Podcast: AI in Asset Management: Lessons from a 30-Year Veteran

This podcast is part of The Assetizer, GenTwo's thought leadership platform.


 

What does it really take to predict markets with AI? And why do 90% of financial forecasting models produce zero predictive value?

In this edition of The Assetizer I'm talking with Miro Mitev, founder and CEO of SmartWealth. Miro has seen it all. He started experimenting with neural networks back in 1997—back when you bought forecasting software for $200 and ran calculations overnight. Today he uses these models to drive real portfolio decisions with real money.

In our conversation, we cover a wide range of topics. Among them: what overfitting actually means and why it's so dangerous, how his system spotted the COVID crash two weeks early by reading market structure instead of headlines, and why decades of experience matter more than raw computing power when you're building AI for finance.

He also shares what keeps him up at night: humanoid robots, quantum computing, and the question of who controls these increasingly powerful tools.

If you want to understand how AI really works in professional asset management—not the hype, but the hard-won reality—this is a conversation you need to hear.
 

Key Topics Covered

•    The 27-year journey from academic curiosity to live trading
•    Why 90% of forecasting models produce zero value
•    The overfitting trap that kills most AI models
•    How AI spotted the COVID crash two weeks early
•    Why the black box problem isn't unique to AI
•    The coming wave of humanoid robots and quantum computing
•    Why GenAI hasn't cracked financial forecasting yet

About Our Guest

Miro Mitev is the founder and CEO of SmartWealth, a leading provider of AI wealth management. After earning his master's degree researching neural networks for stock prediction at Vienna University of Economics and Business Administration, he spent a decade at Siemens building forecasting systems for pension fund allocation and treasury operations. He then spent 16 years providing AI-driven research and portfolio optimization to institutional clients worldwide before founding SmartWealth in 2016 to close the loop—using the same battle-tested technology to directly manage client money. 
 

Full Transcript

Note: This transcript has been edited for readability by AI and human editors. It is not meant to reflect the exact, word-for-word conversation.

Introduction

Tom: Hello everybody, and welcome. My name is Tom Lyons and this is The Assetizer Podcast. Today I'm talking with Miro Mitev of Smart Wealth, and we're going to be talking about a subject that I'm really interested in: the use of AI in asset management and investing. I'm really glad to have somebody with us who's been doing this for a very long time and knows what he's talking about.

So Miro, first of all, welcome to the program.

Miro: Thank you so much for inviting me.

Tom: It's a pleasure. As usual in this show, I'd like to start off with a bit about you. Tell us about your background and take us through the story up to how Smart Wealth got founded.

Early Days With Neural Networks (1997)

Miro: My history goes quite back. I started working with neural networks and artificial intelligence in 1997.

Tom: That's a while ago.

Miro: It's a while, yeah. I was studying economics at Vienna University of Economics and Business Administration, majoring in investment banking and statistics. I was at the final stage of my studies when I got involved with neural networks for application in credit analysis of companies. This changed my perspective because at that time I was deeply involved with quantitative models such multivariate regression, vector-autoregressive and factor models.

Neural networks completely changed my perspective at that time, because I saw how powerful these tools are and how they can be used to analyze vast amounts of data.  So, I decided to write my master’s thesis on using neural networks to predict stock returns. I spent almost two years analyzing and researching this topic, using multilayer perceptrons with different topologies and numbers of hidden layers to forecast stock returns.

The results of the stock return forecasts were very promising and interesting. I compared them with the traditional techniques we had available at the time, and in my master’s thesis I documented that using neural networks produced very good and superior results compared to those traditional techniques.

Tom: This was something fairly new then. When you say traditional techniques, you mean human analysis?

Miro: I compared it to traditional quantitative techniques like linear regression analysis —primarily neural network multilayer perceptron models in comparison with multivariate linear regression models.

Tom: Okay.

Miro: So, it was traditional linear versus non-linear comparison, and the results showed that the neural networks provided almost universally better results than the linear regression models.

Tom: What was the state of the art like then for neural networks?

Miro: It was quite an early stage. Just to give an example, there weren't many neural network simulators you could use at that time. There was a simulator that I bought for $200 - I think it was called Q-Net or something like that.

From Academia to Banking

Miro: I did the analysis using this Q-Net multilayer perceptron software. It was a lot of manual work - preparation of data and doing the analysis. But it was quite interesting to see how fast and how well these tools approximate and analyze the data.

