A curated reading list
A research-grade catalog of 308 papers on financial-domain LLMs — benchmarks, SEC filings, trading agents, alpha mining, RAG over filings, professional exams — each assigned to exactly one primary category, 277 with deep structured summaries.
Survey, review, taxonomy and overview papers that map the finance-LLM landscape.
This survey synthesizes the state-of-the-art in applying large language models to finance, reviewing solution architectures from zero-shot to custom pretraining, proposing a cost-aware decision framework for adoption, and highlighting limitations regarding hallucination and bias.
This survey provides a comprehensive overview of Financial Large Language Models (FinLLMs), comparing their evolution from general-domain models, analyzing training techniques, evaluating performance across six benchmark tasks, and identifying challenges and future directions for the field.
This survey provides a comprehensive review of large language models in finance, categorizing applications into linguistic tasks, sentiment analysis, time series, reasoning, and agent-based modeling, while detailing specific models, datasets, benchmarks, and critical challenges like lookahead bias and data pollution.
This survey synthesizes over fifty primary studies to propose a task-centered taxonomy for large language models in equity return prediction and trading, emphasizing the transition from feature-centric text mining to end-to-end agentic decision systems with rigorous evaluation standards.
This survey systematically reviews 27 studies on LLM-powered trading agents, categorizing them into 'LLM as Trader' and 'LLM as Alpha Miner' architectures, analyzing their data inputs, evaluation metrics, and identifying key limitations such as short backtesting periods and lack of cost considerations.
This survey systematically reviews the evolution of quantitative investment from traditional statistical models to deep learning and large language models, focusing on how AI enhances the alpha strategy pipeline across data processing, prediction, and execution.
This survey bridges the gap between rapid LLM advancements and cautious financial adoption by comprehensively reviewing datasets, models, and applications across trading, risk, and ESG, while outlining future research directions for reasoning, RAG, and multi-agent systems.
This survey reviews the integration of Large Language Models into finance, categorizing applications into engineering, forecasting, risk management, and QA, while empirically evaluating GPT-4's ability to follow instructions across these diverse financial tasks.
This survey bridges financial practice and LLM research by proposing a dual-perspective framework that maps five core financial domains to specific agent tasks, catalogs over 30 benchmarks and 20 models, and identifies critical gaps in numerical reasoning, real-time adaptability, and regulatory compliance.
This 2024 review systematically analyzes 90 papers on generative AI in finance using BERTopic, identifying three main themes: LLM applications for financial tasks, risks and regulatory challenges, and synthetic data generation via GANs.
This survey provides a comprehensive review of large language models in finance, healthcare, and law, categorizing tasks, models, and ethical challenges to guide future interdisciplinary research in high-stakes domains.
This survey categorizes LLM integration in finance into four frameworks, highlighting that hybrid and agent-based methods significantly outperform baselines in stock prediction and portfolio optimization, though challenges in hallucination and real-time adaptability remain.
This review synthesizes LLM applications in stock forecasting from a hedge-fund perspective, critiquing academic practices for ignoring market frictions like data leakage, illiquidity, and regime shifts while proposing rigorous evaluation standards for production deployment.
This 2025 scientific review synthesizes the application of autonomous AI agents across finance, contrasting traditional rule-based and reinforcement learning systems with emerging LLM-driven architectures to evaluate their efficacy, ethical implications, and future research directions.
This scoping review synthesizes 264 recent papers on ChatGPT and LLMs in accounting and finance, identifying three core themes: domain applications, LLMs as research tools, and professional implications, while providing a framework for future inquiry.
This 2025 review synthesizes the landscape of AI agent frameworks like LangGraph, CrewAI, and AutoGen, evaluating their architectures, financial applications in risk and trading, and critical challenges regarding regulatory compliance, explainability, and risk alignment.
This paper presents the first systematic review and taxonomy of LLM-based credit risk assessment, categorizing 60 studies by model architecture, data modality, interpretability mechanisms, and application domains to highlight trends and research gaps.
This 2025 survey synthesizes the landscape of generative AI agent frameworks and their applications in finance, highlighting productivity gains, multi-agent architectures, and deployment challenges.
This paper demonstrates that multi-agent LLM frameworks can effectively automate the cross-verification of underlying assets in structured finance, achieving near-perfect accuracy in matching loan applications to bank statements while offering a cost-effective alternative to manual auditing.
This structured literature review analyzes 14 recent studies to map the current state, application scenarios, and significant challenges of integrating Large Language Models and Generative AI into core banking operations, highlighting a gap between experimental potential and productive deployment due to regulatory and reliability constraints.
