Listen to the audio version of this article (generated by artificial intelligence).
Editor’s note: Yesterday, we looked at how AI is starting to uncover patterns in market data that most investors can’t see.
Today, Keith Kaplan takes this idea one step further.
He points to a new initiative called Project Glasswing, where advanced AI systems are being used to find hidden vulnerabilities in complex systems, problems that have gone unnoticed for years.
This is important because it shows what AI can really do. It can work through huge amounts of data and uncover patterns that humans miss. The same approach can be applied to the stock market.
In a five-year backtest, Keith’s team applied this approach across thousands of stocks and built a model portfolio that found 30,000 hidden signals. I delivered a Return 12Xtogether Winning rate 73%In 2022, it was acquired 16.6% While the S&P 500 index fell almost 20%.
It is difficult to ignore such results.
Below, Keith explains how this approach works and how it can help investors spot opportunities early. He also walks through it in more detail throughout his work Artificial intelligence signals trading event.
If you haven’t watched it yet, you can watch the replay here.
Now, here’s Keith…
*******************
The most important story in the world right now is not the Strait of Hormuz.
…or the price of oil…
…or the Epstein files.
It’s 100 times more important than all of that.
It’s an intensive program called Project Glasswing that brings together top financial officials from the federal government and CEOs of some of America’s most powerful companies.
What sparked it was a new frontier model called Mythos from Anthropic, the private company behind Claude AI. By all accounts, Mythos is the most capable AI model ever created – and what it can do is truly alarming.
It can read the source code of the software that runs your bank, your hospital, or even your power grid… and find security flaws that human experts have missed for decades.
As just one example, Mythos discovered a flaw in OpenBSD—the system that runs sensitive firewalls, government networks, and critical infrastructure—that human developers had missed in more than 27 years of detailed security audits.
Then he wrote the code to exploit it… independently… on the first try.
As Anthropic said in an April 7 press release, “Mythos reveals a stark truth about the state of AI in 2026…
AI models have reached a level of coding ability where they can outperform all but the most adept humans at finding and exploiting software vulnerabilities.
When Washington leaders learned of this, they knew they had to act quickly.
Treasury Secretary Scott Bescent and Federal Reserve Chairman Jerome Powell summoned the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo and Goldman Sachs to the Treasury Building at 1500 Pennsylvania Avenue — just steps from the East Wing of the White House.
Anthropic announced the Glasswing project on the same day. Eleven launch partners – Amazon, Apple, Google, Microsoft, JPMorgan Chase, Cisco, NVIDIA, Broadcom, CrowdStrike, Palo Alto Networks, and the Linux Foundation – will get early access to Mythos.
Anthropic has allocated up to $100 million in usage credits. Mission: Find the flaws before someone else uses a model like this as a weapon.
This may sound like the plot of a Hollywood movie. But Glasswing is very real. And what it tells us about the capabilities of artificial intelligence has implications that reach far beyond cybersecurity — to every wallet in America.
Deciphering the hidden structure of the market
Mythos finds patterns in computer programs that no human can see.
It reads millions of lines of code and identifies certain sets of conditions that indicate a vulnerability. Then he works on them.
Look deep enough, and you’ll find that the stock market contains similar types of hidden structures. There are “signals” buried in decades of data for every stock — specific sets of circumstances that precede consistently big moves.
Just for most of the market’s history, it was invisible. The data was there… but no one had the tools to read it.
But today, thanks to artificial intelligence, it is possible to find signals with historical accuracy rates of 90% or better.
I know this because my team at TradeSmith and I built a new AI-powered trading tool that detected 30,000 of these signals across nearly 2,500 stocks.
As I showed the nearly 9,000 people who joined me Artificial intelligence signals trading event On Wednesday, in a five-year backtest, a typical portfolio of these signal trades generated a 12x return.
And in 2022 — the worst year for stocks in half a century — they posted an average backtested gain of 16.6% while the S&P 500 He falls Nearly 20%.
