{"paper_id":"2410.17266","title":"Temporal Relational Reasoning of Large Language Models for Detecting Stock Portfolio Crashes","authors":["Kelvin J. L. Koa","Yunshan Ma","Yi Xu","Ritchie Ng","Huanhuan Zheng","Tat-Seng Chua"],"paper_date":"2024-10-07","paper_type":"empirical","paper_url":"https://arxiv.org/pdf/2410.17266v2","page_url":"https://thequant.space/flowcharts/temporal-relational-reasoning-of-large-language-models-for-d/","scores":{"empirical_rigor":7.5,"math_complexity":5,"quadrant":"Holy Grail"},"proposes":{"status":"available","statement":"The paper proposes the Temporal Relational Reasoning (TRR) algorithmic framework for zero-shot detection of stock portfolio crashes.","quote":"we propose an algorithmic framework named Temporal Relational Reasoning (TRR)","summary":"Stock portfolios are often exposed to rare consequential events (e.g., 2007 global financial crisis, 2020 COVID-19 stock market crash), as they do not have enough historical information to learn from. Large Language Models (LLMs) now present a possible tool to tackle this problem, as they can…","strategy":"Temporal Relational Reasoning (TRR)"},"assumptions":{"status":"available","items":[{"assumption":"Zero-shot large language models can reason over unseen crash events without specific training data.","type":"model","stated":"explicit","quote":"... Large Language Models (LLMs) now present a possible tool to tackle this problem, as they can generalize across their large corpus of training data and perform zero-shot reasoning on new events"},{"assumption":"Daily online news articles contain sufficient information to infer impacts on portfolio stocks.","type":"data","stated":"explicit","quote":"Using the retrieved daily online news, we brainstorm for all possible impacts on the target portfolio of stocks"},{"assumption":"Transaction costs are 0.02 per reallocation when evaluating portfolio performance.","type":"execution","stated":"explicit","quote":"We use 0.02 as the transaction cost for each reallocation."},{"assumption":"The Reuters financial news dataset is unfiltered and can be used directly for relevance decisions.","type":"data","stated":"explicit","quote":"For news data, we use the Reuters financial news dataset [16], which we also extend to the year 2020"},{"assumption":"Portfolio crash labels are rare, leading to imbalanced datasets.","type":"statistical","stated":"explicit","quote":"As the portfolios are widely diversified and the events that have significant impacts are usually rare, the percentage of data labelled as Crash are low."}]},"datasets":{"status":"available","items":[{"name":"Reuters financial news dataset","provider":"Reuters","kind":"text-news","access":"public","period":"extended to 2020","universe":null,"quote":"For news data, we use the Reuters financial news dataset [16], which we also extend to the year 2020 to cover the selected time periods."},{"name":"Portfolio crash dataset","provider":null,"kind":"other","access":"unknown","period":"2007, 2010, 2020, 2012, 2022","universe":null,"quote":"... 2022 Oct 2021-Mar 2022 Supplementary (Post-Knowledge)"}],"known":[{"id":"yahoo-finance","name":"Yahoo Finance","provider":"Yahoo","kind":"market-prices","access":"public","quote":"9 0.577±0.0854 0.537±0.0129 0.522±0.0709 0.573±0.0352 0.600±0.0514 0.512±0.0108 0.531±0.0230 ToG 0.502±0.0589 0.629±0.0453 0.585±0.0160 0.588±0.0118 0.607±0.0540 0.625±0.0558 0.579±0.0423 0.587±0.0399 TRR (Ours) 0.690±0.0426 0.684±0.0361 0.610±0.0625 0.598±0.0787 0.657±0.0231 0.644±0.0350 0.638±0.0348 0.601±0.0425 To generate the portfolio crash labels, we first retrieve the price data of each constituent stock from ","mentions":1,"used_in":1}],"sample":{"start":null,"end":null,"market":"stock portfolios (country-neutral)","asset_class":"equities","frequency":"daily"},"data_resolution":"daily"},"transaction_costs":{"status":"available","modeled":"unclear","headline_net_of_costs":false,"detail":null,"quote":null,"source":"conflict","rule":"R4b","candidates":{"front_matter":"partial","evidence_card":"yes","evidence_card_quoted":true},"conflict_kind":"graded","card_detail":"0.02 per reallocation (flat fee)","card_quote":"We use 0.02 as the transaction cost for each reallocation.","meaning":"the platform classifier and the paper-text extraction disagree, so no value is asserted"},"results":{"status":"available","evaluation_type":"unclear","findings":[{"finding":"TRR outperforms all baselines, achieving an average AUROC improvement of 8.93% over the strongest baseline (ToG).","kind":"performance","quote":"Finally, our TRR framework was able to outperform all models, by an average of 8.93% over the strongest baseline (ToG)."},{"finding":"TRR achieves the highest cumulative returns and lowest maximum drawdowns during crash periods compared to other methods.","kind":"performance","quote":"The TRR portfolio was able to avoid the most losses, achieving the highest cumulative returns and maximum drawdowns in a market downturn."},{"finding":"TRR maintains strong performance on the 2022 post‑knowledge‑cutoff dataset, indicating results are due to reasoning rather than internal knowledge.","kind":"performance","quote":"TRR was also able to perform well in the 2022 dataset, which is after the post‑knowledge cut‑off date. This shows that the performance can be attributed to the"},{"finding":"In the macro‑economic global crisis detection task, TRR attains AUROC scores of 0.940, 0.747, and 0.961 for 2007, 2011, and 2020 datasets respectively, surpassing traditional indicators.","kind":"predictive-accuracy","quote":"TRR (Ours) 0.940±0.023 0.747±0.013 0.961±0.044"}],"reported_performance":{"strategy":"Temporal Relational Reasoning (TRR)","sharpe":0.2202,"sharpe_benchmark":-0.0486,"annual_return_pct":null,"total_return_pct":14.59,"max_drawdown_pct":3.56},"out_of_sample":{"method":"holdout","headline_out_of_sample":true}},"independent_evidence":{"status":"none_found","supports":[],"contradicts":[],"qualifies":[],"compared_with":35,"citations":{"cited_by_in_corpus":2,"by_independent_authors":2,"how_cited":{"used_as_baseline":0,"uses_method_or_data":0,"extends":0,"discusses_findings":0,"mentions":2,"unclassified":0},"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":"2026-11-14T17:59:47Z"},"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/"}