At that time, this was before starting my next venture with the investment bank of Bank Austria, where I was responsible for analyzing stocks using these techniques for the first time. It was in a very early stage. Traditional analysts at banks were mostly using fundamental models — dividend discounted cash flow models — to analyze stocks, but not really making predictions.

Tom: What's a multi-layer perceptron?

Miro: A multi-layer perceptron is a type of neural network that you're using for analyzing historical data. Multi-layer perceptrons use historical time series. You have a neural network topology that has an input layer and can have a variety of hidden layers, and then there's an output layer that provides your output.

I was using fundamental, macroeconomic, and technical data as inputs into the forecasting neural network models. Then there were one or two hidden layers which were transforming this information using different types of transformation functions – such sigmoid functions, for example. Finally, the output layer produced the forecast by combining all the information.

Tom: And the hidden layers — you don't know what they're doing?

Miro: You cannot really understand how the information is exactly processed in the hidden layers. You can monitor what's on the input side and what's on the output side. It is very important to make sure you have controlled information coming into the models, because at the end you don't want models that are just learning noise.

You want to make sure the data you're using has some kind of significant causal relationship — from the fundamental, macro, and technical side — to the target you're going to forecast.

Tom: I think we'll come back to this question of what goes on in the black box a little bit later. But maybe you want to pick up the story. So you're at the Bank of Austria...

Siemens Years

Miro: I got an offer from Siemens to build up and lead a team focused on developing predictive models.

This was a very big project at Siemens at the time. They had a very strong demand for advanced software and analytical services to support asset allocation for pension funds as well as FX-hedging in treasury operations. Siemens had already developed sophisticated technologies over many years, which were being applied across a wide range of industries - including the optimization of nuclear power plants, electricity demand forecasting, and traffic flow optimization in smart cities. In those early days, this form of AI was already powering numerous large-scale industrial applications.

It was very interesting to have this sandbox where you can start working, testing, doing things with a lot of tools at your disposal, because at that time there was very limited availability of such systems that you could use.

Tom: And you were going to show them how to use this stuff for finance?

Miro: Yes. Our job was to start using this technology and applying it for the financial markets.

Tom: Okay.

Miro: So, we began building forecasting models using neural network frameworks and technologies. This included a wide range of neural network architectures and forecasting approaches, employing different mathematical algorithms.

Models for forecasting like neural networks, factor models, vector-autoregressive models, regression models, genetic optimization algorithms, fuzzy logic - many different types of so-called machine learning techniques.

Tom: Mm-hmm.

Miro: These models were largely based on operations research and machine learning techniques and algorithms. We adapted them specifically for forecasting purposes, and the resulting forecasts were subsequently used as inputs in later stages to optimize portfolio asset allocation.

Tom: Okay.

Miro: Yeah, that was how I first began working with artificial intelligence 25 years ago.

The Path to SmartWealth

Tom: Take us up to Smart Wealth — what that is, how it started, and what you're trying to achieve.

Miro: It was a long journey before I ultimately decided to found Smart Wealth. Over nearly 16 years, I had the privilege — and consider myself very fortunate — to gain deep experience delivering high-level research and software-as-a-service solutions to a wide range of leading institutions around the world.

During my time from 2000 to 2010, and then later another six years, I saw the advantage of all these technologies and the results we provided to many clients worldwide, and also the performance which was generated using this technology. Because I already had this experience and all the tools at my disposal, we knew what was working and what wasn't working. A lot of research has been done over the 16 years, and also a lot of learning by doing and improving the technology.

It's not just one model — it's a technology that has so many components and so many small things that you have to fine-tune over the years. You have to solve a lot of problems over these years. And of course the technology is developing all the time, and we keep building and improving it.

This made the decision quite easy. Let's build up Smart Wealth at its core around artificial intelligence, leveraging technology that had been developed, tested, and proven over many years.  We would use this technology to create our own products and to manage our clients’ capital directly based on these AI-driven processes. That was the decision.

Tom: Okay. Before that when you were advising all these other companies, that was at Siemens?

Miro: Partially at Siemens - 10 years was at Siemens. And then after that there was a wealth management company that I joined with part of my Siemens team.

Then in 2013, I made a spinoff into an index start-up where I continued the old Siemens business by providing research, optimized portfolios and forecasts as a service. In 2016, I decided to take the final step and close the loop - delivering the full investment value chain, from forecasting and optimization to the implementation of investment decisions through proprietary products. At that point, becoming a regulated investment manager was a natural and necessary step.

