Multidimensional Views Behind Target-Price Revisions: An LLM-Based Representation of Analyst Outlooks
summary
The gist
A target price compresses an analyst’s multidimensional views into a single number, and this paper proposes an LLM-based text representation that extracts these distinct views as question-based…
In short
The study uses a Large Language Model (LLM) to convert analyst research reports into ten question-based outlook scores instead of just a single target price. This new representation captures detailed, multidimensional analyst views on earnings, valuation, and industry conditions. The method proves superior at predicting future target-price changes beyond current market momentum.
Key concepts
- Question-Based Outlook Scores (QOS)
- These are ten scores assigned to each research report based on specific investment questions. They range from -3 to +3 and cover areas like expected EPS growth or competitive positioning. Each score is supported by a direct quote from the report, giving users a detailed, evidence-backed view of analyst sentiment.
- LLM-Based Text Representation
- The authors use an LLM to analyze the text of equity research reports. Instead of just summarizing the text, the LLM extracts structured scores that map specific investment dimensions (like valuation or industry conditions) to explicit questions. This turns unstructured text into a compact, interpretable numerical format.
- Multidimensional View
- This refers to capturing various aspects of an analyst's judgment beyond just a target price. Instead of one number, the QOS framework provides separate scores for financial performance, valuation, and industry conditions. This allows users to see how different factors influence an analyst's overall opinion.
- Subsequent Revision Prediction
- This is the ability of the model to forecast how a target price will change after a report is published. The QOS representation outperforms existing methods because these ten scores provide incremental predictive information that goes beyond simple current momentum or sector differences.
Terminology used across episodes
This episode discusses
- Multidimensional Views Behind Target-Price Revisions: An LLM-Based Representation of Analyst Outlooks · Paper Radio
- FinBERT: Financial Sentiment Analysis with Pre-trained Language Models
- M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation
- Do Sell-side Analyst Reports Have Investment Value?
- Hardening Soft Information: Evidence on Analyst Integration Costs · Paper Radio
The paper
Multidimensional Views Behind Target-Price Revisions: An LLM-Based Representation of Analyst Outlooks · Read on arXiv
Kiyoun Kim, Sung-Yeon Nathan Lim
Korea Investment Corporation
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "Multidimensional Views Behind Target-Price Revisions".
Warren: Here's the gist of the paper. A target price compresses an analyst’s multidimensional views into a single number, and this paper proposes an LLM-based text representation that extracts these distinct views as question-based outlook scores to better predict subsequent target-price revisions beyond current momentum and sector differences The Core Proposal The authors introduce an LLM-based text representation that extracts analyst outlooks from research reports as question-based outlook scores with supporting evidence These scores organize reported views on earnings, valuation, industry conditions, and other investment dimensions into a compact representation of analyst sentiment The approach aims to capture information about subsequent analyst actions that is not fully reflected in the current numerical revision By linking each dimension to an explicit investment question and supporting report evidence, the method makes the extracted outlooks available for inspection and interpretation Data and Extraction Framework The study utilizes a sample of 41,640 single-stock equity research reports on US-listed companies issued by one global brokerage between January 2017 and August 2026 For each report, the authors calculate the current target-price revision from the published target price and the associated prior target price The text preparation involves retaining only the title, summary, and body paragraphs while excluding tables and figures that are not material to the analysis Question-Based Scoring Methodology The framework converts each report into a vector of ten question-based outlook scores (QOS) Each question defines the assessment to be extracted, such as expected year-on-year EPS growth or earnings relative to market consensus For report t and question j, the model returns a score st on a seven-point scale from-3 to +3, together with a short quotation or paraphrase supporting that score The questionnaire dimensions cover financial performance, valuation, shareholder distributions, competitive positioning, industry conditions, regulation, and catalysts Evaluation and Predictive Performance The evaluation focuses on two tasks: identifying the contemporaneous target-price revision and predicting the subsequent revision The QOS representation outperforms financial dictionary and FinBERT sentiment measures on all three tasks, including prediction beyond current revision and sector controls Sector-level results show that the relevance of each outlook dimension depends on the sector, with industry conditions (Q8) being particularly prominent for upward revisions in Materials and downward revisions in Industrials The study concludes that the ten scores provide incremental predictive information in a compact representation Applications and Interpretability The question-based scores provide a multidimensional view of analyst judgments that a single target price cannot convey Individual scores can serve as report-based indicators of analyst sentiment toward earnings, valuation, industry conditions, and other investment dimensions The accompanying evidence allows users to inspect the textual basis of each assessment Because the dimensions have fixed meanings, scores can be aggregated to examine how reported conditions vary across industries and over time This structure allows users to examine which reported outlook dimensions are associated with subsequent revisions and inspect the basis of each assessment
Paper summary: Ruth: To summarize this paper, "Multidimensional Views Behind Target-Price Revisions: An LLM-Based Representation of Analyst Outlooks", the authors are tackling the problem that a single target price obscures all those different analyst opinions. The thesis is that they can use a large language model to extract these distinct views from research reports as question based outlook scores with supporting evidence.
