{"paper_id":"2509.01590","title":"Is All the Information in the Price? LLM Embeddings versus the EMH in Stock Clustering","authors":["Bingyang Wang","Grant Johnson","Maria Hybinette","Tucker Balch"],"paper_date":"2025-09-01","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2509.01590v1","page_url":"https://thequant.space/flowcharts/is-all-the-information-in-the-price-llm-embeddings-versus-t/","scores":{"empirical_rigor":8,"math_complexity":6.5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes a generalizable methodology that converts any equity grouping into a real‑time synthetic factor model for out‑of‑sample evaluation.","quote":"a generalizable methodology that converts any equity grouping: manual, machine, or market-driven, into a real-time factor model for evaluation","summary":"This paper investigates whether artificial intelligence can enhance stock clustering compared to traditional methods. We consider this in the context of the semi-strong Efficient Markets Hypothesis (EMH), which posits that prices fully reflect all public information and, accordingly, that clusters…","strategy":null},"assumptions":{"status":"available","items":[{"assumption":"Semi‑strong EMH holds, i.e., prices fully reflect all public information.","type":"market","stated":"explicit","quote":"The semi‑strong form of the Efficient Market Hypothesis (EMH) states that prices fully reflect all public information"},{"assumption":"Number of clusters is fixed at K = 11 to match GICS sectors.","type":"model","stated":"explicit","quote":"we choose k = 11 because GICS is constrained to 11 sectors"},{"assumption":"Linear factor loadings are estimated by OLS on one month of daily returns.","type":"model","stated":"explicit","quote":"At the end of each period, we estimate a linear model for each stock using one month of daily return data and Ordinary Least Squares regression"},{"assumption":"Clusters are re‑computed weekly and predictions are made for the following week.","type":"execution","stated":"explicit","quote":"The linear model for each clustering is reevaluated weekly, and the errors are computed for each day rolling forward"},{"assumption":"Equal‑weighted average returns represent each cluster’s return series.","type":"model","stated":"explicit","quote":"For each cluster, we construct a daily return series by taking an equal-weighted average of the returns of all stocks in that cluster"},{"assumption":"Back‑testing period starts 1 Jan 2022 due to the embedding knowledge‑cutoff.","type":"data","stated":"explicit","quote":"Our backtesting analysis is limited to the period beginning January 1, 2022, due to the September 2021 knowledge cutoff of OpenAI’s text-embedding-3-large embeddings"}]},"datasets":{"status":"available","items":[{"name":"S&P 500 constituents","provider":"WRDS / Compustat","kind":"market-prices","access":"commercial","period":"2022‑2024","universe":"S&P 500 constituents","quote":"Using S&P 500 constituents from from 2022 through 2024","known_id":"compustat"},{"name":"Historical price data","provider":"WRDS","kind":"market-prices","access":"commercial","period":"2022‑2024","universe":"S&P 500 constituents","quote":"Our price data is also sourced from WRDS","known_id":"wrds"},{"name":"GICS classification","provider":"Compustat North America Daily Updates","kind":"fundamentals","access":"commercial","period":"2022‑2024","universe":"S&P 500 constituents","quote":"We use point-in-time GICS classification data from the Compustat North America Daily Updates database","known_id":"compustat"},{"name":"News headline data","provider":"RavenPack","kind":"text-news","access":"commercial","period":"2022‑2024","universe":"S&P 500 constituents","quote":"Our stock-level headline data is sourced from RavenPack","known_id":"ravenpack"},{"name":"LLM embeddings","provider":"OpenAI text‑embedding‑3‑large","kind":"text-news","access":"proprietary","period":"2022‑2024","universe":"S&P 500 constituents","quote":"OpenAI’s text-embedding-3-large embeddings used in our"}],"known":[{"id":"wrds","name":"WRDS","provider":"Wharton Research Data Services","kind":"other","access":"commercial","quote":"Data We use point-in-time GICS classification data from the Compustat North America Daily Updates database [16], accessed via WRDS [2].","mentions":4,"used_in":2},{"id":"compustat","name":"Compustat","provider":"S&P Global","kind":"fundamentals","access":"commercial","quote":"Data We use point-in-time GICS classification data from the Compustat North America Daily Updates database [16], accessed via WRDS [2].","mentions":2,"used_in":2},{"id":"ravenpack","name":"RavenPack","provider":"RavenPack","kind":"text-news","access":"commercial","quote":"Similarity Measure 3: LLM Embeddings of News Headlines We collect press release headlines for each equity from RavenPack, filtering to relevance score 100 (highest relevance), discretionarily removing 20 noisy article types, and applying TF-IDF deduplication by removing later articles from same-day pairs with cosine similarity above 0.3.","mentions":2,"used_in":2}],"sample":{"start":null,"end":null,"market":null,"asset_class":null,"frequency":null},"data_resolution":"daily"},"transaction_costs":{"status":"not_applicable","modeled":"n/a","headline_net_of_costs":null,"detail":null,"quote":null,"source":"front_matter","rule":"R3","candidates":{"front_matter":"n/a","evidence_card":"n/a","evidence_card_quoted":false},"meaning":"the paper reports no trading or execution result"},"results":{"status":"available","evaluation_type":"forecast-accuracy","findings":[{"finding":"Price‑based clustering reduces RMSE by 15.9% relative to GICS and 14.7% relative to LLM embeddings.","kind":"performance","quote":"price-based clustering consistently outperforms both rule-based and AI-based methods, reducing root mean squared error (RMSE) by 15.9% relative to GICS and 14.7% relative to LLM embeddings"},{"finding":"The performance difference between the best returns‑based and best LLM‑based clustering is statistically significant (p‑value = 0).","kind":"statistical","quote":"The resulting test yields a p-value of 0, indicating that the difference in performance between the two methods is statistically significant"},{"finding":"Human‑judged GICS and AI‑derived news embeddings do not provide superior predictive factors over correlation‑based clusters.","kind":"negative","quote":"Human judgments regarding “economic similarity” and sophisticated NLP extractions of news text may be useful in other contexts, but they do not translate into superior predictive factors"},{"finding":"Dynamic correlation groupings are sufficient for portfolio construction and risk modeling.","kind":"theoretical","quote":"For portfolio construction and risk modeling, relying on dynamically updated correlation groupings appears sufficient and essential to capture the cross‑sectional structure of systematic risk"}],"reported_performance":null,"out_of_sample":{"method":"walk-forward","headline_out_of_sample":null}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":34,"citations":{"cited_by_in_corpus":1,"by_independent_authors":1,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":0,"unclassified":1},"independent_papers":[],"scope":"Citations between papers in this corpus only; citations from outside it are not counted."}},"implementation":{"status":"available","public":false,"code_url":null,"note":"No repository presented by the authors as their own was found in the paper. Absence here is not proof that no code exists."},"replication":{"status":"none_found","card":null,"note":"No replication of this paper has been run by thequant.space yet."},"extraction":{"text_source":"fulltext","model":"openai/gpt-oss-120b","version":1,"extracted_at":"2026-11-14T17:28:31Z"},"provenance":{"claims":"Everything here is what the authors state in the paper. Nothing was re-run or reproduced by the platform.","quotes":"Assumptions, datasets, cost detail and findings were extracted by a language model. Every quote is verbatim paper text. Where an item has `support`, \"full\" means every figure in it appears in the quoted sentences and \"partial\" means a figure is stated elsewhere in the paper; items with a figure found nowhere in the paper were discarded.","datasets_known":"datasets.known lists well-known vendors and public datasets by a stable id, found by a dictionary over the full text where the sentence says the data is used; it is independent of the model and covers only names in the dictionary.","flags":"transaction_costs.modeled, results.evaluation_type and results.out_of_sample.method are platform classifications (structured-decision models); see https://thequant.space/score-guide/#penalty-flags","comparisons":"supports / contradicts / qualifies are a language model's reading of two papers' quoted findings, restricted to closely related papers that share no author. They are pointers to read, not verified replications, and coverage is partial.","implementation":"code_url is a repository the paper itself presents as its own (rule-scored, ambiguous cases judged by a model), checked for liveness when it was recorded."},"docs":"https://thequant.space/agents/"}