Share this Article & Support our Mission Alpha for Impact
This is some text inside of a div block.
This is some text inside of a div block.
This is some text inside of a div block.

Algos must be simple - Interview with Prof. Dr. Florian Artinger.

Post by 
Text Link
‍Florian Artinger is co-founder and CEO of Simply Rational, Professor of Digital Business at the Berlin International University of Applied Sciences and an associate researcher at the Max Planck Institute for Human Development. His company Simply Rational combines artificial intelligence with behavioural research to enable customers to make better financial decisions.

Complex algorithms and machine learning models work well when they operate in a stable world. A good example is chess, where the rules will be the same tomorrow as they are today and yesterday. A human decision-maker no longer stands a chance against a computer here. But things look different on the stock exchange. Conditions are more difficult here: High complexity, poor signal-to-noise ratio, regime shifts, excess volatility and fat tails cause considerable problems for human decision-makers and algos alike.

How can investors make fundamentally better decisions?

Prof. Artinger: Many people tend to look for equally complex solutions in complex, highly dynamic systems. But in the financial markets it is helpful to reduce the complexity instead and focus only on a few criteria that are as simple as possible. At the same time, this makes it possible to develop a better understanding - and also to make better decisions. However, there is no standard solution. Decisions should be based on individual investment goals and, above all, take into account the risk preferences of the individual.

In the financial markets, it is helpful to reduce complexity and focus only on a few criteria that are as simple as possible.

Do you have an example of what a simple solution might look like?

Prof. Artinger: Yes, a rather well-known example, actually. Harry Markowitz, the founder of modern portfolio theory, did not follow the mean-variance model he developed when making his own investment decisions. Instead, he applied the simple 1/N heuristic and invested his money equally in different stocks. He did this intuitively, and it turned out to be a clever approach: As later studies showed, this approach actually produced better long-term returns than the mean-variance model or its modern variants.

Intuition is a good keyword. Experienced experts intuitively know what to focus on and what to ignore. But can intuition also work in the financial market?

Prof. Artinger: On a medium- and long-term investment horizon, it shows that experts are quite capable of filtering out the relevant signals. Moreover, they can identify good heuristics and combine them cleverly. In this way, it is also possible for human decision-makers to perform better than an algo. This is especially true in less competitive areas and niches of the market. Nevertheless, chance always plays a role in the objective assessment of individual results on the markets.

Experts are quite capable of filtering out the relevant signals.

What is the difference between good intuition and mere gut feeling - and how much experience does it take to distinguish one from the other?

Prof. Artinger: Developing intuition requires a learning process with successes and failures. By gut feeling, on the other hand, we mean the initial reaction of an inexperienced investor. Advantageous intuitive filters are therefore always based on extensive experience. And with increasing experience, the negative influence of classic behavioural finance effects on investment results also decreases if one has already experienced similar situations several times and learned from them. However, it takes time, the right learning environment and, ideally, professional mentoring or coaching to accelerate this process. Just as people long ago learned through social interaction which berries are poisonous and which are not, today we learn to behave appropriately in the markets.

It is said that the best, but also the worst investors are discretionary decision-makers. How do you see the chances of success of humans and algo in comparison?

Prof. Artinger: Certainly, the dispersion of results is greater overall for human decision-makers than for algorithms. Beginners who only decide according to gut feeling will probably do badly, and experienced investors with simple, systematic if-then rules and processes will probably do well. A systematic approach primarily protects against mistakes and offers a ready-made plan that can be used as a guide. And that is already quite close to algorithms. But the best would be the combination of both, what we call augmented intelligence.

Can you please explain that in more detail?

Prof. Artinger: What is meant here are simple, transparent algorithms that people can understand and interpret. These are then consciously applied in the context of a specific market environment. At the same time, an expert maintains an overview and can intervene if the framework conditions change significantly, for example due to a structural break in the market. This is what a functioning combination of human and algo in trading strategies could look like.

Augmented intelligence is a functioning combination of human and algo.

This combination requires expertise. Does that mean for the average investor that algos are the better solution compared to own decisions?

Prof. Artinger: For the broad mass of investors, rule-based, systematic investment decisions would certainly be optimal. Of course, low costs, diversification and a sufficiently long investment horizon are also important. In the simplest case, however, an "algo" can already consist of regularly investing in ETFs via a savings plan. For more advanced strategies, you need the right tools as well as a strong understanding of statistics. As a basis for this, we have developed a small simulation world together with a partner, in which investors can understand the effects of their decisions in a few minutes and understand them better.

Why should algos in the financial markets be as simple as possible?