Tom: Right. So this is now you're actually managing client funds at Smart Wealth.

Miro: Exactly. We have been managing clients` money since 2019 based on our AI technology, and we have also our own products that we're working with you on together.

Tom: Absolutely.

The Evolution of AI Technology

Tom: I would love to get back to that at the end of the program. But what I'd like to stick with — because the story of AI fascinates me as a layperson — you've seen the technology evolve already for 30 years, at least in a hands-on way. Maybe you could tell us a little bit about how that evolution went. I just want to get a sense of the change, of where the breakthroughs are, of what's different now. Obviously everybody now knows about LLMs, which we're not talking about necessarily here, but I'm just curious about how that trajectory went from somebody who was there.

Miro: Yeah, it's a long journey. At the very beginning, the algorithms and the mathematics behind didn't change too much. But the infrastructure around them changed, and this changed for good. This kind of technology requires a lot of data, but also a lot of history. When we started 20-25 years ago, there was valuable data history, but not so long. Primarily for American stocks there was information, but for many others it wasn't available. To build reliable models, you need data, because everything is based on data. The history over the last 20 years — you got so many additional data points, which makes the testing and developing of your models better than it was 25 years ago.

The second very important change is the data connectivity. At that time there weren't cloud solutions. We used 5,000 - 7,000 computers in Siemens overnight for calculations when employees went home. We were using this for calculations.

The computer chips and the speed of calculations increased dramatically. Nowadays you can — and over the years you could — test much more hypotheses by just having a higher speed of testing.

Tom: Okay.

The Challenge of Overfitting

Miro: Of course, on one side it's good because you can test so many new ideas and hypotheses. On the other side, it's very challenging because a big problem that people very often underestimate is overfitting. It means that the more hypotheses you test, the more likely you will find something that will be working.

When you're building forecasting models — any models — you always have a very strong risk of overfitting a model. You have to train the model, and you have to do this with data —with so-called in-sample data first. That data is known, presented to the model, and you have to calibrate it, you have to fit it, you have to see that the model is learning with this data. Of course you have to select the model finally, and you always tend to select models that are working well.

The more models you have and the more hypotheses you're testing, the higher the risk is that you will be picking up a model that is performing very nicely but is actually an overfitted model — a model that learned the noise, that fitted to the data and is not stable over time. So it will very likely not work in life, in the real environment.

Tom: Okay. So this is something that's telling you what you want to hear.

Miro: Yes, in a way. When you're starting to use it, then it's not usable anymore because it learned the noise in the data. This is true for any kind of models — therefore it's very, very important to select and work with models where you understand what they have learned from the data.

At that time when we started doing this 25 years ago, the term was data mining. This means more or less that you have such a huge amount of data which of course can be perfectly analyzed using neural networks. But if you don't have enough historical data for your time series analysis, then you have so many free parameters in the neural network model that you are actually overfitting by just training the model. So, you have to be very selective about which data you're using.

This is not easy, especially in the financial markets where we're using the technology. All markets are very highly correlated — it doesn't matter which kind of asset classes. The markets are interconnected and interrelated, and this kind of market structure is changing all the time. It's not constant.

So it's very difficult to find stable lead-lag relationships — leading indicators that you have to identify in the data in order to use them as predictors. Because you don't want to say what the price is today of the S&P 500 by knowing that the euro-dollar is strong today and the interest rates go up today. You want to know what is the implication from the today`s market environment — all this data which is interrelated with each other — what is the implication for the price of the S&P 500 for the next one week or next one month.

Tom: Mm-hmm.

The 5% r-square Edge

Miro: Based on this information today, you have to make your decision. And there's so much noise in all this data. It doesn't matter how much data you're analyzing and using —the noise is huge. So you're actually able to explain probably from 100% to be perfectly right in terms of out-of-sample R-square of the predicted returns, probably around 5%.

Tom: Okay.

Miro: So it's very, very few.

Tom: Okay.

Miro: And if you have transaction costs involved and you don't have an efficient system and process working on that, then this very small advantage that you're able to forecast - it goes out.

Tom: Okay, so 5% R-square is the advantage you're giving yourself.

Miro: Normally 90% of the forecasting models of the companies that are involved with forecasting financial markets produce 0% out-of-sample R-square. So, actually they're forecasting nothing - it's 50-50 bet.

