformer architecture on time-series data, and its predictive power improves.

As this Risk article highlights, large language models are bad at financial forecasting—not because the architecture is flawed, but because they weren’t trained on the right kind of data.

A very short recap of the evolution in AI:

The current AI boom began with a 2017 Google paper on how to better translate text. Before that, machine learning generally processed language word by word, and struggled to capture contextual relationships between words. For example, if I say to you, Paris → France → Berlin → ______ you know it’s → Germany — something earlier machine learning struggled to do.

Transformers process whole sentences at once, solving the problem that earlier models had with understanding these relationships. They can also be trained on different datasets.

Researchers are now pursuing transformers trained specifically on time-series data as the next major development.

“The real frontier is in models actually built for time-series rather than stretching LLMs,” Agam Shah, a PhD candidate in Machine Learning at Georgia Tech, told me. “The best part about transformer architecture is that it is one of the most scalable and trainable architectures.”

Some of last year’s NeurIPS and ACL papers showed how ideas from NLP are spilling into financial forecasting, Shah points out. Companies like Stripe are already training transformer models on payments data. Other companies are using it on grocery data.

Jimin Huang, founder of The Fin AI, which develops open-source AI models for financial applications, says his group is experimenting with what they call “decision models” trained to predict actions in real markets rather than the next word. Their ACL-accepted research shows a lightweight LLM trained on trading tasks outperforming GPT-4 with top financial multi-agent systems.

Trading is uniquely complex, Huang noted, because it involves interacting with millions of other participants.

Takeaway: AI models based on transformer architecture are moving beyond language — and early evidence shows they can outperform GPT-4 on financial tasks.

NYC

AI Street Meetup

Next week, I’m in New York to attend Cornell’s Future of Finance & AI conference. And with Jordan Hauer of Amass Insights, we’re co-hosting a meetup with folks interested in the data + LLMs + investing space on Sept 18 ~6:00.

We have about 50+ signed up between the two groups. So come by! We can chat AI, transformer models, or Seinfeld.

Just fill out this form for location details. See you then!

DATA EXTRACTION

When Smaller Beats Bigger in AI

Large language models are often treated as one-size-fits-all solutions for white collar work. The same model is often judged on a large swath of benchmarks like coding, math, image generation, etc. New releases are quickly hailed as a “Game Changer! 🚀 ” --A phrase I hear so often it’s starting to make my eye twitch.

But this fixation on “large” language models may overlook other AI options. For certain tasks, a collection of smaller, specialized AI models may outperform a general-purpose LLM.

That’s the pitch from Thomas Li, CEO and co-founder of Daloopa, who says he’s generally avoided the latest frontier models for data extraction.

Daloopa has built what Li calls a basket of narrow AI models, each tuned for a single job rather than trying to handle everything at once.

“Our models have an IQ of 250 on one task and 2 on something else,” Li said in a Zoom call. “It can’t write you a song or give you a travel itinerary for Paris.”

Li says its strategy has paid off. The company, founded in 2019, claims an over 99% accuracy rate in extracting fundamental data and counts a large majority of the biggest hedge funds as clients. Daloopa says it offers 10 times more historical data than legacy providers such as S&P Capital IQ.

Li said Daloopa can update investors’ Excel models within minutes of the company’s earnings release and highlight any missed numbers they should review.

“If Microsoft posts earnings at 4:00, we can populate investors’ Excel models from 4:02,” Li said.

Small AI can handle jobs that big AI is too big for.

Takeaway: A group of small AI models may outperform general-purpose LLMs in data extraction.

NEWS ROUNDUP

Startup Behind Goldman’s AI Engineer Valued at $10 Billion

San Francisco-based Cognition AI, whose A.I. engineer Devin is used by Goldman Sachs, doubled its valuation after a $400 million raise led by Peter Thiel’s Founders Fund. (Observer)

SimCorp Teams With Axyon AI to Add Predictive Analytics

SimCorp, part of Deutsche Börse Group, will integrate Axyon AI’s predictive analytics later this year, giving portfolio managers AI-driven forecasts, rankings and trading signals. (Hedgeweek)

Italy’s Domyn Hires Ex-BlackRock Exec For Financial AI Division

Domyn, a $1 billion Italian AI startup building tools for banks and asset managers, named former BlackRock managing director Stefano Pasquali to head its newly launched financial services division. (Reuters)

AI Not Affecting Job Market Much So Far, New York Fed Says

"Businesses reported a notable increase in AI use over the past year, yet very few firms reported AI-induced layoffs," New York Fed economists wrote in the blog. "Indeed, for those already employed, our results indicate AI is more likely to result in retraining than job loss, similar to our findings from last year (Reuters)

AI Adoption Surges with 91% of UK Financial Firms Embracing the Technology

Lloyds’ latest Financial Institutions Sentiment Survey shows UK banks, asset managers, and insurers are moving AI from pilot projects to execution, with 59% reporting productivity gains, 33% seeing deeper customer insights, and over half planning increased AI investment in the year ahead.