Prof. Artinger: Complex algorithms are basically subject to the problem that they are strongly adapted to the past (overfitting). The signal-to-noise ratio is relatively poor in the markets anyway, and this overfitting makes it even more lost in the noise of the individual parameters. The reason is the high estimation errors, which in the end overlay everything and no longer allow a reliable signal to be recognised. In comparison, simple models are better and more stable, because they are based on fewer apparent patterns and ignore more noise. In addition, they are more transparent from the point of view of human decision-makers, which is also an advantage if one is thinking about an intervention. The only advantage of complex algorithms, on the other hand, seems to be that they are easier to sell (laughs). But of course there are exceptions here as well: On very short-term time scales with a lot of data or well-quantifiable correlations like high frequency trading or statistical arbitrage, complex algorithms are certainly the best choice.

But simple algorithms are not automatically easy to develop, are they?

Prof. Artinger: Not at all. A lot of expertise is needed for that. Moreover, the rules should be as simple as possible, but not simpler than that. In practice, therefore, several simple models are often used in parallel, each of which fits a specific market behaviour, while experts at the same time make sure that the necessary framework conditions remain intact.

The rules should be as simple as possible, but not simpler.

Can a human being rethink and correct wrong decisions more quickly?

Prof. Artinger: Absolutely. Some algorithms need months or even years of data to react to certain changes. A human being is much faster. For example, it was possible to intervene quite early in the Corona crash in the spring of 2020 that the general conditions had suddenly changed significantly compared to the previous market phase.

A modern alternative to classical algos are machine learning models that continuously and independently learn correlations and patterns of the past. How do you assess the chances of success of these "black boxes" on the markets?

Prof. Artinger: I think that machine learning has pretty bad cards outside of very short-term time frames. This is because the markets are fundamentally subject to a high degree of uncertainty, which by definition - unlike mere risk - is not quantifiable. However, this is exactly what machine learning attempts to do: quantify.

Machine learning has a poor chance away from very short-term time frames.
Prof. Dr. Florian Artinger

Would technical analysis, which is often criticised as useless, be an alternative?

Prof. Artinger: Technical analysis is based on heuristics, where the user at least understands what he is doing and what the decision is based on. This is not the case with machine learning, so they are two different pairs of shoes. By design and concept, machine learning algorithms are black boxes - so the user does not know what rules internally lead to the generation of a signal. Admittedly, there is now a new generation of Explainable Artificial Intelligence (XAI), in which a separate algorithm is responsible for making the decision rules transparent. But one should not expect more than a very rough understanding of the internal processes.

Investors who work visually often claim that algos cannot reliably recognise complex patterns in charts because the variants for specifying them in detail are almost unlimited - a fallacy?

Prof. Artinger: Yes and no. On the one hand, the definitions necessary for this via extensive if-then chains have long been abandoned. But if you feed a computer with a sufficient number of images or patterns pre-selected by humans, it can actually identify them with increasing accuracy through machine learning. However, there are also weaknesses that become apparent especially when new things are added. In facial recognition, for example, this could be that people suddenly wear a cap with a large white dot. Something like that can quickly render the algorithm useless. Something similar is also conceivable with charts. If a human preselects a lot of shoulder-head-shoulder formations, a computer can certainly learn to recognise them quite well - but only as long as all future patterns are at least within the range of the preselection.

In traditional systematic strategies, extensive backtests are made. Is a continuous machine-learning algo superior to this?

Prof. Artinger: The overfitting that is done today by machine-learning algorithms used to be a very similar problem in manual backtests: you built in many parameters and optimised them to a greater or lesser extent to the past. So in the end, both are bad because they are based on the same, problematic basic mechanism.

Overfitting used to be a very similar problem in manual backtests.

Should the final decision to implement trading signals lie with the human?

Prof. Artinger: It depends. On the one hand, a human decision-maker can of course introduce unwanted errors again if he has insufficient experience. On the other hand, a real expert who works systematically and according to predefined criteria can achieve real added value.

It is said that successful stock market trading is both a science and an art. What is your assessment of this?

Prof. Artinger: Away from the very short-term, there is certainly a portion of art to it.

Would an evolution of the markets away from discretionary and towards exclusively systematic decisions be conceivable?

Prof. Artinger: This scenario is of course a utopia, because there will always be people with more and others with less experience and expertise in the markets. But if I assume that there are many different algorithms that largely balance each other out, the markets would then probably be more stable than they are today. I would also assume that there would be less noise, which is mainly caused by human decision-makers. A realistic future scenario would be to understand the decision paths of experts as well as possible and to map them algorithmically via if-then relationships. If such systematics and processes are broadly defined, it would be possible to make very effective decisions.

Finally, a pragmatic question: How do you invest your own money?

Prof. Artinger: Like Harry Markowitz, I follow the 1N method and invest in various stocks and ETFs over the long term. In doing so, I look for the best possible entry points and like to gather a lot of my own valuable experience over time.

The interview was conducted by Dr. Marko Gränitz, Editor-in-Chief of the Alpha for Impact Magazine.

Would you like to use this article - in full or in part - for your purposes? Then please consider the following Creative Commons-Lizenz.