Tom: Okay.

Miro: It's very difficult to be in this 10% that have the ability and are showing over the years to be able to have positive out-of-sample R-square or explanation of the returns that you are forecasting consistently over time. And if you have this kind of 5% out-of-sample R-square, it is enough to be over 50-50. So, we'll go to 55-60% in terms of hit rate. Because this gives you the edge — only this gives you the edge to make an outperformance against the benchmark and against the competitors.

Tom: Okay.

Miro: Doesn't sound like a lot—

Tom: It doesn't sound like a lot, but it's a lot.

Miro: Yeah. Because 80% of actively managed funds underperform the market.

Tom: Yeah, this I know.

Miro: Yeah, and this is a fact. Why? Because they're not using forecasting technology.

Just minimizing risk, or optimizing portfolio based on risk is not enough to outperform the benchmark because you're not covering your transaction costs and you don't have more information than the entire market. So, it's a combination of both. You have to have reliable technology and forecasting models, but you have to have a very efficient and reliable process in order to rebalance the portfolios. It doesn't help you if your execution is too late or wrong.

This was one of the reasons why I decided to set up Smart Wealth, because I've seen over these 16 years of providing valuable information that it's valuable information if you're correctly implementing it — you can generate excess return. But I've seen so many times it wasn't really implemented correctly or was overruled. For example, the forecasts were overruled by human decisions.

Tom: Yeah, okay. So there's more going on than just the analysis part.

Miro: Absolutely. At the end of the day, it is not just an algorithm. When people think, "Okay, it's AI," it is a system that is built on different steps integrated with each other and following — or working in full synchronization with each other.

Dealing With Irrational Markets

Tom: One of the things that always interests me — you mentioned about not having more information than the rest of the market has. What's always interested me, I know it's a bit of a naive question but I like to ask it anyway, when you're doing any kind of quantitative analysis of markets, is how you deal with the irrational side of market behavior. Because there's plenty of that, especially at certain moments of crisis or volatility. So how does your model, or how do you approach the absolutely crazy human side — the behavioral side — of the market?

Miro: Yeah. It's a very interesting question. It's very easy for me to answer because our models are based on hard data — data that as I mentioned has significant ability to forecast. So, this kind of lead-lag relationship is very important. Now, a lot of information is coming onto the market on a daily basis. The financial markets are semi-efficient — they're not perfectly efficient. A lot of the information coming through the media is publicly available — this information tends to be reflected in a very fast way into the price of the assets.

Following this information, in many situations people start overreacting in positive or negative way.

This is something that in many cases generates a lot of noise. Your forecasting or your models need to work with this kind of noise and to be able to just ignore it in very often situations. Even if you have a perfect forecast, the structure of the market is, for example, you have closing hours. You have a lot of information coming after the closing hours—for example, Liberation Day. The information was released after the closing day. So what happens on the next day? You have a gap down with 10%, and what happens one week later? You have again information which is released during, or one hour before the closing of European markets, and then you have a gap up of 10%.

In this second example, you could participate because it was during the day, and if you were reading and hearing the information released from the administration, then you probably would be buying. But this information is just not reliable in the sense of "what happens next?" You have to have a model that has information based also in the history that can prove—you cannot build a model just on one signal, one information.

On the other side, you cannot trade on daily news and daily information very successfully because the nature of the market is very complex. You have gaps that you cannot close, and you have a lot of transaction costs if you're doing daily trades.

Actually our technology is trying to navigate through this information to avoid a lot of noise and a lot of such psychological factors. We are considering technical factors, which are also very important parts of the input factors, but these input factors are in combination with the fundamental and with the macro. So you have to weight them. Sometimes they have a little bit stronger weights, but sometimes they have less weights, but you're not just trading based on only one factor. The strongest thing is that you have a combination of different types of factors, of leading factors considering different types of information. And this makes it stronger and more robust.

Tom: Okay. And the model has seen—it's seen in the past things that happened, for example.

Miro: Yeah.

Black Swans and Market Crashes

Tom: Can it tell you if something is a bubble?

Miro: Yeah. It's very difficult to forecast something that's made by political or some other decisions that are not directly market related. But you can identify situations like in COVID. Our forecasting models changed to negative on February 24th 2020 — so actually two weeks before the markets really crashed. The exposure to equities especially, but also to risky assets, was reduced. The cash was increased. This was good during the COVID-crash in March 2020 — but this was not because our technology was forecasting a black swan.

