Generative AI has dominated headlines for the last year and promises to be disruptive across many industries, including asset management. Bloomberg recently hosted a panel discussion in Singapore to explore how asset managers use generative AI, the associated regulatory environment, and the challenges the current generation of the technology presents, such as attribution, transparency, scalability, and hallucinations.
Firms are exploring numerous use cases
Buyside firms are increasingly using AI for tasks such as coding, analyzing documents, and writing marketing material/email prompts with the goal of improving internal efficiencies. At the same time, experiments are underway to advance more tailored generative AI applications. Buy-side firms have been developing a variety of generative AI applications and exploring how to use it for asset management or investment research, said one panelist.
The most common and most promising use cases, thus far, are:
AI can help summarize and list critical points in a recent 10-K filing or other complex documents to help portfolio managers and analysts quickly scan through reams of information.
- Opinion mining, or sentiment analysis
Using machine learning and large language models, AI can perform text analysis on various sources — web chats, news articles, social media, and more — to produce a summary of overall market sentiment, or produce trading signals derived from unstructured text data.
Imagine being able to ask questions such as “If interest rates were raised by 25 basis points in 2018, how would it impact the economy in 2020?” Portfolio managers could explore different theories or hypotheses, like how a certain CEO candidate might change a company or industry or the case for or against a certain deal.
As a new technology, generative AI’s capabilities and limitations are still evolving. Gary Kazantsev, Bloomberg’s Global Head of Quant Technology Strategy in the Office of the CTO, noted that customers are still struggling with finding real use cases. The current generation of generative AI models, he says, are frequently undertrained and therefore suboptimal, and — because the latest models are so large and therefore expensive to perform inference on — not suitable for real-time workflows.
Attribution, hallucinations and other challenges
Though generative AI opens the possibilities to many exciting new applications, there are numerous challenges to overcome first. The possibilities are potentially groundbreaking, but panelists emphasized caution. Asset managers need to be able to see citations of any given data’s source and confabulated answers, or hallucinations, are still common.
Hallucinations can be mitigated by ensuring citations and references are in place so the source of information is traceable. But this solution is not foolproof. Sometimes, even when the answer appears to be traceable, the references themselves are hallucinations. A panelist noted that the important question is whether investment analysts can accept that.
Bloomberg’s Kazantsev added that technical questions remain about how to design architectures for a particular task and train the models with appropriate objectives that minimize obvious failure modes. Research and the state-of-the-art in this space is evolving extremely rapidly though. For example, the current generation of generative AI models are not capable of reasoning, do not have persistent memory, and are very expensive to retrain, which means they are effectively static. Retrieval augmented generation, where models have access to external knowledge bases or APIs, can alleviate some of these issues, but addressing these limitations in a more rigorous manner will take time.
Reaching peak hype?
Generative AI’s potential is exciting, but there is a risk that the excitement could be premature. The costs of developing the technology are significant, and there is a real fight for talent.
Many generative AI startups have had a retention problem. Once the initial excitement has passed, the daily active monthly users drop off. A panelist noted that many users are still determining the real ROI of investing in this new, much-discussed technology.
An example from the asset management world is that companies put significant resources toward document analysis, but the time saved may not be worth the cost. He noted that reviewing documents could take a few days, but running new documents through an AI application and then checking to ensure accuracy is both time-consuming and expensive. In this case, the benefits may outweigh the costs.
A representative from a financial regulator noted that many financial institutions are exploring the efficacy of using AI for back-end functions, especially for coding or document analysis. Developing newer use cases could be too costly for individual companies to pursue alone, so the regulator is exploring areas where the industry could collaborate for greater efficiency. The cost of developing generative AI is significant, which creates motivation to find ways to cooperate and utilize the technology in ways that benefit all.