Financial-domain LLMs, FinBERT-style models, instruction tuning, and domain adaptation.
BloombergGPT is a 50 billion parameter language model trained on a mixed corpus of 363 billion financial tokens and 345 billion general tokens, demonstrating superior performance on financial benchmarks while maintaining competitive general NLP capabilities.
FinGPT introduces an open-source, data-centric framework for financial LLMs that utilizes lightweight LoRA fine-tuning and market-driven reinforcement learning to achieve superior financial sentiment analysis compared to proprietary and baseline models.
PIXIU introduces FinMA, an open-source financial LLM fine-tuned on the FIT dataset, and FLARE, a benchmark evaluating financial NLP and stock prediction, showing FinMA outperforms general LLMs on NLP tasks but struggles with quantitative reasoning.
FLaME introduces a holistic benchmarking suite for financial NLP, evaluating 23 foundation and reasoning-reinforced LMs across 20 datasets to reveal that no single model dominates all tasks, with significant performance gaps in numeric reasoning and causal analysis, while highlighting cost-performance trade-offs for open-weight models.
FinBERT, a BERT-based language model further pre-trained on financial text, achieves state-of-the-art performance in financial sentiment analysis on the Financial PhraseBank and FiQA datasets, significantly outperforming traditional machine learning and other transfer learning baselines like ULMFit and ELMo.
FinBERT is a domain-specific BERT model pretrained on 4.9 billion tokens of financial text that significantly outperforms generic BERT on financial sentiment classification tasks.
The paper demonstrates that multi-task fine-tuning on a cocktail of financial and general datasets significantly enhances LLM performance on specific downstream tasks, enabling small models like Phi-3-Mini to surpass larger models like GPT-4-o, though this does not necessarily improve broader domain knowledge.
The paper introduces FinCDM, a cognitive diagnosis framework for evaluating financial LLMs at the knowledge-skill level, revealing hidden gaps and specialization patterns that aggregate benchmarks miss.
FinGPT introduces a cost-effective instruction tuning benchmark for open-source LLMs, demonstrating that task-specific and multi-task tuning significantly improves financial NLP performance, while zero-shot generalization remains challenging for most models except ChatGLM2.
The paper introduces FLANG, a financial language model using preferential masking of financial terms and phrases, and FLUE, a comprehensive benchmark suite, demonstrating that domain-specific pre-training significantly outperforms general and prior financial models on sentiment analysis, news classification, and question answering tasks.
FinGPT introduces an open-source framework for democratizing financial LLMs by automating the collection of real-time data from 34+ sources and enabling lightweight fine-tuning via LoRA and a novel Reinforcement Learning with Stock Prices (RLSP) method.
Instruct-FinGPT improves financial sentiment analysis by instruction-tuning LLaMA-7B, significantly outperforming FinBERT and ChatGPT in numerical sensitivity and contextual understanding.
The paper demonstrates that instruction fine-tuning smaller open-source LLMs like Mistral-7B and Llama3-8B significantly improves financial text classification, with model merging via MergeKit effectively mitigating catastrophic forgetting on unseen tasks.
FinBERT is a domain-specific BERT model pre-trained on large-scale financial corpora using six multi-task self-supervised objectives, achieving state-of-the-art performance on financial sentence boundary detection, sentiment analysis, and question answering benchmarks.
The paper introduces FinBERT-FOMC, a fine-tuned FinBERT model enhanced with a Sentiment Focus preprocessing method, which significantly improves sentiment classification accuracy on complex FOMC Minutes sentences containing contrasting conjunctions.
This study demonstrates that GPT-4o, when optimized through systematic prompt engineering, outperforms the domain-specific FinBERT model in financial sentiment analysis across business, health, and technology sectors by up to 10%.
This study demonstrates that zero-shot prompting of ChatGPT 3.5 significantly outperforms FinBERT in classifying sentiment from forex news headlines and correlates more strongly with market returns, highlighting the value of prompt engineering in financial NLP.
The paper evaluates LLM reasoning in financial sentiment analysis by introducing a Financial Attribute Prompting framework, revealing that models like GPT-3.5 and PaLM-2 lack inherent capabilities in numerical and comparative reasoning despite improvements from explicit prompts.
The paper demonstrates that zero-shot financial sentiment classification aligns better with human judgment when using intuitive, direct prompting (System 1) rather than deliberative chain-of-thought reasoning (System 2), which often leads to overthinking and reduced accuracy.