The information was already existing in the market, and COVID didn't start in March. In fact, it started already in October 2019 the year before. Okay. So, signals were already in the market.

Our technology analyzes the relationships between the different asset classes, between the different fundamental and macroeconomic and technical factors with respect to what we want to forecast. Based on causal relationships the information in the hard data is reliable over time. It's changing of course, but still those causal relationships are important. At some time, they may be disturbed.

For example, we have seen some kind of regime change at the beginning of 2025 before Liberation Day, when there was asymmetric information on the market and a lot of uncertainty. But the market couldn't really react normally, and you couldn't see that in the hard data. The volatility, for example, which is also an interesting factor for risk, was actually not so high at the beginning, but there was a lot of uncertainty. In such market environments, the relationships that the models have learned over the years are disturbed for a short period of time.

Then it's important that the models can work also during this time on the risk management side — that this kind of market environment is identified that the quality of forecast in such periods is probably not going to be as good as usual.

Then of course we've seen negative forecasts and the exposure to the risk assets was reduced again. We had also an increase in the cash positions, which was also good because reducing risk in such uncertainty is very good. And then in the middle of April 2025, the forecasts changed to positive again.

Tom: So basically you just get out for a bit.

Miro: Yeah. For a couple of weeks we were increasing the cash positions. And then starting from middle of April 2025, we increased the positions again to the risk assets. And even in the booster product, we increased also the leverage a little bit.

Generative AI and LLMs

Tom: That's all very fascinating. Another question that I had on my mind—what you've been talking about up to now is really machine learning and stuff like that. Is there room now in your type of analysis for data or signals coming in from gen AI, from large language models? I don't know, different—is that a technology that can add to your data sets in a meaningful way, or not?

Miro: Yeah, very good question. Absolutely. As you know, we built also a scientific board, and we are very strongly involved with research. Part of the job of the scientific board is to develop also talents and cooperation with universities where we can test or we are testing and stress-testing the existing technology. It means not changing the winning horse but trying to make it better by introducing additional technology which could potentially lead to significantly better results.

So yes, I think that generative AI technologies, LLM models, can add some value, especially in making the technology probably more adaptive on the short run. Meaning that LLM models is something that we have tested many, many years ago. We haven't seen that they can be entirely used as forecasting in the financial area, but of course they have very strong applications in many other fields.

In the financial field, they have to be tested probably again. Now we have much more data, we have much faster possibility to calculate and test these models, they're much more advanced now. But the evidence so far — and we're not the only ones — recently I heard from Citadel that they also confirmed that they're not really seeing the positive impact that they were expecting.

But we are just testing and we have to test it anyway in order to see if there is a significant improvement over an entire economic cycle – including a recession and a recovery. So this is what I mean. Not just in a very short period of time. Then of course we will test and run it. And as soon as there is a successful live run with live money for at least two years, then of course we could consider applying it in the future as an important part of the technology.

Tom: Right. With a different approach, obviously.

Miro: Of course. At the end of the day, what we are trying to do is we are trying to forecast. There are so many ways you can go wrong, and there are so many ways you can get to the right forecast. But there is only one right forecast.

Tom: Yes.

Miro: So now you have to test your models and see which of these models are actually producing the right forecast? And we have technology that produces good forecasts, but of course if we find an additional variable which can add more value to improve the out-of-sample forecast quality, of course we will be using it. But it's not so easy at the moment. We don't see this evidence, but of course we will be testing it again.

Tom: I know.

Miro: I was so happy when ChatGPT came out. I mean, it's an incredible efficiency improvement tool.

The Importance of Experience

Tom: Yes. I mean, I use this stuff for marketing communications since it came out — every single day. And I find it a mixed bag because if you don't already know what you're doing and have that experience — and I don't say that just because I'm an old guy — but I've noticed these things can lead you in the wrong direction. I also notice with writing or with communications stuff, you can go — so I always feel bad for the young ones because I'm thinking, okay, I can spot this because I've been doing this 30 years. And if I was just coming up now, I'd be more afraid I could be led astray.

Miro: Mm-hmm.

Tom: Because the models that I've worked with — I mean, they have a tendency to tell you what you want to hear a little bit, or you can almost influence the answer in ways you don't know you're doing by how you ask the questions, how your prompts are, and all this other stuff.

