AI Trading: What Could Possibly Go Wrong...

The popular press is abuzz with articles about AI's transformation of investing, financial research and analysis, and stock market forecasting. At MarketPsych we work in the field of AI with large language models (LLMs) including chatGPT, and in today’s newsletter we explain some LLM trading basics as well as our own successful AI agent designed for stock price forecasting. To illustrate the value of AI in investing, we'll start with the human weaknesses it is designed to exploit…

Investing on fear

“Fear is driving the world like never before…. Humans act in very predictable ways when they're frightened.” ~ The Fear Index (2022) Trailer, Sky TV series Robert Harris’ book The Fear Index (2011) - remade as a TV miniseries - describes the development of a financial trading AI that profits from human emotion and pursues its sole objective function - profit - without compassion or any regard to human consequences. [SPOILER ALERT] The AI is adaptive and a fast learner. First it learns to write news to move market prices. Then it learns to cause events from which it can profit (the book being in the “airport thriller” genre, it caused an airplane crash to profit from shorts on the airline firm’s stock). Eventually it engineers the murder of its creator to prevent itself from being shut down. Given the emerging capabilities of chatGPT and other LLMs, fear about AI’s turning against their well-meaning human masters is an increasingly common theme in the media. But the reality is much less dramatic because current AI is not self-aware AGI.

AI vs AGI

chatGPT is a non-sentient model (it's not conscious), and there is a big gap between it and the type of artificial general intelligence (AGI) described in the Fear Index. As physicist David Deutsch noted: “AI [artificial intelligence] has nothing to do with AGI [artificial general intelligence]. It’s a completely different technology, and it is in many ways the opposite of AGI … An AGI can do anything, whereas an AI can only do the narrow thing that it’s supposed to do.”

By training and learning on the large corpuses of internet text, LLMs absorb and learn patterns in human language and communication – a super-fast binary analog of the biology of our toddler-selves as we learn language and how to think. Overviews for learning more about the technical achievements of AI and LLMs (the software foundation of chatGPT) are available in these expert literature reviews and in Google’s education series on the topic. Perhaps because it is not self-aware, AI still requires human operators and conductors with domain expertise (that is, experienced quants) to operate it. And when there are not experienced operators, it often doesn't go well…

chatGPT is not a stock market guru

LLMs learn patterns in language, understand context, and predict future words and phrases. That's all. Such limited functionality hasn't deterred several start-ups and at least one academic paper from jumping on the AI Trading bandwagon.  But without experienced operators, such models rarely work in the real-world out-of-the-box. See the failure notice on the webpage of one such model below, screenshot in June 2023. 

Using chatGPT for stock price forecasting is good marketing, but it’s not good business (unless your business is selling subscriptions).

However there is a swathe of quant finance where AI models are used productively. For example, LLMs are excellent at finding patterns in high frequency order flow data. But where data is more sparse, or where conditions are changing rapidly, LLMs are unable to see over the horizon. In such cases human domain experts must guide the model to extract the best predictive features from large datasets. 

Our AI trading bot

At MarketPsych we launched an AI-based media-based stock predictive model in Jan 2020 called the StarMine MarketPsych Media Sentiment (MMS) model via Refinitiv. The model uses features derived exclusively from financial media sentiments and themes, themselves extracted by trained AI models. 

Our MMS model was approved for production in Aug 2019 (red dotted line in the image below), and it has been sold commercially since Jan 2020. The model ranks stocks daily according to their expected relative performance for the following 30 days. The decile spread return is plotted below, and it is generated by hypothetically going long the top decile of stocks and shorting the bottom decile, so it's an absolute return model. No transaction costs are included in this monthly-updated model. The equity curve (return) is plotted below as the blue line.

We have a more detailed white paper on the model as well as lots of background research, and we encourage institutional investors to reach out for more information.

Bottom line, there are predictive patterns embedded in the emotions and themes expressed about companies in news and social media. While optimized LLMs (and certainly AGI) may be able to find them, we believe human domain experts will remain in the lead in the near-term. Linked here is a brief promo video on our MMS model.

The AI advantage

“What disturbs and alarms man are not the things, but his opinions and fancies about the things.”
~ Epictetus

While we should not overreact to the emergence of powerful AI tools (no, they won’t steal your job, at least not yet), we should keep track of how they exploit human “opinions and fancies”. As the Fear Index prophesied, AI may find the most profit (alpha) by arbitraging the gap between human opinions and reality. In social media, video gaming, and some governments’ social control efforts, we already see AI deployed for this purpose. Consumer, financial, and political manipulation has already reached problematic levels with today's more sophisticated AI.

To help our clients understand market opinions, we launched a new website with (limited free) sentiment analytics on our Data App.

Behind the login are screeners, charting tools, and research covering 100s of themes and emotions and 100,000+ global assets. Please contact us if you'd like more information about sentiment or our AI model.

Happy Investing!
Richard and the MarketPsych Team