Study: Foundation models learn market microstructure via Bayesian inference

A new study published on arXiv explores what foundation models, specifically GPT-style transformers trained on limit order book data, learn about market microstructure. By using synthetic data from simulators with known parameters, researchers could directly compare the models' learning to optimal predictors and data entropy. The findings indicate that next-token prediction effectively performs amortized Bayesian inference over market regimes, with latent market states being decodable from the models' activations. The study suggests that training diversity, rather than sheer scale, is key for generalization to unseen markets, and that these models can transfer knowledge to real-world NASDAQ data. AI

IMPACT Demonstrates how foundation models can perform complex inference tasks, potentially improving financial market analysis and prediction.

RANK_REASON Academic paper detailing a controlled study on foundation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• DagsHub
• generative pre-trained transformer
• Gotit.pub
• Hugging Face
• IArxiv
• Litmaps
• Nasdaq
• ScienceCast
• Scite

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →