LISTEN TO AUDIO VERSION:
About Tungsten Capital
Tungsten Capital is an independent, owner-managed investment boutique based in Frankfurt/Main and Cologne. The TRYCON AI Global Markets fund, managed by Tungsten Capital, is one of the first funds to rely entirely on artificial intelligence for its investment process. It invests in 60 global liquid markets, consisting of stocks, government bonds, currencies and volatility. The multi-asset long/short strategy can profit from both rising and falling price movements.
Pablo Hess is one of the two fund managers of Tungsten TRYCON AI Global Markets. In addition to portfolio management, he focuses on research and the development of multi-asset trading strategies, especially for exchange-listed derivatives, with special emphasis on sophisticated quantitative models.
Hello Mr. Hess, about thirteen years ago, TRYCON AI Global Markets was launched as one of the first AI funds in Germany. How did the idea of a purely artificial intelligence-based fund emerge?
Our origins lie in systematic trading strategies, more precisely in quantitative methods and trend following. Trend following, in particular, is a very inefficient way to handle the available data. Therefore, we have been carrying out research since 2000 on how to apply AI technology to financial markets to allow for far more in-depth analyses. Our goal was to evaluate the data more comprehensively and thus gain a different view of what was happening in the markets. Simply put, our AI can both make use of momentum and anticipate directional changes in the markets, achieving a near-zero correlation to global stock markets.
What can artificial intelligence contribute to asset management?
There are numerous quant funds that rely on sophisticated mathematical models. Where exactly is the boundary between a “traditional” quantitative investment strategy and an artificial intelligence-based one?
Especially in machine learning, the distinction is indeed somewhat blurred. For one thing, AI can handle massive amounts of data. Furthermore, it’s particularly capable of determining nonlinear correlations in addition to linear ones. It also gets by without a hypothesis that would need to be examined and thus distort the model with its own presuppositions and vagueness. And last but not least, the AI algorithm itself generates the trading strategy that will be applied in the end, whereas in traditional quant models this is often defined by the developers. I would say that this is an entirely different approach.
During the coronavirus pandemic, stock-market prices collapsed all around the globe. Actively managed portfolios also suffered heavy losses, whereas the TRYCON AI fund impressed with positive returns. What can artificial intelligence do that portfolio managers cannot?
First, AI can calculate probabilities in a strictly rational way. I think it’s totally understandable that this ability was lost on many human portfolio managers in the panic of the Covid-19 crash. What’s more, the algorithm can dig deep into the past and at the same time consider countless variables in a decision. AI can ask which periods, however far back in the past, are comparable in abstraction to the current situation and provide clues as to which upcoming market reaction is the most likely. And – as already mentioned before – AI can make use of nonlinear correlations. We humans are not very good at handling large numbers of variables when nonlinearity comes into play.
Limits and challenges of using AI
You mentioned that AI can dig deep into the past, but doesn’t this also mean that only trends known from the past will be recognized by AI? To what extent can AI identify or predict future trends?
I would say: decisions made by humans are also based on experience and what they have learned, so data from the past is processed to make abstract statements about the future.
However, price movements are often due to emotional factors and herd instinct; consequently, sharp jumps in stock price may not appear logical and comprehensible. Can AI handle this “human component” of investing at all, or might discretionary decisions by portfolio management be reasonable now and then?
The markets are driven not least by fear and euphoria. In our experience, this isn’t problematic for AI, on the contrary: AI takes advantage of the market participants’ irrationality but acts quite rationally. Therefore, this peculiarity of the markets doesn’t require any intervention by portfolio management; it would probably even be counterproductive.
Let’s move on to the challenges of using AI. The successful use of AI systems strongly depends on the “signal-to-noise ratio”. If you feed algorithms with unfiltered data, the model is disturbed by random correlations (noise), while actual cause-effect relationships (signal) are lost in relation. As a result, the model makes inaccurate predictions. Can you explain how you deal with this issue?