The paper introduces an open, continuously updated FinLLM leaderboard developed with the Linux Foundation and Hugging Face to standardize the evaluation of financial language models and agents across multimodal tasks, addressing the static nature of existing benchmarks.
General finance LLM benchmarks, evaluation suites, exams, leaderboards and broad task collections.
FinBen is a comprehensive open-source benchmark evaluating 15 LLMs across 24 financial tasks, revealing that while models excel at information extraction, they struggle with complex reasoning, forecasting, and trading, with GPT-4 leading in IE and trading while Gemini leads in generation.
The paper introduces FLAME, a comprehensive Chinese financial LLM evaluation system comprising certification and scenario benchmarks, demonstrating that Baichuan4-Finance outperforms other models in professional knowledge and practical application tasks.
FinanceReasoning is a rigorous benchmark for evaluating large reasoning models in financial numerical tasks, demonstrating that combining reasoner and programmer models with structured knowledge retrieval significantly improves accuracy on complex financial calculations.
FinDABench is a comprehensive benchmark evaluating LLMs on financial data analysis across foundational, reasoning, and technical skills, revealing that even GPT-4 achieves only 32.37% accuracy in zero-shot settings, highlighting significant gaps in domain-specific analytical capabilities.
This survey systematically categorizes financial foundation models into language, time-series, and visual-language types, reviewing their architectures, training strategies, datasets, and applications while identifying key challenges in data, algorithms, and infrastructure.
The paper demonstrates that data-centric post-training strategies, specifically high-quality Chain-of-Thought distillation for SFT and difficulty-aware sampling for RL, enable an 8B-parameter financial LLM to outperform larger and specialized open-source competitors across general, sentiment, and numerical reasoning benchmarks.
The paper introduces CALM, a fine-tuned LLM for credit scoring, demonstrating that LLMs can generalize across diverse financial risk tasks but exhibit significant demographic biases requiring ethical oversight.
This empirical study benchmarks ChatGPT and GPT-4 against domain-specific models and fine-tuned baselines across eight financial NLP datasets, revealing that generalist LLMs excel in sentiment analysis and numerical reasoning but lag in structured prediction tasks like named entity recognition and relation extraction.
This study evaluates the open-source financial LLM FinMA-7B-full on the FLARE benchmark, demonstrating strong performance in sentiment analysis and classification but significant weaknesses in numerical reasoning, entity recognition, and summarization compared to general-purpose baselines.
BizFinBench introduces a business-driven Chinese financial benchmark with 6,781 real-world queries and the IteraJudge evaluation framework, revealing that while proprietary models lead in reasoning and extraction, no single model dominates all financial tasks, highlighting significant gaps in complex cross-concept reasoning.
PIXIU introduces FinMA, an open-source financial LLM fine-tuned on the FIT dataset, and FLARE, a benchmark covering financial NLP and stock prediction tasks, demonstrating that domain-specific instruction tuning significantly improves performance on most financial NLP tasks compared to general and proprietary models.
RealFin introduces a bilingual benchmark evaluating LLMs' ability to detect under-specified financial questions, revealing that general models over-commit to guesses while finance-specialized models fail to identify missing premises.
This study benchmarks ChatGPT and GPT-4 against domain-specific models and fine-tuned baselines across eight financial NLP datasets, revealing that generalist LLMs excel in sentiment analysis and numerical reasoning but lag in structured prediction tasks like named entity recognition and relation extraction.
RiskLabs is a multimodal framework that leverages LLMs to extract features from earnings calls, news, and time series data to significantly improve financial risk prediction compared to traditional baselines.
The paper demonstrates that using Sentence-BERT embeddings to encode non-semantic categorical features in general ledger journal entries significantly improves anomaly detection performance compared to traditional one-hot encoding, particularly when paired with Logistic Regression and Neural Networks.
This position paper identifies five critical biases in financial LLM evaluation—look-ahead, survivorship, narrative, objective, and cost—and proposes a Structural Validity Framework with a checklist to enforce rigorous, deployment-ready standards.
The paper proposes MANAGER, a multimodal LLM framework that integrates external financial knowledge and adaptive graph-based fusion of text, video, and audio from Monetary Policy Conference calls to outperform baselines in predicting asset price movements and volatility.
This paper demonstrates that LLMs can effectively detect money laundering in financial graphs via in-context learning on serialized subgraphs, achieving 63.7% accuracy while providing interpretable rationales, though scalability remains a challenge.
Multimodal, multilingual, bilingual and non-English financial LLM resources and evaluations.
FinEval is a comprehensive Chinese financial benchmark evaluating 19 LLMs across academic knowledge, industry tasks, security, and agent capabilities, revealing that while top models surpass ordinary humans, they still lag behind experts, with general models often outperforming domain-specific ones.
Golden Touchstone introduces a comprehensive bilingual benchmark for financial LLMs across eight NLP tasks, revealing that while models excel at sentiment analysis, they struggle with complex reasoning and stock prediction, and proposes Touchstone-GPT as a competitive baseline.
The paper introduces BBT-FinT5, a large-scale Chinese financial pre-trained language model enhanced with knowledge extraction, supported by a 300GB corpus and a comprehensive benchmark, demonstrating superior performance over existing domain-specific models on financial NLP tasks.
FLAME introduces a comprehensive Chinese financial LLM evaluation system comprising certification and scenario benchmarks, revealing that Baichuan4-Finance outperforms other models in professional knowledge and practical application tasks.
XuanYuan 2.0 is a 176-billion parameter Chinese financial chat model built on BLOOM-176B that utilizes a novel hybrid-tuning method to integrate general and financial knowledge while mitigating catastrophic forgetting.
CFGPT introduces a Chinese financial LLM framework comprising a large-scale pre-training dataset (CFData), a 7B-parameter model (CFLLM) fine-tuned via continued pre-training and supervised instruction tuning, and a deployment system (CFAPP) supporting multimodal inputs and complex financial reasoning tasks.
DISC-FinLLM is a Chinese financial large language model built on Baichuan-13B using a Multiple Experts Fine-tuning Framework with four specialized LoRA modules, demonstrating superior performance over baselines in financial NLP, computation, and retrieval tasks.
The paper introduces ICE-PIXIU, an open-source framework comprising the ICE-FIND bilingual instruction dataset, the ICE-INTENT fine-tuned LLM, and the ICE-FLARE benchmark, demonstrating that bilingual fine-tuning significantly enhances Chinese financial NLP performance, often surpassing GPT-4.
FinTral is a multimodal financial LLM based on Mistral-7B that achieves GPT-4-level performance on nine financial tasks by integrating domain-specific pretraining, RLAIF alignment, and tool-augmented retrieval, while introducing FinSet, a comprehensive benchmark for evaluating financial hallucinations.
Open-FinLLMs introduces a suite of open-source multimodal financial LLMs, including FinLLaMA and FinLLaVA, which outperform general and specialized baselines like GPT-4 and BloombergGPT across text, tabular, time-series, and chart data tasks.
The paper introduces Toisón de Oro, a bilingual Spanish-English financial LLM framework including the FinMA-ES model, FIT-ES instruction dataset, and FLARE-ES benchmark, demonstrating that FinMA-ES outperforms GPT-4 in Spanish financial tasks through cross-linguistic transfer.
The paper introduces Plutus-ben, the first Greek financial NLP benchmark, and Plutus-8B, a fine-tuned Greek financial LLM, demonstrating that domain-specific training significantly outperforms general multilingual models in low-resource financial contexts.
The paper introduces CFLUE, a comprehensive Chinese financial LLM benchmark with 38K+ knowledge questions and 16K+ application tasks, revealing that while GPT-4 leads in knowledge, lightweight models often match or exceed it in application tasks, and specialized financial LLMs underperform general ones.
This position paper surveys Multimodal Financial Foundation Models (MFFMs), detailing their data-centric approach, agentic applications, and critical challenges in transparency, reasoning, and governance.
MME-Finance is a bilingual multimodal benchmark evaluating 19 MLLMs on financial charts, revealing that while top models like GPT-4o and Qwen2VL-72B show competence, they struggle significantly with fine-grained visual perception, spatial awareness, and domain-specific chart interpretation.
The paper introduces FinMME, a large-scale multimodal benchmark with over 11,000 high-quality financial samples across 18 domains, demonstrating that state-of-the-art MLLMs struggle with financial chart reasoning and introducing FinScore to penalize hallucinations.
The paper proposes PreBit, a multimodal model combining FinBERT-embedded Twitter text with technical analysis data to predict extreme Bitcoin price movements, demonstrating that social media content improves prediction accuracy and enables a lower-risk trading strategy.
Question answering, numerical reasoning, financial table/text reasoning, discrete reasoning.
