Paper: arXiv 2611.22277

Authors: Bryan Kelly, Yinan Su, Chao Zhang

Abstract

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.

Complexity vs Empirical Score

  • Math Complexity: 5.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper presents a highly novel approach to explaining market returns using generative AI and textual analysis, addressing a long-standing puzzle in finance. Its empirical methodology is robust, featuring out-of-sample R2 evaluations and careful masking of tautological language. The clarity and completeness are strong, though full reproducibility details would depend on the availability of the specific AI models and data used.

Research Flowchart

  flowchart TD
    A["Research Goal: Explain Realized Returns with Modern Empirical Exercise"] --> B{"Conventional Methods Analysis"};
    B --> C1["Input: Macroeconomic, Cash-flow, Expectations, Flow Variables"];
    C1 --> D1["Process: OOS R^2 Calculation, Monthly Data"];
    D1 --> E1["Finding: <10% Explained, Often Negative R^2, No Daily Data"];
    B --> C2["Input: Wall Street Journal Market Summaries"];
    C2 --> D2["Process: Generative AI Extraction of Reasons (Filtered for Price Language)"];
    D2 --> E2["Finding: 30-40% Explained (Daily/Monthly OOS), Thematic Content (Corporate, Macro, Monetary, Sentiment, Geopolitics)"];
    E1 & E2 --> F["Key Outcome: Generative AI Reasons Significantly Improve Explanatory Power"];

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes extracting reasons from Wall Street Journal market commentary using generative AI and using these news‑based reasons to explain a substantial share of daily and monthly U.S. equity return variation.

Source sentence
These reasons explain around 30-40% of daily and monthly return variation out of sample.

Assumes 6

  • Explanatory variables must be observed independently of asset returns (criterion 1). data
    Source sentence
    Both components are therefore derived from concurrent return data and do not satisfy criterion 1.
  • All non‑news baseline series are observed at monthly or lower frequency, limiting their explanatory power at higher frequencies. data
    Source sentence
    All non-news explanations considered in this paper are observed at monthly or lower frequency, whereas news can be used at the daily frequency.
  • Quarterly macro, cash‑flow, expectations, and flow data can be transformed to monthly frequency by linear interpolation. data
    Source sentence
    In particular, the following simple example illustrates the linear transformation: YMar = Y (4Q) Q1 , YApr = 2 3 Y (4Q) Q1 + 1 3 Y (4Q) Q2 , YMay = 1 3 Y (4Q) Q1 + 2 3 Y (4Q) Q2 , YJun = Y (4Q) Q2 . Macro statistics.
  • The out‑of‑sample R² is evaluated via five‑fold cross‑validation with strict hold‑out folds. statistical
    Source sentence
    We evaluate the explanatory power of the estimated relationship with out-of-sample R2 using a five-fold cross-validation split. Specifically, we split the sample period into five folds.
Show 2 more
  • Zero‑shot inference contains a persistence bias that can be removed by HAR filtering. model
    Source sentence
    Therefore, filtering does not introduce realized-return leakage: the filtered series remains consistent with Criterion 1 because it is not derived from asset returns. Filtering consistently improves R2 for both methods and at both frequencies, indicating both the prevalence of persistence bias and the effectiveness of HAR filtering in removing it.
  • News‑based reasons are filtered to remove any leakage from realized returns. data
    Source sentence
    Therefore, filtering does not introduce realized-return leakage: the filtered series remains consistent with Criterion 1 because it is not derived from asset returns.

