Today AI is often associated with chatbots and applications built on large language models (LLMs). However, artificial intelligence has long encompassed other approaches: machine learning (ML) and deep learning (DL) are established techniques used for forecasting, classification and many other tasks in finance.
Asset management, trading and market-making — and particularly hedge funds — have been heavy users of these techniques. Hedge funds tend to adopt new methods quickly because even a small edge matters; by edge we mean any procedure that systematically extracts profit opportunities from movements in financial and capital markets.
Why LLMs are not a universal solution
Although asset managers work with very large datasets, that alone does not make the domain ideally suited to LLMs. Two interrelated issues are prominent:
- Hallucinations: LLMs sometimes produce false or fabricated information.
- Context rot: even with a large context window, a model can lose track of earlier context over time and fill gaps with hallucinated content as scale increases.
These behaviors make reliable information extraction difficult and mean that simply providing more data does not guarantee better outputs. This is an inherent characteristic of how many LLMs operate, and control routines can reduce but not fully eliminate these tendencies.
Practical uses of chatbots and LLMs in asset management
LLMs are not without value. With careful prompting and oversight, chatbots can be effective for certain tasks. Concretely in asset management:
- ML and DL models are likely to remain the primary tools for systematic forecasting.
- LLMs can speed up the development of other models, help surface bugs and inconsistencies, and thus indirectly support investment decisions.
- They can assist learning and research, but domain expertise is still required to spot LLM hallucinations and verify outputs.
A key factor: data hunger and feature engineering
All ML, DL and LLM models are data-hungry. Training often involves creating millions or billions of explanatory variables (features). If these features are poorly designed or irrelevant, model performance will suffer. This applies to LLMs as well, with the additional complication that we frequently lack full visibility into the data and processes used to train large proprietary models.
The article uses a simple baking metaphor: features are the ingredients. Flour, eggs and sugar can make a cake; but if the ingredients list included face cream, toothpaste and garlic, one would rightly doubt the result. Similarly, poor feature sets do not produce reliable predictions.
Conclusion
LLMs and chatbots can play a useful supporting role in asset management: accelerating development, aiding data validation and assisting decision-making. However, because of hallucinations, context rot and opaque training data, they do not replace well-established ML and DL models. Careful control, robust feature engineering and domain expertise remain essential when deploying LLMs in finance.
