---
title: Explaining Realized Returns: Roll's $R^2$ 40 Years Later
url: https://www.ml-quant.com/papers/arxiv/2611.22277/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-11-20
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2611.22277
source_url: https://arxiv.org/abs/2611.22277
featured: 2026-11-20
citations: unknown
topic: LLMs & Text
---


# Explaining Realized Returns: Roll's $R^2$ 40 Years Later

Using AI to extract reasons from market news summaries, the research shows that narrative explanations capture 30-40% of daily and monthly equity return variation, far exceeding conventional variables.

- Source: https://arxiv.org/abs/2611.22277
- Identifier: arXiv:2611.22277
- Released: 2026-11-18
- First featured: Quant Letter No. 140 (2026-11-20): https://www.ml-quant.com/issues/2026-11-20/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: LLMs & Text
- Authors: Bryan Kelly, Yinan Su, Chao Zhang

## Abstract (arXiv, CC0)

Can we explain asset price movements even with hindsight? Reinterpreting Roll's (1988) AFA Presidential Address, "$R^2$," as a modern empirical exercise, we show that conventional macroeconomic, cash-flow, expectations, and flow variables explain less than 10% (often with negative $R^2$) of monthly U.S. equity return variation out of sample and lack daily observations. We use generative AI to extract reasons stated in Wall Street Journal market summaries while filtering out price-related language. These reasons explain around 30-40% of daily and monthly return variation out of sample. Their content spans interpretable themes, including corporate news, macroeconomic conditions, monetary policy, sentiment, and geopolitics.

## Related

- [Simple and Effective Masked Diffusion Language Models](https://www.ml-quant.com/papers/arxiv/2406.07524/): The performance of diffusion models in language modeling has been enhanced by using an effective training recipe and a simplified objective, setting a new standard among diffusion models.
- [LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models](https://www.ml-quant.com/papers/arxiv/2305.13655/): Enhancing Text-to-Image Models: The study suggests a two-stage process using a pretrained language model to improve image generation accuracy in diffusion models, enabling multi-round scene specification in various languages.
- [Predicting Liquidity-Aware Bond Yields using Causal GANs and Deep Reinforcement Learning with LLM Evaluation](https://www.ml-quant.com/papers/arxiv/2502.17011/): The paper introduces a new method for predicting bond yields using Causal Generative Adversarial Networks and reinforcement learning, which improves forecasting performance by generating synthetic bond yield data.
- [Generative AI Impact on Labor Market: Analyzing ChatGPT's Demand in Job Advertisements](https://www.ml-quant.com/papers/arxiv/2412.07042/): A study reveals the growing demand for ChatGPT-related skills in the U.S. labor market, identifying five key skill sets and emphasizing the widespread use of Generative AI in various sectors.
- [Multimodal Generative Models for Bankruptcy Prediction Using Textual Data](https://www.ml-quant.com/papers/arxiv/2211.08405/): The research presents multimodal learning in bankruptcy prediction models to tackle the problem of missing MDA section in Form 10-K, showing improved classification performance and addressing the limitation of previous models.
- [ChatGPT, Generative AI, and Investment Advisory](https://www.ml-quant.com/papers/ssrn/4519182/): The article shows that AI like ChatGPT can generate portfolio recommendations based on policy announcements, potentially outperforming markets unlike traditional textual analysis.