SPONSORSHIPS

Reach Wall Street’s AI Decision-Makers

Advertise on AI Street to reach a highly engaged audience of decision-makers at firms including JPMorgan, Citadel, BlackRock, Skadden, McKinsey, and more. Sponsorships are reserved for companies in AI, markets, and finance. Contact me ([email protected]) for more details.

ICYMI INTERVIEW

Spotting Accounting Shenanigans with AI

I spent six years writing about white-collar crime for Bloomberg News. In that time, I learned that accounting fraud cases were among the longest for the SEC to investigate and the hardest to bring.

I was surprised. I naively thought, “Well, if the company is cooking the books, eventually folks will find out, right?” But that’s not always the case. Bad actors can use events outside their control—like COVID—to bury years of weak numbers.

It’s just hard to police accounting statements. And even if you latch on to what you think is a significant issue, sometimes it doesn’t matter because the company is massive. I once wrote about a company that stuffed six months of revenue into a quarter and the market basically shrugged. Granted, this detail does not inspire confidence.

To dig deeper into these challenges, I spoke with Hamish Macalister, co-founder and CEO of Transparently.ai, which uses traditional AI and large language models to assess signs of accounting manipulation.

Transparently.ai rates the accounting health of 80,000+ public companies on an A-to-F scale, flagging early signs of manipulation and potential failure. Founded in 2021, the Singapore-based company counts two of the Big Four auditors as clients and money managers overseeing $4 trillion in assets.

Macalister worked as a macro strategist at Citigroup, led quantitative strategy in Asia at Deutsche Bank, and later served as chief data scientist at Firth Investment Management. He also earned a PhD in finance, where his doctoral research on analyst forecasts laid the groundwork for Transparently.ai’s approach.

In this interview, you’ll learn:

• Why accounting manipulation is more common than most investors think.

• How avoiding high-risk companies based on these scores can generate meaningful alpha.

• Why auditors and analysts miss red flags—and how AI can surface them. 

This interview has been edited for clarity and length. 

How does Transparently.ai help investors evaluate accounting nuances across industries?

This is a perfect problem for machine learning because it’s very complex and multidimensional, but also one for which there’s a great deal of data. That combination makes it well suited to machine learning.

There may be relationships a person would struggle to identify, but a machine can. Another advantage is that the machine isn’t wedded to traditional ways of thinking. For example, it might pick up on signals that an activist short seller would look at, but from a different angle.

One of the red flags might be unusually high margins—possibly a sign the company is faking revenue or hiding costs. That’s a classic example of what an activist short seller might look for.

From a machine learning or AI standpoint, the system might learn something similar: unusually high margins can be a warning sign. Our system does flag that from time to time. But it can also flag unusually low margins if it detects that certain combinations of features—low margins alongside other factors—may indicate a company is doing something unusual.

Machine learning can identify very complicated patterns that may not be intuitively obvious. The one thing I’ll add to that is it cannot just be a black box—unless you’re a quant and all you care about is the black box. In that case, all you want is the numerical output: the risk, the number, the indicator.

But for most of the users we deal with, they want some sort of explanation behind this. So it’s critical to design the system not only to provide an indication that something unusual may be happening in a company’s accounts, but also to explain why and how. It should guide what you need to do next: what questions to ask management, what areas to investigate, and what procedures to implement if you’re an auditor, given the specific features of that company.

Read the Full Interview Here

WHAT ELSE I’M READING

• The Rise of Hudson River Trading (Rupak Ghose newsletter)

• OpenAI Backs AI-Made Animated Feature Film (WSJ)

• Meet the Guys Betting Big on AI Gambling Agents (Wired)

• With AI Leadership, Thematic ETFs Will Reach Record Flows in 2025 (Funds Society)

• Citi poaches IBM exec to accelerate its AI ambitions (Business Insider)

CALENDAR

Upcoming AI + Finance Conferences

• Cornell Financial Engineering Manhattan 2025 Future of Finance & AI Conference – Sept 19, 2025 • New York (I’ll be attending)

  A one-day forum on AI, quantitative finance, and hedge-fund strategies, attracting leading quants and industry practitioners.

• Bloomberg-Columbia ML in Finance Conf – Sept 25, 2025 • New York

  Academic–industry event hosted by Columbia University and Bloomberg, focused on ML applications in finance including asset pricing, market forecasting, and LLM risk.

• GAIIM Conference 2025 – Sept 30, 2025 • New York

  Forum on practical applications of AI in investing, featuring tools for research, valuation, and portfolio workflows.

• AIFin Workshop at ECAI 2025 – October 26, 2025 • Bologna, Italy

  One-day academic workshop on AI/ML in finance, covering trading, risk, fraud, NLP, and regulation.

• AI in Finance 2025 – October 27–30, 2025 • Montréal

  Academic event covering ML in empirical asset pricing and risk.

• ACM ICAIF 2025 – November 15–18, 2025 • Singapore

  Top-tier academic/industry conference on AI in finance and trading.

• AI for Finance – November 24–26, 2025 • Paris

  Artefact’s AI for Finance summit, focused on generative AI, future of finance, digital sovereignty, and regulation 

• NeurIPS Workshop: Generative AI in Finance – Dec. 6/7 • San Diego One-day academic workshop at NeurIPS focused on generative AI applications in finance, organized by ML researchers.