This is indeed a major pitfall when using AI, especially in financial markets, therefore you should not approach this task expecting a “plug & play” functionality. Successful use requires deep knowledge of the specifics of financial markets and of how AI algorithms work. As demonstrated by our long-term track record, we have addressed this problem with a set of measures. How to tackle this challenge is certainly part of the key to AI’s success in asset management.
Using AI models often means having to struggle with the non-explainability of predictions. They are therefore often referred to as black boxes. Can you explain each investment decision made by your AI model?
In most cases, using AI means taking the route of increased complexity. When AI is used in other areas relevant to everyday life, such as autonomous driving and navigation, in medicine, and many others, the focus is not on explaining how the results have been obtained, but on their concrete benefit. In almost ten years of history, we have given proof of this benefit. By the way, this does not only apply to the use of AI: hardly anyone understands how exactly the alternating current we use every day works. My point is this: we use technology in many areas to enjoy its benefits, even if the processes in the background are not easy to grasp.
However, the lack of transparency of AI models is considered one of the reasons why self-driving cars are not yet on the market. Although AI usually delivers good results, it makes fatally wrong decisions in individual cases. In rare cases, stop signs are confused with other road signs, for example. Such errors can only be prevented if I as a user know the reason behind the AI’s decision. With regard to AI-based trading systems, isn’t it therefore important to understand how the system arrived at its decision in a given case?
I wouldn’t refer to our models as black boxes, but rather as gray boxes. Because for each model, we know the factors that can influence its decisions. By the way, not only artificial intelligence is a black (or gray) box, but especially the human brain. We are currently unable to comprehend a large part of decision-making processes in the human brain. The decisive factor is: who makes fewer mistakes, AI or humans? In the case of autonomous driving, humans currently appear to make fewer mistakes – by now. With regard to financial markets, we are convinced of the added value that AI models can provide.
Of course, we also give the AI a strict risk management framework within which it’s free to operate. We believe it’s a prudent approach to separate signal generation and risk management. I would like to give an example of how this can be useful elsewhere, and it’s from the field of IoT: if the refrigerator reorders butter, I don’t want it to order the exorbitant amount of 800 kg of it either. Admittedly, in that case, there would probably be something wrong with the underlying model …
So, the AI is not capable of engaging in trades autonomously, which means these must first be approved by a human?
In our case, the trades are still approved by a human. We do this by checking for plausibility and not by questioning the respective trade.
Future of AI in asset management
The fund industry is a highly competitive market. Investors can choose from an indeterminable number of investment options – active and passive investing, many different asset classes, various investment strategies, sustainable funds and much more complicate the investment decision. Pure AI funds are still rather niche products. What is your expectation for the coming years? Will you be able to stand out from the competition with the combination of classic investment areas and AI?
Indeed, the field of AI-driven funds is still quite easy to survey at the moment. Today, this may even be an advantage for investors, because they can choose the right product for them from a (still) manageable range.
What is special about our way of using AI isn’t only the positive alpha, but also the low correlation with stocks and bonds as well as the moderate volatility of about 5 to 8%. This can hardly be achieved when using AI with traditional products.
Is the fund suited for institutional investors and private investors alike?
Thanks to its moderate volatility and outstanding diversification characteristics, the fund has proven to be beneficial from an economic perspective for most of the portfolios we have examined – this is demonstrated by the Sharpe ratio that has improved over the course of our now almost ten-year history.
Does the fund only take full effect as an addition to classic investment products for risk diversification, or does it also make sense as a stand-alone investment?
Although the fund is also attractive as a stand-alone investment, it really does perform best in conjunction with traditional investments.
However, there is one restriction: unfortunately, in the past, people did not always embrace innovative and progressive investment processes. Yet, since ChatGPT has been receiving considerable attention, we have felt a significant change. People seem to be becoming more and more aware that they are already using AI as a matter of course in other areas of their lives and that it can offer them many advantages. So why not also in asset management, which, after all, is essentially a domain of data processing?