Warren: What they claim is that these scores organize reported views on things like earnings, valuation, and industry conditions into one compact representation of analyst sentiment. This approach aims to predict subsequent target price revisions better than just looking at current momentum or sector differences in a walk-forward evaluation setting.
Iris: The paper sets up a framework where each report is converted into ten question based outlook scores. Each question defines an assessment, like expected year on year EPS growth, and the model returns a score from negative three to positive three with supporting text for that score.
Ruth: These dimensions cover financial performance, valuation, shareholder distributions, competitive positioning, industry conditions, regulation and catalysts specifically. The goal is to capture information about subsequent analyst actions that isn't fully captured in the current numerical revision.
Warren: They used a sample of forty-one thousand six hundred forty single stock equity research reports from a global brokerage between two thousand seventeen and two thousand twenty six for their study. They focus on extracting these views from the title, summary, and body paragraphs, excluding non material tables and figures.
Iris: The evaluation shows that this question based scoring method outperforms financial dictionary and FinBERT sentiment measures across three tasks: identifying the current target price revision, predicting the subsequent revision beyond current momentum or sector differences, and even controlling for sector effects.
Ruth: The key finding is that these ten scores provide incremental predictive information in a compact representation of analyst judgments. This suggests that structured questions can capture hidden information about what analysts are going to do next with a stock's price target.
Conclusion: Warren: Looking at the title, "Multidimensional Views Behind Target-Price Revisions: An LLM-Based Representation of Analyst Outlooks", the paper is proposing a way to move beyond just looking at the final target number. It’s about capturing the full picture of what analysts are actually thinking.
Iris: It suggests that by linking each dimension to an explicit investment question and supporting report evidence, you can make these extracted outlooks available for inspection and interpretation, which is a big step toward understanding analyst sentiment deeply.
Ruth: The implications are that we get a multidimensional view of analyst judgments that a single target price simply cannot convey. Individual scores can become report based indicators of analyst sentiment toward specific investment dimensions like earnings or industry conditions.
Warren: Because the dimensions have fixed meanings, you can aggregate those scores to examine how reported conditions vary across different industries and over time, which gives us a structure for analysis that's useful for tracking revision patterns.
Iris: The paper’s limitation is that while it provides this structured view of analyst sentiment, the authors themselves note that the relevance of each outlook dimension actually depends on the specific sector you are looking at. That means you can't just treat all ten scores as equally important everywhere.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2601.05274-On the use of case estimate and transactional payment data in neural networks for individual loss reserving
- 2009.00868-Loss-Given-Default Modeling by Post-Last Passage Time Process
- 2606.30070-Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures
- 2212.12687-Measuring price impact and information content of trades in a time-varying setting
- Daily Summary for 2026-11-17
- 2401.08323-Dynamic portfolio selection under generalized disappointment aversion