Can investors really use market sentiment, rather than financial data, to decide which stocks to buy or sell?
Yes, definitely. If people are angry with a company, for exam ple, you can expect its share price to fall in the short term but then rise later on. That’s often what happens when the company’s management has done something morally wrong. Shareholders tend to overreact based on their emotions and sell their shares at a price that is not economically rational. So crisis periods often represent an opportunity. A few years ago, we set up a strategy that consisted in buying shares in the 10% of companies that caused the most anger on social media. These companies tended to outperform the market the following year. It was a sort of anti-ESG strategy... We bought shares in companies that everyone hated. But the fact is that these companies often improve after that and sometimes even become leaders in sustainable development and governance.
In terms of ESG, we will soon come out with a predictive index based exclusively on these criteria. Take a look at Apple stock (ed. note: he’s sharing his screen). You can see how many people are talking about Apple on social media in relation to ESG topics. You can watch the activity minute by minute. This activity and the crowd’s opinions can guide future share prices.
You believe that the sentiment expressed in the news and on social media influences share prices. But isn’t it actually the other way round?
It works both ways. If a share price suddenly falls by 10%, the media will produce a narrative to explain why. If enough people believe it and negative sentiment emerges, this will end up generating a future decrease in the price. It’s a self-perpetuating scenario. But sometimes the sentiment expressed precedes the market movement. Investors are always telling themselves a story. For instance, they tell themselves that if interest rates rise, share prices will fall. This pessimistic scenario is picked up in the news and on social media, and it winds up influencing prices.
Social media, in particular, acts as an accelerator. There have been some salient examples of this recently. We recall when the share price of video game retailer GameStop surged following all the buzz on social media.
The cryptocurrency market is notoriously volatile. Is it more sensitive to market sentiment than equities?
Yes, that’s what we’ve seen. In the crypto sector you can very frequently see sentiment emerge before a price movement occurs. For example, before the LUNA token fell last autumn, we could clearly see sentiment moving in a negative direction, even before the last spike upwards. For each crypto, we track a huge number of parameters based on keywords, such as adoption, developer sentiment, code updates, transaction speed, FOMO (ed. note: fear of missing out) or potential cyberattacks. Generally speaking, the top three cryptocurrencies tend to outperform the market.
In practical terms, how do you go about analysing market sentiment?
We use a lot of open-source natural language processing tools, such as spaCy or more recently Meta’s Llama 2, but we adapt them to finance. Especially on social media, messages are very short, so the tools have to take into account the economic context.
I can show you some of the indices we offer (ed. note: he shares his screen). Our customers have access to this platform. In this window, for example, we extract the information every time social media or news outlets take an interest in a company. The large language models (LLM) collect all the comments and news from the last hour and summarise it in a very short text of a few lines for each firm. This basically reflects the pulse level, i.e., the extent to which the market’s attention has been stimulated about a given company, and the general feeling about it. Users can set their preferences based on several criteria, for example by selecting a specific country. Take Swiss companies listed on the stock exchange: here we can see the sentiment expressed about each of them almost in real time.
"If I can make 10% consistently with low volatility, that's great"
So does it work?
As various studies have shown, our data is mostly predictive for currencies, bonds, commodities and cryptocurrencies. We provide detailed global sentiment indicators on these areas.
In 2020, we also launched a product designed to predict share prices. Most companies fail in this area after launching their product, as it is difficult to develop signals that remain consistently predictive over a long period of time. However, since its launch, our model has remained useful in predicting price changes in US and Japanese stocks over the next 30 to 90 days.
However, we do not claim to be able to predict price movements systematically. There is currently no academic consensus in this area. Several academic studies suggest that there is no proven predictive power for prices based on market sentiment data.
The important thing when trying to predict stock prices is to do it well. You have to let go of the dream of getting annual returns of 30% or 40%, and just say to yourself, okay, if I can make 20%, that’s great. If I can even make 10% consistently with low volatility, that’s great too.
Do your solutions work best at predicting the market over the short, medium or long term?
Different approaches are possible. Data can be collated and updated every minute, every day or monthly. It all depends on what the client needs, whether they are traders, pension funds or insurance companies. Our next product, due out in January, will analyse data within 140 milliseconds.
It’s worth pointing out that the shorter the averages are, the more extreme the peaks and troughs will be. If you only look at the daily change in sentiment, it can be difficult to understand what’s going on, but when you take an average over longer periods, for example one to three months, the picture becomes much clearer.
A study published last April showed the accuracy of ChatGPT-4 in stock picking based on market sentiment. To what extent is this tool a "game-changer" and a competitor for you?
These tools are game-changers because they have a chat interface. Investors can therefore interact with them and ask very specific questions, enabling them to make better use of the data. They also have an increasingly detailed understanding of the context in which a text is written, and can then interpret the meaning of a word much more accurately.
However, as far as ChatGPT goes, it is not specifically designed to have predictive ability. It uses historical data and will therefore be inclined to produce an estimate based on that data, which often results in over-reliance on the past. We believe there are other algorithms out there that are better at building long-term models.
How do you see the future of sentiment analysis in the markets, given the rapid evolution of technology?
The tools for extracting useful information from text are advancing quickly. We can now use large language models (LLMs) to extract deeply embedded signals from 100-page regulatory reports and large volumes of seemingly meaningless comments on social media. But these GPU-based technologies are still expensive due to the computational power required. We expect costs to come down over time and capacity to increase. So the needle in the haystack will be easier to find.