The technology behind ChatGPT is inherently versatile. It can be trained on large, complex datasets beyond language.

Researchers have begun applying transformer-based architectures to any domain where data follows consistent structure, rules, and sequencing, such as weather forecasting, autonomous driving, and payments.

Weather

DeepMind predicted Hurricane Lee’s landfall 9 days in advance, outperforming the industry-standard "gold-standard" physics models by three days.

Driving

Waymo shifted from traditional AI to a unified transformer architecture, resulting in an 8% jump in motion prediction accuracy.

Payments

Stripe created a "Payments Foundation Model" that treats transactions like sentences. The model boosted fraud detection for card-testing attacks to 97% from 59%.

I’ve written about a few instances where researchers have applied transformer models to financial data and they’ve had encouraging early results. Financial markets have similar structure. Returns, transactions, and order book messages follow repeatable patterns that can be learned in the same way language models learn grammar.

In November, I highlighted work from Eghbal Rahimikia at the University of Manchester on how he trained a transformer model on daily stock returns that showed substantial forecasting and economic improvements over traditional methods.

This week, I spoke with Juho Kanniainen, a professor at Tampere University’s Data Science Research Centre, on how he and his colleagues trained a transformer-based model on limit order book messages with superior predictive results. Check out the interview below.

I expect more research and use cases exploring frontier models and financial markets. The main bottlenecks in building these models are access to sufficiently large financial datasets (you need billions of datapoints), GPUs, and specialized expertise. I think this has created an arms race for the world’s biggest funds, a topic I plan to cover in more detail this year. University of Manchester’s Rahimikia summed it up well:

AI is moving faster than companies can scale.

I often end interviews with this open-ended question: what’s been most surprising to you about AI?

I asked ~100ish experts in 2025 this question and the most common answer was how quickly the technology was advancing.

And these are not bombastic folks. They weren’t selling me anything. They were/are genuinely surprised by what AI can do.

This has created a gap between what AI can do today and how Wall Street is currently deploying it.

This is a human problem. Anyone who’s ever worked at a big organization knows that it’s hard to get everyone on the same page. It’s even harder when there are no agreed-upon standards. Governance and standards are just emerging.

Last year, I predicted that more standards would emerge in 2025. Some did, but not many, which makes me realize that this is not a problem that will be solved in a year.

Virtually every week there’s a news story about how AI is going to lead to mass layoffs, but there’s no real data showing that this is happening. (If you’ve seen any credible analysis or data showing AI-induced job losses, please drop me a line, [email protected] because at this point I’m wondering what numbers these AI doomers are looking at.)

My favorite counterpoint to this narrative is this article from MIT written in 1938(!) by the university’s President at the time on the “Bogey of Technological Unemployment.” He was pushing back against John Maynard Keynes’ thesis that technological progress was “outrunning the pace at which we could find new uses for labour.”

So, we’ve been here before. And the pattern repeats across decades: each major wave of technology arrives with confident predictions of mass displacement that rarely play out as expected. The historical record is full of these false alarms. An often-cited example is ATMs. In the 1970s, everyone predicted that bank clerks would be made obsolete. (The opposite happened.) In the 1980s, teachers complained about giving students calculators! The trope that technology makes you dumb and takes your job has been around for a while.

Also, AI can create jobs. OffDeal, a startup that uses AI to speed up small-business M&A, says the technology makes small sell-side deals economical, bringing transactions into the market that previously did not happen at all, according to the FT.

I’m not saying AI won’t change how we work. It will. I’m saying it’s not a foregone conclusion that it leads to mass unemployment.

LLMs are often treated as one-size-fits-all tools, but smaller language models can be a better fit for specific enterprise tasks.

“Small” is relative, and there is no agreed-upon definition. What matters is scope. Models trained on narrowly defined data can be more accurate, easier to control, and more cost-effective than their larger counterparts.

These models excel at tasks that are repetitive and precision-heavy rather than reasoning-intensive: extracting numbers from filings, normalizing entity names, classifying transactions, or flagging anomalies. None of this requires a model trained on the entire internet.

Large enterprises are already leaning into this approach. JPMorgan uses smaller models to classify and normalize messy payments data. Meta relies on them to deliver and optimize ads at scale. Airbnb uses them to automatically resolve routine customer service issues. IBM has shown that tightly scoped models can be constrained to produce consistent, non-drifting answers.

One of the strangest things about the AI boom to me is how much money is flowing into compute, literally hundreds of billions of dollars globally, without anything resembling a clear market price. 

Buyers of computing power still have almost nothing to benchmark against. There is no Expedia for compute the way there is for flights, and no universally accepted spot price like you would see for oil, natural gas, or electricity. If a company wants GPU capacity today, the process is mostly manual. You call around. You talk to cloud providers, resellers, or data center operators. You compare quotes that are hard to normalize because the terms, hardware, utilization guarantees, and performance metrics all differ. 

As a result, two buyers can be paying very different prices for what appears to be the same underlying resource.

That starts to matter more as compute shifts from being an occasional capital expense to a recurring operating input. For AI labs, hedge funds, and large enterprises running models in production, compute may no longer be a one-off cost. It is closer to rent, fuel, or freight. And markets generally do not tolerate opacity for long once volumes reach this scale.

In 2025, I interviewed Carmen Li,* CEO of Silicon Data and Simeon Bochev, CEO of Compute Exchange to impose structure on what is still a largely opaque market.

*Li is now CEO of both Silicon Data and Compute Exchange.

For most of the past few decades, Wall Street investors largely fell into two camps. Fundamental investors dug through 10-Ks, earnings calls, and meetings with management. Systematic investors built models and let algorithms trade on their behalf. Until recently, the two groups mostly operated in parallel.

Historically, quantitative strategies were constrained by data availability. Most models relied on structured inputs such as prices, volumes, and accounting ratios, because much of the information that fundamental investors cared about was either not digitized or difficult to process at scale. Long-form text does not fit neatly into traditional statistical models.

Large language models have begun to change that constraint. They make it possible to analyze text directly, allowing earnings calls, corporate disclosures, and other unstructured data to be converted into signals that can be queried, compared, and tracked over time.

At the same time, discretionary investors now have access to tools that do not require deep engineering skills. Methods that once sat exclusively inside quantitative teams are increasingly available as standard software. As a result, the two approaches are converging. By 2026, the distinction between quant and fundamental investing is likely to matter less than it once did.

LLMs are prediction machines. They excel by forecasting the next token, typically, thought of as the next word.

But you can train the same transformer architecture on the weather, genes and financial transactions if you have a large enough dataset.

And there’s growing research that just as transformer-based models excel at language, they can excel in other domains.

You just need a lot of computing power and terabytes of data. And financial markets are brimming with these data streams.

I recently spoke with Juho Kanniainen, a professor at Tampere University’s Data Science Research Centre, on how he and his colleagues trained a transformer-based model on limit order book messages with superior predictive results.

We talked about his new pre-print, “LOBERT: Generative AI Foundation Model for Limit Order Book Messages,” co-authored with Eljas Linna, Kestutis Baltakys, and Alexandros Iosifidis and presented at QuantMinds and the NeurIPS Workshop on Generative AI in Finance. Here’s what they did:

What follows is a conversation about the potential future of “foundation models for markets.” 

This interview has been edited for clarity and length.

Matt Robinson: Could you explain the core idea behind using these models for market data and how you address the question of market efficiency?

Juho Kanniainen: The models themselves don't really care if the sequences are about words or limit order book messages, as long as there are patterns. When it comes to limit order book messages and the whole stream of data, patterns emerge because trading and market making algorithms react to past events, continuously creating chains of reactions. This is why (illegal) spoofing exists in markets, where manipulators induce desired reactions using orders that are never intended to be executed. The key point is that certain patterns emerge in the limit order book, whether unintentionally or intentionally created, enabling short term prediction with ML models that are capable of capturing them.

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. Email me ([email protected]) for more details.

Vanguard CIO Nitin Tandon says AI is the only realistic way to bridge the gap between 50 million clients and 20,000 employees, boosting efficiency and enabling personalized advice at scale, with a long-term goal of AI acting as a trusted digital advisor. (Yahoo)

Despite heavy investment in AI and high automation potential, experts say Wall Street job cuts are mostly tied to overhiring and economic uncertainty, with AI acting more as a productivity boost and convenient scapegoat than a true replacement for finance workers for now. (Fortune)

Minotaur Capital is preparing a potential ETF to bring its AI-driven global equities strategy to the broader public, following strong performance and growing interest in its proprietary research system, Taurient. (Minotaur)

Minutes from the Bank of England’s Artificial Intelligence Consortium meeting in October 2025 details concentration risk, AI explainability, AI-driven contagion, agentic models, and emerging standards in UK financial services. (BofE)

I imagine I’m missing some events, if there’s one you think I should add, reply to this email or reach out: [email protected]