Research Partnership with the Institute of Financial Services IFZ Zug: Chat GPT for Financial Advice.

(The Combination of Large Language Models and Rule-Based Systems)

Large language models (LLMs) like OpenAI’s GPTs or Google’s Bard belong to the latest developments in natural language processing and are currently used and explored in a wide range of applications, from chatbots to machine translation. In simple terms, LLMs are computer programs that are able to generate text word by word based on what they have learned from large data sets. The program predicts the most likely next word based on the previous words in the conversation. Such LLMs are also increasingly prevalent in the financial sector and are likely to influence it even more in the longer term. However, the compatibility between such solutions and certain financial services requirements is not fundamentally guaranteed.

A relevant example from the Swiss financial industry is the investment advice, where traceable and explainable recommendations are required, which, however, may conflict with the probabilistic character of LLMs.

The objective of our condensed study is to explore how LLMs and rule-based systems can be combined for investment advice. Furthermore, the study aims to create a prototype that showcases the benefits of integrating these two concepts, without, however, meeting all regulatory requirements (e.g., from the Banking Act or the Federal Data Protection Act) for AI-based investment advice. Hence, the focus lies on demonstrating the technical feasibility of using LLMs to support deterministic investment recommendations, rather than discussing potential obstacles or non-technical enablers.

The solution obtained shows that the combination of probabilistic LLMs and rule-based systems, while retaining the advantages of both approaches, is possible and that the prototype works well in most cases.

The findings from this study and the prototype can be used by the financial sector as a starting point for discussion and the development of more sophisticated solutions.

Condensed Study:

A Python implementation of the prototype and more technical details can be found in the following public GitHub repository: