Systematic asset manager Quoniam Asset Management is expanding its use of AI on the desk to derive signals from increasingly complex news flows and to inform trading decisions. Speaking to Trader TV, Markus Ebner, head of multi-asset and liquid alternatives at Quoniam, says the firm processes over 1.5 million news articles daily, using large language models to extract market sentiment more precisely than traditional “bag of words” methods—stretching across equity index futures, bonds, and FX. Ebner breaks down how the model functions, the challenges in identifying sentiment and forward-looking signals; he unpacks how the buy side has automated processes and the firm’s plans to extend the AI-driven strategy’s emerging market coverage.
Full Interview
Josephine Gallagher – So we’ve had you on Trader TV This Week before, but for anybody that’s not familiar with Quoniam Asset Management, tell us who you are and what is your investment strategy?
Markus Ebner – Yes, Quonium is a systematic asset manager with offices in Frankfurt and in London, and we’re managing in total about 25 billion euros, and what we’re focusing on is active portfolio management for institutional investors. So two years ago you had me on your show, and we were talking about our GDS strategy, Global Data Sentiment. And this is one of our new strategies at Quoniam, and we’re having a life track record of five years now, sharp above one, and very low correlation figures to equity indices to fixed income indices and also to hedge fund indices. And what we’re doing in this strategy on a daily basis, we collect more than 1.5 million news articles, filter them by those topics for which our analysis have shown that they’re really driving the markets, then we’re deriving the sentiment out of these articles, and based on this sentiment, we’re going long or short in equity index futures, fixed income futures, or FX forwards.
Josephine Gallagher – So, as you said, when we had you on last in 2024, we discussed how your firm is extracting positive and negative signals from news articles and large unstructured data sets, so in the last two years, have you had to develop those models further?
Markus Ebner – Yes, two years ago we were using the so-called “bag of words” approach for extracting the sentiment out of an article. So the bag of words approach in simple words just counts positive connotated words, negative connotated words, and builds a score out of it, and this is a very robust approach, especially if you build averages over hundreds or thousands of news articles. However, this score is not very precise. So think about like quarter earnings announcements of a company, and the companies may be coming from a very poor past, and the earnings announcement was like oka-ish however, the outlook is very good. Then applying the bag of words approach comes to the result that most of all of the article is more negative because they’re coming from negative past, the current earnings are okay-ish, however, it ignores totally that the outlook is very positive. So this is a disadvantage by just counting words of an entire article. So large language models are much more powerful because large language models, you can tell, okay, what was in the past, what was published today, this is already reflected in the stock price. Focus on the outlook, and this is very positive. And tell me, how the sentiment of this outlook is. And so, because of this, because of having much more precise measurement of sentiment, we enriched our strategy also by AI, by large language model-based signals.
Josephine Gallagher – Okay, now in today’s very news-heavy environment, there seems to be a bit of a disconnect between what’s actually like a scary headline to what’s actually been reflected in the market. So, how do your large language models precisely measure or identify market sentiment in these kinds of scenarios?
Markus Ebner – This is definitely true. We see lots of mixed signals from news articles and from the markets. On one side, there is still the US-Iran conflict, Strait of Hormuz is still closed, oil is still above 100 per barrel, and also we see already the inflation rates rising, and hand in hand with the interest rates. So this is quite challenging for the markets. The other side of the coin is that we see a very positive narrative for AI, which is still ongoing. The narrative that AI will fundamentally change the entire economy globally, and this narrative even got momentum with companies like Nvidia or Entropic exceeding expectations,. And in our GDS strategy, what we’re doing, we are collecting those articles, putting into major themes like company-related news, like political or macroeconomic related news, and then we are deriving the sentiment, and what we see at the moment, is that on one side company-related news are very positive, and on the other side, the macroeconomic and political news have a sentiment which is at least optimistic. It’s not positive for the current situation, but it’s very optimistic for the future, and combining this leads us to a position where we’re still having a significant long exposure in equities, currently about 100%, and we have also seen a very interesting dynamic during the year. The beginning of the year was quite positive; we got started with a long position, then the sentiment shifted strong negative, however, just for a short period in time, and then optimism came back, the sentiment picked up, and this was also reflected in stock prices.
Josephine Gallagher – Understood. How much of this process has actually been automated from the investment stage right through to the trade execution, and what asset types are covered?
Markus Ebner – So our investment universe are derivatives, they are equity index futures, fixed income futures, FX forwards. We’re talking in total about more than 30 assets, and it is derivative-based, because we also want to go active-short when we see negative sentiment for a specific company or specific asset class. And yes, indeed, automation is the key. What we are doing at 8 am central european time, we collect news articles from the last 24 hours, process them until 8.30 then we receive the signals, we double check the signals, we double check if the data load is correct and then we derive the trades out of those sentiment signals and pushing them to our trading desk, so we are ready to trade before the physical markets opening at 9 am and normally we are executed before 10 am. And yeah, velocity is a key, because the half-life of our signal is just a day, so we have to be really timely at the market.
Josephine Gallagher – Now, looking ahead, like, how do you plan to actually build out the data analysis part and these data models even further?
Markus Ebner – So, at the moment, we’re running two research projects. The first is about to enlarge our investment universe, also to non-developed market countries. This is quite challenging, because we have to translate from many languages, articles into English, and then we derive the sentiment out of those articles. What we’ve seen is that, especially from non-developed market currencies, the local press has real value in it, so this really reflects the sentiment for non-developed market countries. This is one research project, and the other research project is a more general one, because AI is coming up more and more. And we try to add also AI agents to our research team in a way that we learn the AI, how our models are functioning, to get at some point in time in the situation that AI really can support our research process, and maybe also hints where there is room for improvement for one or another strategy.
