I had the opportunity to discuss with my friend Derek some of the implications of AI in finance on his podcast, What I Think.

Derek invited me on his podcast to discuss how artificial intelligence is reshaping quantitative finance. We covered a lot of ground in about 35 minutes, so here's a condensed roadmap of what we explored.

Foundations

There are some building blocks behind all statistical learning systems, which include classical machine learning techniques and modern-day LLMs. As a rule of thumb, older models are easier to use, faster, and are more interpretable. Robustness is also a parameter.

We traced the arc from classical statistics to modern deep learning. An inflection point was the introduction of "memory" in models designed for sequential data, which eventually led to the transformer architecture. Transformers process language in parallel rather than sequentially, and that shift unlocked the large language models we see today.

Models in markets

There is a polarization in the choice of models: practitioners will usually either pick:

• simple ones for speed and robustness - typically high-frequency trading;
• complex ones to assist decision making - typically at the research (also called "data inference") stage.

Large language models like ChatGPT make information more accessible and arguably push markets toward informational efficiency, but predicting the most likely outcome isn't the same as generating alpha. Capacity is also an issue when more market participants use the same models to generate alpha.

When it comes to actual trading bots, these ones usually have an emphasis on explainability, but the data fed into these systems can be derived from either simple or more complex models. Full automation of critical decisions remains rare, largely because accountability has to sit with someone.

You can listen to our full conversation on Youtube.