Me, my job, and artificial intelligence

Artificial intelligence has become an indispensable part of my day-to-day work. A report on my experience.

In customer service and among programmers, artificial intelligence (AI) is already the norm. In the day-to-day work of a financial analyst, however, the use of artificial intelligence is still rather a vision than actual reality in many places.

As a research analyst, I am constantly on the lookout for ways to use AI to make my day-to-day work more efficient and increase the quality of my analyses. In the process, I quickly learned the lesson: Finding meaningful fields of application is easier said than done.

Many investors dream of working with an AI that can predict the long-term development of stocks and markets fully automatically and accurately. Many start-ups are working relentlessly to make this dream come true. The range of AI tools is large, and I regularly receive emails with more and more appealing AI tools.

However, one quickly awakens from this dream when one asks oneself the question: Why is this product being offered for sale when the inventor could better use it himself and exclusively to skim off the return? The answer is usually obvious. The dream does not come true that easily. As an analyst, I must take a more pragmatic approach to find processes where AI tools can be of help.

When someone asks me what my day-to-day work as an analyst looks like, I like to answer: "Reading, thinking, writing". Put a little more precisely, this means:

• Gathering data
• Formulating theses and providing evidence
• Communicating results

Generally, the steps are not clearly separated from one another, are carried out in mixed time sequences and, depending on the subject area, one is occupied with one or the other step for a longer period of time. Nevertheless, the three-step division is ideal for identifying use cases for the application of AI.

Gathering data

AI applications built on language models such as ChatGPT, BingChat, Bard, or HuggingChat are promising candidates to provide assistance in gathering information. Anyone who has seriously studied language models, also called chatbots, knows that prompt engineering (getting the input right) is a challenge. In addition to your own experiences, you can find various help online with the help of which you can learn to make your input more precise. It is not uncommon that the reason for an inappropriate answer lies in the imprecise input.

Popular theories, economic relationships, and information on companies can be gathered more quickly and comprehensibly with the help of chatbots than by approaching the question with the help of a search engine. However, if you want to go into more depth, the information quickly becomes scarce. If you ask questions about topics that are outside the realm of general knowledge, e.g., details about medium-sized companies, the chatbots quickly start hallucinating. They come up with factually incorrect, but plausible-sounding statements.

Chatbots that access an information database and do not primarily evoke the information from their own memory offer a significant improvement in quality. Such an information database can be created, for example, by the chatbot performing an Internet search. The results of the Internet search flow into a temporary information database, which the language model draws upon to formulate a response. With the combination of language model and information database, hallucinations occur much less frequently.

For many of my analyses, the relevant information can be found in publications by companies and institutions. It is contained in texts and tables that are often neither available in a standardized form nor can be accessed via databases. This is where chatbots come in handy, allowing one to search through documents. The bot is then supposed to formulate its answer primarily based on the information in the document rather than on its memory.

In some cases, such queries work very well, but often, unfortunately, they don't work at all. Whereas information can still be read precisely from short PDFs, annual reports can quickly comprise more than 200 pages. Such documents exceed the input capacities of the language models. Instead of a precise answer, one receives vague statements or is presented with hallucinations.

Hallucinations are a serious problem. Nevertheless, the use of language-model-based AI should not be rejected outright. After all, a similar problem arises when a newly hired research assistant is tasked with gathering information. Neither do human assistants have unlimited capacities, nor does information intake and processing always proceed without error. Therefore, one has to decide whether to accept the found information unchecked or verify it. After some time, however, one gets a good impression of the quality of the work of the new assistant and can - in the best case - accept the information after a plausibility check. If, on the other hand, one spends more time checking the information than researching it independently, neither chatbot nor assistant offer any added value.

Formulating theses and providing evidence

When it comes to formulating theses based on the information collected, AI systems can only offer limited added value in my experience. In the case of chatbots, this is not surprising when you look at the way the language models work. The bots do not have a body of knowledge but put together words that best fit together given the input.

However, the language models can be used excellently to put one's own theses to the test. You enter your own thesis and then ask what arguments contradict your thesis and what alternative explanations exist. The answers of the bots often anticipate what readers might later criticize. So you can address these points yourself in the post.

When analyzing data, other AI systems, such as machine learning algorithms, can help identify correlations. Artificial intelligence can indeed be used here, but the data sets that I collect are usually not so extensive that the effort of using complex algorithms would be worthwhile, let alone that the algorithm learns from the data to independently develop theses. Statistical methods, starting with averaging and ending with random forest estimation, are sufficient for my evaluations.

Communicating results

Chatbots are true linguistic artists. In communication, they can be used excellently for efficiency and quality gains. This starts with translations, for example. The e-mail in English is quickly formulated with the help of a chatbot. Even the translation of a long article, which would take many hours without tools, can be done efficiently within an hour with the help of AI-based translation software like DeepL. The resulting text is free of grammar and typing errors. Only phrasing needs to be checked and, if necessary, technical terms need to be replaced with the correct words.

The language models are also well suited for creating summaries and headlines. However, one should not expect the suggestions of the chatbots to meet one's own requirements and ideas. They serve more as inspiration.

The suggestions of a language model can be used to predict what a reader will pick up from my articles. If a bot reproduces content in its summary that is not central to the contribution, one must ask oneself whether one has given enough space to the main arguments.

Language models are also suitable for searching for alternatives for your own formulations. Grammar and typing errors can be quickly identified. However, experience has shown that they regularly miss typos, just like the human eye.

If you want a chatbot to formulate entire texts from scratch, you face the problem that chatbots need extensive information to write a text with a given hypothesis. If one additionally wants one's own found connections to be used as evidence, the input becomes extensive and almost resembles one's own text. Keeping the input short means that the resulting text is generic and, in my view, has little or no added value for a research institute.

Conclusion

Artificial intelligence can facilitate the work of analysts in the financial industry. Particularly in information gathering and communication, language models can make processes more efficient and increase the quality of work.

I do not perceive artificial intelligence as a threat to the job of a financial analyst. Rather, I see AI systems as a creative, much-knowing, but also fallible assistant. I am excited about the possibilities that technology will offer me in the future.

Legal notice

The information contained and opinions expressed in this document reflect the views of the author at the time of publication and are subject to change without prior notice. Forward-looking statements reflect the judgement and future expectations of the author. The opinions and expectations found in this document may differ from estimations found in other documents of Flossbach von Storch SE. The above information is provided for informational purposes only and without any obligation, whether contractual or otherwise. This document does not constitute an offer to sell, purchase or subscribe to securities or other assets. The information and estimates contained herein do not constitute investment advice or any other form of recommendation. All information has been compiled with care. However, no guarantee is given as to the accuracy and completeness of information and no liability is accepted. Past performance is not a reliable indicator of future performance. All authorial rights and other rights, titles and claims (including copyrights, brands, patents, intellectual property rights and other rights) to, for and from all the information in this publication are subject, without restriction, to the applicable provisions and property rights of the registered owners. You do not acquire any rights to the contents. Copyright for contents created and published by Flossbach von Storch SE remains solely with Flossbach von Storch SE. Such content may not be reproduced or used in full or in part without the written approval of Flossbach von Storch SE.

Reprinting or making the content publicly available – in particular by including it in third-party websites – together with reproduction on data storage devices of any kind requires the prior written consent of Flossbach von Storch SE.

© 2024 Flossbach von Storch. All rights reserved.