If you missed it, the replay is still available online. It’s packed with trading examples, strategy details, and on-screen demonstrations. Go here to watch it now.
Today, I want to share something I didn’t have time to cover on Wednesday — the amount of work and creativity that went into building our new system.
The answer is: more than I expected.
Can the weather in Paris move the stock market?
Our lead developer, Mike Carr, has been writing code for 40 years.
He spent 20 years in the US Air Force — coding nuclear missile trajectories, working on cryptography for the National Security Agency, and helping install an early version of the Internet at the Pentagon.
When he left the military, he continued to manage more than $200 million in client funds. He also became a certified market technician – a certification that only 4,500 people in the world have.
Two years ago, he joined TradeSmith to help us develop new analytics and strategies. He brought with him the kernel of an idea he had been working on for more than 20 years.
In 2003, Mike began conducting preliminary studies on signals. Every time he noticed a recurring pattern in the data that tended to precede the stock’s movement, he recorded it. He has traded this way for years, constantly testing what worked and what didn’t.
Then in 2016, he read a Bloomberg profile of Jim Simons’ famous hedge fund, Renaissance Technologies. One detail stuck with him: Simons had once found a tradable weather signal in Paris.
Mike realized that if he could find a signal in the Paris weather, the signals hiding in ordinary market data must be almost limitless. That was the moment he decided to stop looking for signals manually and start building a system that could catch them at scale – a system that ordinary investors could actually use.
Last year, we started feeding the roughly 150 signals it had collected into the AI system and pushed us to generate more signals like it. We’ve processed over a trillion database rows, and run each stock through 847 individual calculations. We tested every combination of price patterns, technical indicators, and calendar conditions we could find.
Then we ran into a problem we didn’t expect.
On any given day, the AI-powered signal generator was delivering a stream of 30,000 trade setups – all with historical accuracy rates of 75% or better. It was too much for any trader to handle, regardless of their level of experience.
So we spent the next six months solving a different problem: How do you take so many high-quality signals and deliver something that an investor can actually use?
The answer Mike came up with was Quality Score. It is a rating from 0 to 100 that takes into account each signal’s win rate and average returns and uses machine learning to see how effective it has been during similar market conditions in the past.
Pair that with a focused model portfolio, and Flood becomes an actionable shortlist.
Introducing our three-stock strategy
At any given moment, you hold three stocks in the S&P 500 — each stock selected by an algorithm based on its Quality Score and other key factors. When an exit signal sounds on one of the three, a new trade recommendation takes its place.
This is the whole strategy. Three jobs, always live. And each one was chosen not because a human liked a chart or a story… but because the algorithm chose the mathematically optimal trade.
We tested it again from January 1, 2020 through January 30, 2026 – a stretch that covered the Covid crash, the 2022 bear market, two years of historic inflation, rising interest rates, and two wars.
It was not a friendly period to stress test the trading system against him. But here’s what the main signals wallet produced:
- A 54% compound annual return – Compared to about 15% for the Standard & Poor’s 500 index over the same six years
- A Winning percentage: 73.4% Across hundreds of deals
- A Maximum withdrawal 18.1% — Less than the S&P 500 maximum drawdown of 25.4%
The maximum withdrawal limit from the model portfolio is worth stopping at.
Many trading systems can generate a high compound return in backtesting. Few can create an index that holds up better than a benchmark like the S&P 500 during its worst periods. This is the true test of whether the system is managing risk effectively or is just riding luck.
And that’s the goal of what we’ve spent the last 12 years (and in Mike’s case, over 20 years) developing. Hedge funds have been doing this kind of work for decades — pattern recognition, machine learning, and disciplined rotation in and out of short-term trades. But until now, there has been nothing like this for ordinary investors.
I went into all the details during the launch event on Wednesday. So, if you haven’t already, make sure you do check it out While still online.
Keith Kaplan
CEO, Tradesmith
