AI agents promised to revolutionize day trading. Then the losses started piling up
Artificial intelligence showed up promising to revolutionize financial markets for anyone with access to a smartphone. And for a while, it actually looked like that promise might hold up. Social media feeds were overflowing with screenshots of insane profits, videos teaching people how to set up trading bots in minutes, and the irresistible narrative that easy money was just one well-crafted prompt away.
But reality, as it almost always does in financial markets, has a peculiar way of settling the tab.
Jake Nesler, a 29-year-old software engineer living in Scranton, Pennsylvania, bought into the idea enough to spend two and a half weeks teaching Anthropic’s Claude model to think like him. The goal was to build an agent that could analyze risk, identify entry signals, and manage position sizing in the stock market — all while Nesler handled his day job as a programmer.
The inspiration came from an unlikely place. Nesler had seen an experiment by Anthropic itself where Claude controlled an office vending machine. The idea that popped into his head was straightforward: what if, instead of serving up snacks, Claude could be trained to buy and sell stocks?
The result was… mixed at best.
In the first week, the bot made one solid call and avoided an estimated $10,000 loss. When Nvidia’s earnings sent the stock soaring in late November, the agent debated with itself whether it should chase the rally. Fortunately, it decided against chasing the inflated price — a move that would have blown a sizable hole in the simulated portfolio. 😅
That same week, though, the bot also lost money on a string of speculative trades that simply didn’t pan out. After five days of live operation, the scorecard showed one good decision sitting next to a sequence of losses.
Nesler’s story isn’t an isolated case. It represents something much bigger happening right now across financial markets worldwide, where a new generation of investors is betting that AI agents can handle the dirty work of automated trading more efficiently than any human ever could. Except the numbers, more often than not, are telling a very different story from the viral promises making the rounds on social media.
The dream of money on autopilot
The idea behind automated trading with artificial intelligence is seductive and makes perfect sense on paper. You set up a smart agent, define your risk criteria, connect it to your brokerage, and let the machine do the work while you sleep, travel, or just enjoy life. This appeal isn’t new, but it’s taken on an entirely different dimension with the rise of large language models and generative AI tools that anyone can access today — no coding skills or deep knowledge of quantitative finance required.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Open-source platforms like OpenClaw already let users chat with their AI agents through messaging apps like WhatsApp and Telegram, drawing in a crowd of aspiring traders without the need for a computer science degree. All someone has to do is hook up an AI model to the system and turn it loose with simple instructions. That ease of access is both the greatest strength and the greatest risk of the current ecosystem.
Over the past two years, platforms dedicated to this kind of automation have popped up all over the internet, many of them aimed directly at retail investing — the segment of individual investors, regular people trading with their own money and, more often than not, without institutional backing. These platforms sell a powerful narrative: that artificial intelligence can level the playing field between the small investor and the big Wall Street funds. It’s a compelling argument. And it works really well as a marketing strategy, especially when the creators of these tools share screenshots of winning trades on social media without showing the full history of losses that came before or after.
Trading platforms are also riding this wave. Companies like Public Holdings are looking to offer their own AI agents directly to customers. In the crypto world, exchanges like Polymarket, OKX, Bybit, and Kraken have rolled out interfaces in recent months that make it easier for bots to execute trades. The incentive is straightforward: bots trade at high frequency, and exchanges live off transaction volume.
What actually happens in practice, though, is that most people who jump on this idea don’t have a clear understanding of what they’re building or buying. Setting up an AI agent to trade financial markets requires a lot more than feeding the model a few prompts and hoping it reads charts correctly. You need to understand how markets actually work, what the real risks of each trade are, how the model interprets historical data, and most importantly, what the limits are of what AI can reliably predict — which, in practice, are much narrower than YouTube videos tend to suggest.
The problem lies in the technology itself
Nesler ran into a recurring issue with his agent that reveals something fundamental about the nature of language models applied to trading. The bot kept insisting on being too responsible, naturally gravitating toward blue chips and S&P 500 stocks — the kind of positions that barely move over the course of a week. Nesler had to intervene repeatedly, pushing the model to seek out riskier trades that matched his own profile.
This behavior isn’t a bug. It’s a direct consequence of how large language models are built. Models like Claude are trained on massive volumes of financial content, risk management literature, and market commentary. Without specific instructions to the contrary, they absorb the consensus view on what responsible investing looks like and start behaving like the average of every financial advisor blog on the internet. Some traders building agents on top of these models end up fighting a real battle against this default conservatism, trying to extract risk appetite from a system that was specifically trained to avoid it.
After tuning the agent to his liking, Nesler ended up with a return of roughly 7% over 30 days, beating the S&P 500’s roughly 4.5% gain over the same period. On the way there, though, the bot severely tested Nesler’s tolerance for volatility, with drawdowns of up to 22%. Even though he published his code online for others to experiment with, Nesler makes a point of saying he doesn’t recommend anyone put real money into this.
It’s totally possible to make money doing this, he said. But anyone could do that with pure luck on options. It doesn’t mean you won’t lose that money too.
When financial losses become just a statistic
The financial losses generated by poorly configured automated trading bots or bots based on strategies without a solid foundation have a particular quality that makes them especially dangerous for individual investors. They happen gradually, through small trades that seem acceptable on their own, until the cumulative balance reveals a significant hole. This pattern is exactly the opposite of what automation platforms show in their promotional materials, which typically highlight one-off gains and sweep the losing streaks under the rug.
On X, claims of extraordinary returns through AI agents have become their own content genre. One viral post, viewed 4.7 million times, boasted a 5,860% return in two days on the prediction market platform Polymarket. The story was later debunked by another account operated by an AI agent, which demonstrated that the claims were impossible. Other similar posts went so far as to direct users straight to malware, posing a real security risk for unsuspecting investors.
