In this interview, Stefan speaks with Béatrice Guez, co-founder of AI for Alpha, a company that uses artificial intelligence to help investment professionals understand financial markets and make better investment decisions. Some of the key points discussed include:

‍

• The use of decoding strategies based on graphical models to identify investment drivers and replicate successful strategies
• The application of AI to classify funds in terms of ESG stance and stress risk analysis
• The importance of machine learning and reinforcement learning to adapt to new market environments and optimize asset allocation

‍

If you're interested in learning more about the intersection of artificial intelligence and finance, tune in to this insightful interview with Béatrice Guez.

‍

‍Tune in now and get the latest insights on structured products, technology and investment management:

Schedule a call with us for additional questions.

‍

No time to listen? We summarized the podcast for you:

‍

Stefan: I'm here with Béatrice Guez, she is the co-founder of AI for Alpha. Thank you very much for taking the time today. What is AI for Alpha? I presume AI stands for artificial intelligence. But anything more you can elaborate on?

You're right, AI stands for artificial intelligence. We help investment professionals better understand the main drivers of financial markets and help them in their investment decision. We help our clients in the asset allocation decision, which is based on market signals. We provide some analysis on their fund investments, and help them to identify macro-risk sensitivity factors, buyers or any other risk exposure. And last, we develop the tailor-made investment strategy based on our signals. And, in particular, we launched at the start of the year, AI CTA decoding strategies with 29% performance year to date.

‍

Wow, that's impressive. You use the word decoding strategy. Can you tell us a little bit more about what that means funding decoding solution and how it adds value?

It is based on a path of machine learning, which is called graphical models. What graphical models enable to do is to detect the main factor of a specific strategy based purely on its historical NAV. The technology itself and its advantage is that when we analyze or help our clients analyze the funds or the portfolio of funds, we do not need to go into the transparency of each fund investment, but based on the NAV, we are able to identify the main investment driver to explain the performance and to replicate the strategy. When I say investment factors, this means: Which asset class? What is the geographical exposure? Is there a style bias in the investment strategy? What are the main sector investments? And finally, how to categorize funds in terms of ESG stance?

‍

ESG is one of the latest things at the moment. Some people are also a bit cynical about ESG calling it greenwashing. Is that something that you address?

This is definitely one of the applications. When we use or decoding solution, we can spot the type of ESG investments. Is the investment more toward carbon reduction? Or is the investment more about social responsibility or environment? And to do that, we use some benchmarks. We use usual benchmarks such as MSCI or stock benchmarks and we can identify what the benchmarks are that are closest to the fund identification. When we speak about greenwashing, we are not labelling a fund as greenwashing. But we can rank funds in terms of ESG intensity and we can identify funds that do not map with standard ESG benchmarks. For us fund selectors, being able to provide our clients with this fund classification, is a way to guide our clients in their fund analysis. A fund that does not match with any benchmark does not necessarily do greenwashing, but has an investment style, which is a bit different than the benchmark. And it can be interesting for the fund selector to dig a bit more. That is one of the applications of the decoding regarding ESG. And we do it as well for stress risk analysis: What is the interest rate exposure of the portfolio of funds that you have constructed? Which is the exposure sensitivity to inflation or the sensitivity to the US dollar? Usually, the asset manager has quite an idea of what the exposure is, but it's not so easy to quantify it at the fund level. And it is even more difficult to quantify it at the portfolio level. That is where the decoding tool is a real help for any fund manager, who is doing portfolio construction - to be able to have a global top-down view of his/her investments.

‍

How did you actually come about and who are the people involved in it? What's the background?

My background is investment banking, so I have worked for more than 20 years in different investment banks such as JP Morgan, Deutsche Bank and Société Générale, where I had different financial engineering responsibilities. And I decided three years ago to leave the corporate world of big banks to participate in the transformation of the asset management industry, notably through new technology, and in particular, artificial intelligence. The other co-founder of AI for Alpha is Eric Benhamou. Eric used to be Head of Quant with different investment banks. More than five years ago he decided to specialize in artificial intelligence. So from going from pure quantitative finance to artificial intelligence, he is completely convinced that quantitative finance will need some more tools to be able to be more efficient to analyse more data, and that artificial intelligence will be more and more necessary in the industry, and especially asset management industry. Jean-Jacques Ohana joined us a bit more than one year ago, first to be a scientific adviser and then to develop the new product business.

‍

Given that we have increasingly more data and fixed rules don't really work anymore: What is your vision how you address this with AI for Alpha?

