The model that does not read the news

When most people think of AI, they think of ChatGPT. They think of a model trained on language, on words, on text scraped from the internet. Keith Kaplan’s TradeSmith presentation makes a clear distinction: Predictive Alpha is not that kind of AI.

“Instead of being trained on words and images, which is how ChatGPT is built,” Kaplan says, “ours is trained specifically on numbers.” The model is called a Time Series AI, and it is a fundamentally different approach to machine learning than the language models that dominate the headlines.

As we explain in our full review of the PRA presentation, this distinction is not just a technical detail. It is the foundation of the entire thesis. And it has roots in a genuinely impressive lineage of AI research.

The Jim Simons precedent

Kaplan grounds the Time Series approach in the history of Jim Simons and Renaissance Technologies, the greatest hedge fund of all time. A $1,000 investment in Renaissance’s Medallion Fund in 1988 would have been worth $90 million 34 years later.

Simons was not an economist. He was a codebreaker. He hired physicists, computer scientists, and mathematicians, not Wall Street analysts. And he was an early adopter of AI, or the primitive versions that existed in the 1980s.

The lesson Kaplan draws is specific. AI and sophisticated models can see things in market data that normal people cannot. The hundreds of billions Simons made prove it is possible to use AI to make money in the stock market. The Time Series approach is the modern expression of the same idea.

How Time Series AI works

Time Series AI is built for sequential data. Stock prices are sequential. Each price update is a data point in a time sequence, and the model learns to predict the next point in the sequence based on the patterns that came before.

This is different from language models, which learn to predict the next word in a sentence. Language models are optimized for text. Time Series models are optimized for numbers, and that makes them better suited for financial data, which is purely numerical.

The London Stock Exchange Group shares 7.3 million price updates per second. Over a year, that is 220 trillion data points. No human can process that volume. No team of human analysts can either. The advantage of Time Series AI is not that it is smarter than a human. It is that it can review amounts of data that are physically impossible for a person to see.

The Google connection

Kaplan notes that the Time Series model his team uses was “pioneered inside Google.” This is a meaningful claim, and it connects Predictive Alpha to one of the most serious AI research programs in the world.

Google has published extensively on Time Series models for forecasting, and the company uses them in production for applications that range from traffic prediction to resource allocation. The same family of models that helps Google predict what will happen next in a data sequence is the foundation of Predictive Alpha’s stock prediction system.

The adaptation to finance is the work of Kaplan’s team, led by Mike Carr, a former Air Force missile coder who worked on cryptography for the National Security Agency and installed the first version of the Internet for the Pentagon. Carr earned the CMT certification, which Kaplan describes as “the trader equivalent of a PhD.” The combination of military cryptography experience and technical market expertise is unusual, and it is the kind of background that shows up in the model’s architecture.

The Pentagon parallel

The presentation makes a striking parallel to explain why Time Series AI matters for financial markets. The Pentagon is using similar models to analyze battlefield data. At UC San Diego hospitals, predictive AI that monitors heart rate and blood oxygen levels saves at least 50 lives per year by alerting doctors before a crisis shows up.

The common thread is that these systems find patterns in sequential data that humans cannot see. In medicine, the pattern is a heart rate trend that precedes a cardiac event. In markets, the pattern is a price trend that precedes a significant move.

The model does not need to know why the pattern exists. It just needs to recognize that it does, and to alert the user in time to act. In the stock market, that means flagging a stock as a buy before it moves, not explaining why it will move.

The evidence from the backtest

The presentation walks through several specific examples. Devon Energy, flagged on February 1, 2022, with a 6% projection. The stock went up 19% that month, an annualized 252% gain. Block, flagged in August 2020 with a 5.61% projection, went up 26%. Trade Desk, flagged in November 2020 with a 13% projection, went up 58%, a 732% annualized return.

The most recent examples are the most compelling because they are closest to the present. Seagate Technologies, flagged in September 2025, went up 42% that month. Lumentum Holdings, flagged on November 1, 2025, went up 50% in the next month.

Each of these is a specific stock, a specific date, and a specific projection. The AI did not read the news. It did not analyze the company’s earnings. It looked at the data, found a pattern, and produced a number.

Whether that pattern holds in live trading, and whether the 71% win rate survives outside the backtest, are the open questions. But the Time Series approach is a genuine AI methodology, not a marketing construct, and it is the same family of models that Google, the Pentagon, and leading hospitals use for sequential data prediction.

NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.