FOR IMMEDIATE RELEASE
AI for the Buy-Side: Preparing for the Next Wave
Mike Desanti, President, LightPoint Financial Technology
According to a 2022 survey by Greenwich Associates, 54% of capital markets firms are already using AI to assist in trading and managing securities, and 64% have made increasing automation through AI and Robotic Process Automation (RPA) one of their key longer-term goals. Most of the work to date has been done using rule-based systems, which have been used in capital markets space for about 30 years in trading and risk management, or in the development of neural networks, which have been used for the last ten years or so mostly to develop adaptive algo trading strategies and to mine data.
The new kids on the AI block are the natural language-based transformer systems and the one that is getting the most attention in this space is ChatGPT from OpenAI. The GPT acronym stands for generative pre-trained transformer. Transformer architectures are different then earlier recurrent neural networks in that they can process their entire input all at once in parallel and this reduces training times and costs. Transformer models are what has enabled these types of systems to evolve as fast as they have.
ChatGPT 4, release in May 2023, goes a long way towards giving us humans the capability to have a conversation with our computer where we can ask it write text or code for us, mine data and perform data transformations, as well as perform and/or explain various types of math. The technology is the fastest growing consumer application to date with over 100 million users. It will be incorporated into products that we use every day such as our Internet search engines, email, office productivity tools like Microsoft 365. Expect competitors like Google Bard and other products that leverage the core ChatGPT engine like Jasper.ai.
This article examines some specific ways that hedge funds, asset managers and mutual funds on the buy side, can leverage this technology, as well as what limitations they may face as they embark on their GPT journey.
On the buy side one of the main ways that generative AI will make it into their firms will be in the form of using it as a part of another vendors SaaS offering. BloombergGPT will probably become a dominant force in this space because they have built a 50 billion parameter large language (LLM) model from scratch specifically for financial firms. This proprietary model will be offered with various Bloomberg products. This type of approach requires a significant investment, in addition to a large amount of financial data that is used to train the model.
Broadridge Financial’s LTX subsidiary recently announced the BondGPT which extends the 175 billion parameter model that comes with ChatGPT to include corporate bond information. It is designed to answer bond-related questions and assists users in their identification of corporate bonds on the LTX platform. It also incorporates real-time liquidity information from the LTX Liquidity Cloud®. Its goal is to simplify the bond selection process and to help streamline the workflows associated with bond acquisition. This approach uses a private instance of ChatGPT and this will likely be the architecture that many SaaS firms will use in order to offer ChatGPT to their users.
The main advantage that Bloomberg and Broadridge have is that they have access to large amounts of quality data and training resources. This is very important because these systems will hallucinate or fill in the blanks in order to compensate for data ambiguity or lack of training. Good data quality and sufficient training are essential.
Programmers at many firms are already starting to use ChatGPT to generate code. The main Uses Cases currently involve eliminating repetitive programming tasks that are associated with cleaning and processing data for relational data warehouses and data lakes. Expect this trend to build steam as firms debut their own competitive open source LLM’s like Databricks Dolly. The main thing to remember is that the developer needs to be able to describe what they want ChatGPT to do. If the prompts are vague the results will be disappointing.
For buy side firms that want to create their own LLM extensions with their firm’s data there is not a lot of environments that are readily available yet. For example, firms have to apply to get access to ChatGPT4 running in Azur’s OpenAI environment as Microsoft works through capacity and risk issues.
Even though the tooling is not yet widely available to create your own LLM extensions, firms should be looking at the state of their data warehouse environment. Is the data clean and in a central place? Do I have a sufficient amount of quality data to support training my own model? They should also be starting to train their users on how to write good prompts which is essential in order to get quality output from a GPT system.
Business Intelligence tools and dashboard technology will also likely play a key role in the future of this technology. Once a user requests a specific set of data using a natural language prompt they may want to render it graphically or put it into an environment where the user can analyze the data themselves.
While there are issues to be addressed like hallucination and data security issues. The development of GPT technology represents a significant step forward in bridging the man machine divide.