First, there is the learning part, which is crucial. The idea of machine learning is to be able to digest huge amount of data and to learn from historical information and from this learning to be able to adapt to the new environment. Opposed to purely quantitative strategies, that are based on a fixed rule, machine learning can do continuous retraining and adapt to new environments, and it might change its decision based on new information. Little by little the model is learning on switching the allocation based on new market conditions. For example, one year ago, one of our models spotted the peak of inflation and that was one of the reasons of the shift from a bull to a bear market. That is just one of example out of 20 that we can quote.

‍

You already mentioned deep reinforced learning. Could you please explain it a bit more in detail?

It is quite simple, actually. Supervised learning is called supervised because we guide the model, we know the answer. For example, quite intuitively for image detection, we give an image of cat and dog and the machine little by little understands how to distinguish what a cat and what a dog is, because it had been supervised bythe right or wrong answer. So out of sample, it can categorize properly cats and dogs eventually. We use supervised learning for crisis detection, for example. And just the same way a machine learns from the past what the drivers of crisis were and given a sample it can learn and adapt to the environment and therefore anticipate crisis using supervised learning. That is the most used AI technique. There is also unsupervised learning, where we do not know the right or wrong answer. Here we provide a lot of information to the machine, and the machine itself detects some patterns. Usually, it is used to do some classifications. For example, you might have seen some dendrograms. This is typically unsupervised learning. The last type of machine learning that is used in games is reinforced learning, which is something between supervised and unsupervised learning. For example, in a game we do not know the right way of playing, but we give the machine an objective, which we call a reward, and some rules of play. And by iteration, the machine will adapt to the new environment and know how to play to optimize to achieve the reward. Obviously, it has been used by Google DeepMind to beat the best Go game player. In our case we use this type of ML to find the best allocation depending on the market environment. Our allocation model uses reinforcement learning to be able to adapt to the crisis environment and to see what the best allocation is based on that.

‍

When you say allocation, that is asset allocation?

That is correct.

‍

What does it detect in detail - Does it detect market regimes? Is it trying to time the market?

Yes, exactly. What we have developed is a model that detects market regime. Why did we choose to focus on market regime and not try to anticipate the price return or profitability, for example? The reason is that we realized that financial markets are very noisy. Trying to anticipate the short-term or even long-term return would be a very difficult exercise. On the opposite we see that there are patterns in crisis. We ask the machine to detect the similarities between these patterns and to anticipate market regime based on new market conditions. This is what we use in the asset allocation model. We use more than 150 features in order to help the machine find what the most predictive features are. That is the first step, and it is very important to see what the most predictive market information is.

‍

Would one such feature be the implied predicted outcomes from the option market that become more fat-tailed, for example?

We rather use raw data like price indicators, meaning the evolution of price not only on market indices, but also on commodities. Or for example, data on parallel markets that can have a big influence on the market in general. We also use technical indicators such as put-call-ratio or market brace that can give a trend in the market. But we also use macro factors such as economic surprise, inflation expectations or interest rates evolution. We obviously have some risk perception indicators that we have developed based on risk premia on each asset class in order to have a sense of the financial liquidity in each market. In general, we leverage between 100 and 150 features, and then during the supervised learning task, the machine will detect at each point of time the most predictive features. That is a way for the manager to learn what the model expectation behind the allocation is and what the probability of crisis is, but also to understand what the main market drivers behind the model results are.

‍

How do you check that the judgment of your model was correct? And how is there proof of success that what you brought to them actually added value to them versus if they wouldn't have used your model?

Whenever we do an allocation for clients, we have a benchmark, and we follow that allocation versus the benchmark. In one example, we would have a basket of the 10 best CTA hedge funds. We managed to have on the life performance a very high correlation with this basket with an over-performance of these funds. Why over-performance? Because the indices do not charge the same fees as a hedge fund on CTA providers. Today the strategy has almost 30% performance year to date, which is very high compared to the main CTA asset managers.

‍

How do you deal with inflation, in particular with inflation exposure in the asset allocation?

We did a conference on this subject a few weeks ago. Not only on inflation. We are in stagflation environment. Our view today is that central banks do not longer have a grip on financial markets, because they need to cope with the inflation issue. We see one of the consequences of that macro-trends are taking over financial markets: In our view and something that we also see in the model is that two types of strategy that make sense in the current environment is first, commodities and second, following strategies. This is why we launched AICT index vicia which received very high interest from investors and investment banks.

‍