The current AI boom stems from the breakthrough 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 billions of datapoints.

Netflix trained a model on viewer data. Stripe trained a model on payment data.

Results look promising: Netflix says the model learns more from a viewer’s history to provide better user recommendations, while Stripe says the model is better at spotting fraud.

The biggest bottlenecks to training these models beyond language are having enough data and computing power.

A new study cleared these hurdles by training a transformer model on up to 2 billion datapoints on daily stock returns over 34 years and 94 countries for about 50,000 GPU hours.

Researchers at Manchester, UCL, and Shanghai University say this is the first comprehensive study of Time Series Foundational Models in global markets. Time series data is any set of observations recorded in order over time (think temperature, daily sales figures, etc).

Here’s what they did:

The researchers took two popular "foundation models" for time series forecasting (Chronos from Amazon, TimesFM from Google) and asked them to predict next-day stock returns.

They trained these models from scratch by using only financial data. They then compared the model they built against off-the-shelf versions trained on generic time series data. They compared all of this against simpler, well-established methods that quants already use (specifically ensemble models like gradient-boosted trees, which are basically very sophisticated decision trees).

Here’s what they found:

The off-the-shelf models flopped. When you just download these foundation models and point them at stock data, they perform terribly — worse than much simpler techniques. Fine-tuning helped a bit, but not enough to close the gap.

Training from scratch worked surprisingly well. When the researchers trained these same architectures using only financial data, performance jumped dramatically. The models started generating meaningful trading signals that translated into actual portfolio returns.

A Chronos model pre-trained on financial time series achieved a 36.84% annualized return and a Sharpe ratio of 5.42, compared to losses when used out of the box. The traditional quant benchmark still edged it out at 47.25% in this test, but the results suggest transformers can become more competitive with more data.

What’s interesting is that synthetic data improved accuracy, eventually allowing specific large models (like TimesFM 20M) to approach or match traditional benchmarks.

Expanding the training set to global markets and adding generated data pushed TSFM performance even higher. The researchers released their models publicly through FinText.ai and Hugging Face, which should help with follow-on work.

I asked lead author, Eghbal Rahimikia, for his big picture takeaway:

Transformers, the same technology behind ChatGPT, do seem well-suited to financial forecasting. You just have to train them on lots of financial data from the start.

AI has made me realize how much white collar work is just manual labor at a keyboard. 

Point. Click. Copy. Paste. Drag.

Many firms are targeting this pain point. This week, Model ML raised $75 million in Series A led by FT Partners alongside YC, QED, 13Books, Latitude, and LocalGlobe.

The company builds AI-driven workflows that create Word, PowerPoint, and Excel materials from a firm’s data while matching the clients’ formatting style.

I asked the company’s co-founder, Chaz Englander, who started the company with his brother, Arnie, what makes them different?

Model ML isn't alone in the race to automate the junior associate. Lots of capital has flowed into this vertical over the last ~18 months: Hebbia raised $130 million last summer; 73 Strings secured $55 million in February; Rogo closed on $50 million in funding in April; and Finster raised $15 million last month, among others.

There’s lots of investor demand for AI companies that can really automate the rote processes of bankers.

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.

You know AI in finance is mainstream when NPR is doing features about it. It’s a good roundup of JPM is using AI to automate tasks like pitch books, cut back-office headcount, and shift junior banker work.

The order announced the “Genesis Mission” for AI, a new federal push to mobilize the federal government's research and data to create artificial intelligence models. (AP)

An SEC advisory body is eyeing how the agency could compel companies to disclose details on their use of artificial intelligence. (BBG Law)

Recently added events in BOLD