There’s also a relevant psychological component in this equation. When a human loses money on a trade, they feel the weight of that decision and, in theory, learn from it. When a bot loses, the investor often doesn’t process that loss the same way, because the sense of distance created by automation reduces the immediate emotional impact. The practical result is that many people keep bots running in the red for weeks or months because they can’t confront the accumulated losses with the same clarity they’d have if they’d hit the sell button themselves.
Jay Malavia, co-founder of Kairos, a Chicago-based company that operates a trading terminal for prediction markets, has heard versions of this story many times. Trading is a zero-sum game, he explained. A competitive edge, by definition, stops existing when it’s shared with the masses. If I had a bot that actually worked, I wouldn’t give it to you. And if you had one that worked, you definitely wouldn’t post it online.
The threat to prediction markets
Beyond individual financial risk, AI trading agents raise a broader concern about the future of prediction markets — platforms like Polymarket and Kalshi where people bet on outcomes of real-world events like elections and sporting events.
The concept behind these markets is that they function as a kind of collective thermometer. The people betting on them supposedly know something — or at least believe in something with enough conviction to put money on the line. This aggregation of knowledge and convictions is what makes these markets, in theory, useful forecasting tools.
But an AI agent placing bets based on what it can find on Google isn’t adding new knowledge to the mix. It’s just recycling information that’s already publicly available. If enough bots push out the humans who actually have insights into how a given election or sporting event might unfold, the contract stops being a forecasting tool and turns into something more like an echo chamber. The result is a machine that merely averages out what the internet already thinks, stripped of the contrarian judgment that’s precisely what makes crowds intelligent.
Annanay Kapila, a former quantitative trader who now runs the derivatives exchange QFEX, doubts that AI trading bots work at scale for retail investors. Contracts on prediction markets typically have low volume, which makes it tough for AI agents to deploy capital with any real speed or scale. Sports and elections are the most popular areas to bet on, but trading these events puts traders face to face with highly specialized players in arenas where AI simply can’t compete with the same effectiveness.
The kind of modeling you need to do is exactly like the modeling needed to predict a stock price, Kapila said. You don’t ask an LLM what a stock price will be one second from now.
The business models behind the promise
Understanding who actually makes money in this ecosystem is critical for anyone considering entering this space. The business models of AI automated trading platforms rarely depend on the financial success of their users. Most operate on monthly subscriptions, fees based on trading volume, or commissions on allocated capital, which means the platform profits regardless of whether the investor is in the green or in the red. This misalignment of incentives is a critical point that very few people discuss openly when it comes to artificial intelligence applied to finance.
There are exceptions, of course. Some more sophisticated platforms adopt performance-based business models, where the company only takes a cut of the profits generated, creating a real incentive for the product to actually work. But this format is less common in the retail investing segment precisely because it’s harder to scale and requires the product to deliver verifiable, consistent results over time. It’s far more profitable, from a business standpoint, to sell the dream of intelligent automation than to deliver real, auditable results.
Another important element in these business models is the use of backtesting — the practice of testing a trading strategy against historical data to show how it would have performed in the past. The problem is that backtesting is notoriously susceptible to what specialists call overfitting, a situation where the model learns the specific patterns of a historical period so well that it becomes incapable of generalizing to new data. Automation platforms frequently present impressive backtesting results as evidence that their AI works, while leaving out that the historical performance rarely repeats when the bot encounters market conditions that weren’t present in the training data. For the retail investor without the technical background to question this methodology, those numbers look like a guarantee. In practice, they’re anything but.
What this story tells us about the real state of AI in the market
What makes Jake Nesler’s case interesting isn’t the financial outcome — which was ambiguous — but the process he followed. Nesler didn’t just download an app and deposit money expecting results. He spent weeks refining the model’s behavior, testing different parameters, adjusting risk tolerance, and closely monitoring every single trade. That level of involvement is the exact opposite of what most automated trading platforms encourage their users to do, because the product’s appeal is precisely the promise that you don’t need to do anything once the system is set up.
Nesler also tested his agent on prediction markets. He gave the bot about $30 to bet on Kalshi, with instructions to research sporting events and bet on the most likely outcome. The performance was terrible. The bot did slightly better predicting bitcoin price ranges, getting about 60% of the trades right. But eventually, it lost everything.
It feels like a slot machine, Nesler said. People win and lose.
Sumer Malhotra, co-founder of Fireplace, a trading terminal used by professional prediction market bettors, recognizes the appeal but also sees the limit clearly. Agents are very emotionally cold, he said. They make decisions purely based on objective reasoning and their own constraints. That emotional coldness can be an advantage in certain situations, but it doesn’t replace the human ability to interpret nuances and contexts that data simply doesn’t capture.
The practical lesson that emerges from all of this is that artificial intelligence applied to financial markets works better as a decision-support tool than as a complete replacement for human judgment. The bot that avoided a $10,000 loss didn’t do it because it’s infallibly smart. It did it because Nesler had trained the model with well-defined risk criteria and was monitoring the results closely enough to know when to step in. That’s very different from turning on a bot and hoping it works, which is what the vast majority of automation platform users end up doing in practice.
For the retail investing market, this kind of more mindful use of artificial intelligence represents a more realistic path than the passive wealth narrative that dominates social media. It’s not glamorous, it doesn’t generate the kind of viral screenshot that racks up followers, but it’s an approach that acknowledges the real limits of the technology and uses its strengths intelligently. AI can process volumes of data that a human could never analyze manually, identify statistically relevant patterns, and execute trades with a speed and discipline that eliminates emotional errors. What it can’t do — at least with the tools available today — is fully replace the human ability to contextualize information within a complex, constantly shifting macroeconomic landscape. 🤖📉
