Last quarter, I revealed MarketMasterAI, a machine-learning algorithm designed to beat the stock market.

Its backtests were successful: $10,000 invested in A+ rated stocks in 2010 would have theoretically grown to over $50,000 today, a full third better than the market average. Low-rated F-grade stocks — the short part of a long-short portfolio — would have underperformed by a wider margin. (Even long-only investors need to test both sides of an algorithm).

Since then, MarketMasterAI’s real-world performance has been even better. The 201 A+ stocks initially selected have outperformed the average Russell 3000 by 470 basis points, a 20.1% compound rate of return. Pairing A+ and F stocks in a long-short portfolio would have returned 9.7% in three short months.

These are stunning figures by any reasonable metric. Most mutual funds are lucky to cover their 1%-2% annual fee, never mind generating alpha. And MarketMasterAI’s outperformance was achieved with a large and highly diversified portfolio.

The average daily standard deviation of A+ graded stocks was 31% lower than the average Russell 3000 company in this period.

And here’s the interesting thing. MarketMasterAI’s performance also highlights a separate fact:

Getting rich with AI investing is eerily similar to getting rich without it. 

Artificial Intelligence: A Reflection of Its Creator

MarketMasterAI is a product of my investing habits: long holding periods, a focus on fundamentals, and a general trust of domain experts regarding earnings projections. Its proprietary data inputs are factors that I chose. And its deep learning design comes from my observations of what works in data science. Even though the algorithm “learns” independently, a human still built it.

The result is a system that acts much like I would. A+ companies are largely under-the-radar growth stocks with plenty of stable upside, while F-rated ones are moonshot biotech bets that have historically underperformed. MarketMasterAI essentially “discovered” decades of academic research by itself and proposed investors switch to blue-chips like Mastercard (NYSE:MA) and Autodesk (NASDAQ:ADSK) to ride out a coming storm.

Meanwhile, algorithms designed by high-frequency traders will look far different. By optimizing for shorter holding, allowing more significant risk-taking, and using second-by-second data, these traders will end up with systems that prefer executing hundreds of trades a minute.

Even the same firm can sport multiple systems. Luke Lango’s Prometheus Project was initially designed to beat the market with a 5-day holding period! Meanwhile, our sister company, TradeSmith, used its deep knowledge of technical analysis to create an algorithm that seeks to perform on a longer 1-month window.

Of course, not every algorithm succeeds. Some machine-learning algorithms are better at telling the difference between images of cats and dogs… not picking which stock will rise or fall. And even the most advanced algorithms created by nuclear scientists can become “bamboozled” by noisy markets if you throw too much irrelevant data at it. Every autonomous system will eventually overfit the information and find spurious correlations. The historical relationship between per capita consumption of chicken and U.S. crude oil imports, for instance, might cause an algorithm to base predictions of Exxon (NYSE:XOM) on how much fowl we’re eating.

That means the way to actually get rich with AI investing involves three key steps.

1. Find an Algorithm You Understand

Most high-frequency trading algorithms are surprisingly straightforward. Some make money on the tiny differences from the bid-ask spread by simply placing a limit sell order above the market price and a limit buy order under it. Others use rules like “sell a stock if it crosses its 5-minute moving average.”

AI algorithms work the same way. They take historical financial data, perform appropriate transformations, and create rules to achieve a specific outcome.

There are variations, of course. Some algorithms are programmed to execute short-term trades, while others target longer holding periods. Some optimize for high batting averages, while others only look at raw performance figures, and so on.

Nevertheless, all good algorithms should have an easily understandable end goal and plenty of historical back-tested data. Even “black-box” methods like neural networks can be tested with historical or dummy data to see how they perform. If your algorithm’s creator can’t explain 1) what the product does, 2) how to use it, and 3) what data it’s built on, be sure to ask the algorithm where the nearest exit is (and double-check the answer).

2. Use the System to Gain an Edge… Not to Win the Lottery

Most AI investors aren’t looking for “market-thumping returns,” as a recent Bloomberg article notes. Instead, they’re looking for a slight edge — which on Wall Street can mint billions.

“In finance, you can be very successful by just being a little bit better than 50%,” said Michael Kharitonov, CEO of quant finance firm the Voleon Group.

That’s because AI algorithms are far better at predicting the average, rather than finding rare outliers. Machine learning requires an immense amount of training data, and abnormalities (by their definition) don’t happen often. That means predicting moonshot successes like GameStop (NYSE:GME) and meme-coin Dogecoin (DOGE-USD) is extremely difficult for artificial intelligence.

Consider giving an image recognition algorithm millions of images of handwritten digits as training data. After a while, the algorithm will get pretty good at telling the difference between a “5” and an “8.” But if you give that same algorithm a picture of an elephant, you should expect a nonsensical response.

Some techniques will eventually change this. Researchers are already using large language models to optimize their own prompts, so it’s only a matter of time before algorithms can extrapolate from situations never previously seen. But until then, it’s far safer to use AI to gain an edge, not win the lottery.

3. Stay for the Long Haul

AI investing isn’t immune from making mistakes. Of the 201 A+ rated companies in July, 46 dropped more than the market average. It’s much like a Las Vegas gambler — no poker player goes into a tournament expecting to win every hand.

That means even the best AI algorithms need rules for position sizing, risk management, and the ability to stay invested for the long haul.

• Long-Term View: AI algorithms, especially those designed for investing, often work best with a long-term perspective. There might be periods of underperformance, but an effective algorithm should demonstrate its value over time.
• Avoid Overriding Constantly: Constantly overriding or second-guessing the system can undermine its effectiveness. While human intuition is valuable, it’s essential to recognize when to let the algorithm do its job.
• But, Stay Vigilant: Regularly review the AI’s performance; if you notice consistent anomalies or underperformance, it may be time to reassess or recalibrate the system.

One of the best examples is in 2020-2021, a period where MarketMasterAI’s backtests show F-rated stocks outperforming A+ ones. The system had never seen a junk rally before, so it assumed firms like near-bankrupt clothing retailer Express (NYSE:EXPR) should disappear. Covid-19 relief checks, however, meant that Express would rise 500% in the first half of 2021.

The system would eventually learn to navigate junk rallies and incorporate longer-shot bets into its A+ list. Top-rated stocks have since outperformed. But staying invested during these periods is an absolute must if you want to enjoy the good times that come after the bad.

The Future of AI Investing

When OpenAI launched ChatGPT last year, high schoolers and traders were both quick to experiment. If the seemingly human chatbot could make coherent sentences, why not have it do the heavy-lifting homework too?

The results were… well… not great.

The system hallucinated — stringing together words just because it “looked” right. And the performance of ChatGPT’s stock picks has been mixed. It turns out that large language models are not optimized for large amounts of relevant number-crunching.

But AI is so powerful because its techniques can be reused. The cat-dog recognition algorithm mentioned earlier (known as convolutional neural networks) has since been adapted for predicting stock momentum. Stock charts, it turns out, have a lot in common with 2-dimensional images of household pets.

Systems also improve as more historical data gets added. As a general rule of thumb, doubling the amount of data will double the quality of the output.

That means AI is constantly getting better at what it does… whether in identifying handwritten numbers, chatting with users, or picking stocks. And if you’re willing to get rich slowly, AI investing can help guide the way.

As of this writing, Tom Yeung did not hold (either directly or indirectly) any positions in the securities mentioned in this article. The opinions expressed in this article are those of the writer, subject to the InvestorPlace.com Publishing Guidelines.