The internet was becoming mainstream in the late 90s, but Miro Mitev was head-down exploring something that wouldn’t become popular for decades: AI.
Now an asset manager, Mitev was an early adopter of AI in finance after discovering the capabilities of neural networks in 1997 while studying at the Vienna University of Economics and Business.
He told CNBC he saw the potential of neural networks for financial forecasts. “I fell in love with these kinds of possibilities,” he said.
Mitev spent his 25-year career forecasting for banks and tech firms like Siemens. He founded SmartWealth Asset Management, whose decisions are made entirely by a network of AI systems. Its latest fund, IVAC, is eyeing $2 billion in assets under management and has an annualized returns target of 14-15%.
Despite no human involvement in the AI’s decisions, Mitev said that “humans are the most important part of the equation” as they’re the ones selecting training data, inputting variables, building the parameters, and consistently tweaking the model.
Once a model is created, “it’s very dangerous to start intervening,” Mitev said. Indeed, trusting the model is his golden rule, he added.
Instead, humans should ensure that there are no errors in the data or calculations, and introduce new data so that the model is up to date.
“The worst is to overrule the results, and this is what happens very often,” Mitev said, adding that people “don’t trust” AI at first. “Even if we, as humans, don’t see the result now, if we look back after two months, three months, we say, ‘Oh, actually, we were wrong,'” he added.
The forces driving the market — optimism, pessimism, speculation — are very human. Even the European Central Bank has warned that the current AI bull run may be driven not by detailed technical analysis but by fear-of-missing-out.
Mitev said that taking the emotion out of investing proves better results; SmartWealth Asset Management has seen gains of 407.63% across a 10-year period to Nov. 1 2025, compared with an industry benchmark of 145.34% over the same period, according to a graph a representative for the firm shared with CNBC.
It’s “not possible” to know what will happen in one year, Mitev said, but he can see up to one month ahead with his model. “Evaluating this information and making informed decisions based on this consistently proves to be providing better results than the human.”
The constant monitoring and introduction of new data are important points, given that AI systems do “hallucinate”: generating false information. Mitev said models’ mistakes were down to “overfitting,” data issues or model misspecification.
Overfitting is where the algorithm pays too much attention to what Mitev called “noise.” He said this was data “which is not meaningful” because it doesn’t reveal a true cause-and-effect relationship with stock performance.
Rigorous design, validation, and live environment testing, serve as an antidote to this, Mitev added. It means that, although his fund strategy is executed entirely by a series of algorithms, humans still play a crucial role in making sure it’s effective.
“It’s actually a process that evolves over years … and this is the reason why in-house development of these kind of technologies is very important,” he added – especially for anyone looking to differentiate their AI play.