Maybe that brings us to another question we wanted to talk about. From what you've been describing up to now, it seems to me — I was expecting a very technical conversation and we're having that — but it seems to me there's a lot of art involved. And to what extent is your experience important, to sort of know when to say, "Wait a minute, this can't be right"? Because it seems it's not just analytical, or am I wrong?

Miro: Yeah. I mean, experience is incredibly important. Experience and the domain knowledge. I've been doing this for almost 30 years and nothing else — I didn't change the industry or main field of business.  So, this kind of consistency certainly helped me to understand not only the technical part of the equation, but also the fundamentals - how the markets and the economy are working together.

The combination of these two—the technical application and the microstructure of the capital markets—is incredibly important. I think building forecasts, building models for financial markets, you have to start from the very beginning. This is what I'm very happy about—starting from the very beginning with the data analysis, the models, then going to forecast, then going to optimizations, then going to products, then going to executions and so on. So all this kind of journey is so incredibly important in order to find the right application, the right model for the right target, because there are so many.

And you don't have the time as a researcher, and also in a company you focus normally on creating a good product, you focus on margins, profit, whatever. But you don't have the time for testing and developing and these kinds of things.

Therefore, experience can really help you to save so many years, even if you have so many tools at your disposal, because there are so many possibilities. And you have to narrow down and say, okay, for example, just to give an example: you can use an algorithm to find a local minimum. What kind of algorithms will you be using? What kind of environment will you be applying in order to come to this? And then you can test one, two, three, four, five, and then you can decide. But you already know what is to be used, and this saves you a lot of time.

The Perils of AI

Tom: Fantastic. You know, what worries you about AI? I like to turn the conversation to that, and I guess on two aspects. First of all, in terms of just your analyses and stuff like that, but maybe stepping back into the bigger picture.

Because you've written about—there's power and there's perils in this technology, and a lot of what goes on we don't understand. That's the nature of the tech. So what worries you?

Miro: Well, so actually what's worrying me—I mean, AI and all algorithms is something that you can program. You're programming and you're controlling as a human. And of course, even if you don't understand everything which is inside, you still have some kind of control of it. But we have to think about the future in the sense that we will be experiencing next year and the next maybe two, three years a lot of robots, a lot of humanoid robots that will be coming.

They will be of course programmed for specific things, following specific rules—not harming people, behaving as they should, and so on. But of course you never know. There might be somebody hacking them or changing this. And the next war, which we will be definitely thinking about very much in the near future—I hope not soon, unfortunately—I think it will be soon. But it's going to be—you and I, we've seen this many, many years ago, Star Wars. So it is going to be—we have experience also now with drones and so on. So a lot of these powerful tools can be used also in a very bad way.

This scares me a little bit, that we should be having some kind of protection so that this cannot happen—that very few people can be actually in the power of this kind of control, because this is very powerful if you're controlling this kind of tools and algorithms.

But of course, we have to think about the next generation, which is not just AI—AI is the first step maybe. We have the next six Gs coming, the humanoids are coming, and then in five to 10 years we have quantum computing coming. And that's the real breakthrough, I think. Of course it's changing the way how you are solving issues, problems that you need many, many years to solve—you're just solving this in a second or a minute. And this is changing the way how we are living.

And of course it can be also used for non-human or non-compliant reasons, unethical reasons. And this is something that sometimes I'm a little bit scared about.

Tom: Yeah, me too.

Miro: Yeah.

Tom: But it's like—not much you can do. It's going to happen.

Miro: Yeah, we just need to make sure that there are people in power of these decision-makings. It's very important to have the control of it—not keep it out, but that the machine is actually not starting to program you as a human. That would be something that is quite scary.

Tom: Which, by the way, it's not hard to program a human if you know how to do it. It's quite easy.

Miro: Yeah, absolutely.

SmartWealth Products and AMCs

Tom: So Miro, this has been an absolutely fascinating conversation so far, and I could keep going all night. We're already at 44 minutes, which is long for us, but that's also nice. But I do want to get back to Smart Wealth and talk about the product.

So you guys are a client of ours here at GenTwo?

Miro: Yeah.

Tom: And maybe we can discuss a little bit about how that came about and what the AMC is that you have with us. I think it's interesting—this is not here to market GenTwo, but I do like to talk about innovative products. And obviously the people we work with, other people we know, know what's going on. So I'd love to hear a bit about that.