Data

  • Wall Street Journal market‑summary reasons dataset · Wall Street Journal text-news commercial
    2002–2024
    Source sentence
    We use generative AI to extract reasons stated in Wall Street Journal market summaries while filtering out price-related language. The full time series sample period is 2002–2024, aligned with our news-based reasons dataset.
  • FRED‑MD macroeconomic statistics · FRED‑MD (McCracken & Ng, 2016) macro public
    Source sentence
    Macro Macroeconomic statistics: output, labor, etc., FRED-MD dataset excluding financial market indexes (McCracken and Ng, 2016) 99 Cash flow Dividends and earnings, aggregated at the S&P 500 level, from Shiller’s (2015) webpage 4 Subjective expectations Analyst forecasts of dividend and earnings, aggregated at the S&P 500 level (Delao and Myers, 2021) 4 Trading flows Net equity flows of various investor sectors,
  • S&P 500 aggregate dividends and earnings · Robert Shiller website fundamentals public
    S&P 500
    Source sentence
    Macro Macroeconomic statistics: output, labor, etc., FRED-MD dataset excluding financial market indexes (McCracken and Ng, 2016) 99 Cash flow Dividends and earnings, aggregated at the S&P 500 level, from Shiller’s (2015) webpage 4 Subjective expectations Analyst forecasts of dividend and earnings, aggregated at the S&P 500 level (Delao and Myers, 2021) 4 Trading flows Net equity flows of various investor sectors,
  • Analyst forecasts of dividend and earnings (subjective expectations) · Delao & Myers (2021) via Sean Myers’ webpage macro public
    S&P 500
    Source sentence
    Macro Macroeconomic statistics: output, labor, etc., FRED-MD dataset excluding financial market indexes (McCracken and Ng, 2016) 99 Cash flow Dividends and earnings, aggregated at the S&P 500 level, from Shiller’s (2015) webpage 4 Subjective expectations Analyst forecasts of dividend and earnings, aggregated at the S&P 500 level (Delao and Myers, 2021) 4 Trading flows Net equity flows of various investor sectors,
  • Net equity flows by investor sector · Gabaix & Koijen (2021) methodology alternative public
    12 investor sectors
    Source sentence
    Macro Macroeconomic statistics: output, labor, etc., FRED-MD dataset excluding financial market indexes (McCracken and Ng, 2016) 99 Cash flow Dividends and earnings, aggregated at the S&P 500 level, from Shiller’s (2015) webpage 4 Subjective expectations Analyst forecasts of dividend and earnings, aggregated at the S&P 500 level (Delao and Myers, 2021) 4 Trading flows Net equity flows of various investor sectors,

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • Conventional macro, cash‑flow, expectations, and flow variables explain less than 10 % of monthly U.S. equity return variation out‑of‑sample, often yielding negative R². predictive-accuracy
    Source sentence
    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.
  • News‑based reasons explain roughly 30–40 % of daily and monthly return variation out‑of‑sample. predictive-accuracy
    Source sentence
    These reasons explain around 30-40% of daily and monthly return variation out of sample.
  • Zero‑shot inference in raw form fails at the monthly frequency, but after filtering out temporal persistence it achieves an out‑of‑sample monthly R² of nearly 50 %. predictive-accuracy
    Source sentence
    We also show that the zero-shot inference method fails when aggregated to the monthly frequency in its raw form, but can achieve a monthly out-of-sample R2 of nearly 50% with a simple adjustment that filters out the temporal persistence in the raw zero-shot predictions.
  • HAR filtering improves zero‑shot monthly R² from –115.03 % to 49.84 % and daily R² from 25.58 % to 31.88 %; embedding‑regression also improves after filtering. predictive-accuracy
    Source sentence
    At the daily frequency, zero-shot R2 improves from 25.58% to 31.88%, exceeding the embedding-regression benchmark of 29.82%. The largest improvement is at the monthly frequency, where zero-shot R2 rises from −115.03% to 49.84% after filtering.
  • Flows have some explanatory power, but macro, cash‑flow, and expectations series show little robust explanatory power for market fluctuations. statistical
    Source sentence
    12 tent with mechanisms in which flows have price impact or correlate with investor sentiment and other fundamental drivers of market movements (Shleifer and Summers, 1990; Gabaix and Koijen, 2021).11 The other three explanations (macroeconomic statistics, cash flow, and subjective expectations) do not have robust explanatory power for market fluctuations. These findings may simply reflect that the measured

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-18 · same card as JSON · A language model chose these items: check the source sentence before relying on one.