Miro: Yeah, I mean, actively managed certificates are—as they also very often say—they revolutionized the way of how asset managers in Switzerland are doing business. So it's a very cost and tax-efficient way for managing our clients' money.

This is the reason why we, in a very early stage, decided to be part of the ecosystem and one of the pioneers by using actively managed certificates for managing client money. This is how we started working with you together, and we are very happy with this cooperation. We are constantly increasing the number of actively managed certificates that we do together.

It is not only the efficiency, it's also for the clients the security, which makes a lot. Using actively managed certificates, you can test in comparison to a normal fund which is very expensive and not such cost-efficient. It's a very nice way to test some ideas, to build track record, and to provide fast access to different topics and ideas to your clients. This is very, very interesting.

With respect to our certificate or to one of our product certificates, we are very—not only early stage in the use of the actively managed certificates, but we are using the entire automation on our side, which was very important. Our fully automated investment process from forecasting generation, optimization of portfolio, but also execution and so on—it's also fully automated. Also generation of the setup of the AMC and also reporting and all this stuff. And this is also one of the very important reasons that we see this kind of automation and innovation on your side with your system. Therefore we are very happy with the cooperation.

Tom: Excellent. And the product itself is?

Miro: One of my favorite products follows a multi-asset strategy using a so-called core–satellite approach. Around 50% of the assets are managed through highly liquid ETFs from leading ETF providers. The remaining portion consists of the so-called satellites, which are built from individual equities — specifically, the most liquid single large cap stocks in the world.

Tom: I see. Core-satellite approach.

Miro: The asset allocation is driven by our proprietary AI forecasts and optimization technology, with fully automated order execution.

We also have the ability to increase exposure to individual stocks by using leverage of up to 35% on the portfolio level, which becomes particularly attractive when our forecasts are strongly positive. In such cases, we can enhance performance by increasing leverage to these single stocks. Conversely, when our forecasts are negative, we actively increase the allocation in cash, treating cash as an important asset class and a safe haven.

Tom: One of the things—you don't see market dip as a buying opportunity?

Miro: Not necessarily. We don't see also the market highs as overbought or bubbles. No. We see our forecast — they are telling whether it's a buy or it's a sell opportunity. It's based only on the forecast.

Tom: Okay. Yeah.

Miro: Just remember, people are talking about Nvidia. I know that many important and good top-tier asset managers—they're talking for years that Nvidia is overbought, overbought. But Nvidia is still going up and up and up.

There are so many factors that are influencing—I wouldn't say AI bubble. It's not a bubble. It's just starting and it's real. It's infrastructure. You need this infrastructure, and the Americans are doing a very nice job because they're taking the lead here and they're getting so much investments in this infrastructure. I saw this at Siemens when we were building infrastructure. You need it in order to be successful over the next 20 or 30 years.

And this is—I think now it's the beginning of that. I wouldn't say of course it's normal. There will always be recessions, there will always be some kind of market corrections. And this is the nature of the market. It doesn't have nothing to do with an AI bubble. It's just the normal way how the economy is working.

And there are many other factors that are influencing the entire economy—like unemployment rates, like interest rates, like stock market prices, commodity prices, and many different factors. If they're influencing the entire economy and the economy needs to go a little bit in a recession in order to solve the issues of unemployment or whatever issue of high inflation, then it will do this. This was always in the past. But we have to be in a position to focus, to see such kinds of signs that this market environment will come. And if we see something like three to six months ahead, then of course we'll be reducing the positions to risk assets at the right time.

Tom: Okay. Got it. And did I understand correctly—the algorithm or whatever is doing the trades itself?

Miro: Yeah, of course. We built a proprietary fully automated trading and execution platform. It's executing all portfolios that we have in just 15 seconds. We only trade the most liquid stocks and ETFs in the world. Our portfolio managers are just monitoring the execution and rebalancing process, making sure that everything is going well. Sometimes trades may be rejected — the reason needs to be investigated on the broker’s side. But normally 99% of the trades go through, fully automated. So, it's a nice feature to have, and this makes the entire process very efficient.

Closing

Tom: Well Miro, this has been an excellent, excellent conversation. I really enjoyed it. Thanks so much for coming across the lake to visit us tonight.

Miro: Thank you so much, Tom. It was a great pleasure. Thank you so much.

Tom: You're welcome. Thank you.