The paper introduces FinQA, a large-scale dataset of 8,281 expert-annotated financial question-answer pairs with executable reasoning programs over S&P 500 earnings reports, demonstrating that current LLMs significantly lag behind human experts in complex multi-step numerical reasoning.
The paper introduces TAT-QA, a benchmark for financial question answering over hybrid tabular and textual data, and proposes the TAGOP model which uses sequence tagging and symbolic aggregation operators to achieve 58.0% F1, significantly outperforming baselines but lagging behind human experts.
ConvFinQA introduces a large-scale dataset for conversational financial question answering that requires complex, multi-turn numerical reasoning chains, revealing that both neural symbolic and prompting-based LLMs significantly lag behind human experts in handling long-range dependencies and domain-specific calculations.
The paper introduces MultiHiertt, a benchmark for numerical reasoning over multi-hierarchical tabular and textual financial data, and proposes MT2Net, a retrieval-reasoning model that significantly outperforms baselines but still lags far behind human experts.
DocFinQA extends the FinQA dataset by grounding 7,437 financial numerical reasoning questions in full-length SEC filings (averaging 123k words), revealing that even state-of-the-art LLMs and retrieval systems struggle significantly with long-context financial document analysis.
FinanceBench is a benchmark of 10,231 financial QA questions from public filings that reveals state-of-the-art LLMs like GPT-4-Turbo and Claude-2 still fail or hallucinate in over 20% of cases even with retrieval or long-context augmentation.
BizBench introduces a comprehensive benchmark for evaluating LLMs' quantitative reasoning in finance through program synthesis, quantity extraction, and domain knowledge tasks, revealing that current models are bottlenecked by limited financial background knowledge rather than just code generation capabilities.
The paper introduces a cognitive fine-tuning framework using a curated MCQ dataset, structured chain-of-thought reasoning, and robustness augmentation to train open-source LLMs for stable financial decision-making, demonstrating that smaller models can approach frontier performance in bullish and mixed market regimes when evaluated via a two-stage static and sequential trading simulation.
FinChain introduces a symbolic benchmark for verifiable multi-step financial reasoning, revealing that even frontier LLMs struggle with complex chain-of-thought verification despite strong final-answer accuracy.
The WWW'18 Open Challenge established benchmarks for aspect-based financial sentiment analysis and opinion-based question answering, revealing that while sentiment prediction is feasible, aspect classification and opinion QA remain difficult tasks with limited performance.
DianJin-R1 enhances financial reasoning in LLMs through a structured SFT and GRPO reinforcement learning pipeline on a curated dataset, achieving state-of-the-art accuracy on Chinese financial benchmarks while significantly reducing inference costs compared to multi-agent systems.
FAMMA is a multilingual, multimodal benchmark for advanced financial reasoning that reveals significant gaps in LLMs' domain knowledge and demonstrates that fine-tuning on distilled reasoning trajectories improves performance on unseen financial questions.
SEC filings, annual reports, XBRL, accounting, credit & risk, disclosure and document analytics.
SEC-QA introduces a dynamic benchmark for multi-document financial QA using SEC filings, demonstrating that standard RAG fails on complex quantitative tasks while a program-of-thought approach with document selection significantly improves accuracy.
SECQUE is a benchmark of 565 expert-written questions on SEC filings that evaluates LLMs on financial reasoning, with results showing GPT-4o leads but significant gaps remain in complex insight generation.
Fin-RATE introduces a benchmark for evaluating LLMs on SEC filings through detail, cross-entity, and longitudinal tasks, revealing that performance degrades significantly in multi-document reasoning due to hallucinations and entity mismatches.
FinTagging introduces a two-stage benchmark for XBRL tagging that decomposes the task into Financial Numeric Identification and Financial Concept Linking, revealing that while LLMs excel at extraction, they struggle significantly with fine-grained alignment to the full US-GAAP taxonomy.
This paper presents a three-stage pipeline combining LLM extraction, embedding-based mapping, and LLM-as-a-judge validation to extract structured risk factors from 10-K filings into a hierarchical taxonomy, demonstrating autonomous taxonomy refinement and strong industry clustering validity.
This paper demonstrates that LLM agents, by integrating textual disclosures with company-specific financial and operational context, can more accurately assess the materiality of physical climate risks in SEC filings than traditional keyword or embedding methods.
This paper demonstrates that GPT-4, when prompted with standardized financial statements and chain-of-thought reasoning, outperforms human analysts and matches specialized machine learning models in predicting earnings direction, generating actionable alpha in trading strategies.
This study quantifies the automation potential of equity research reports by mapping 4,964 sentences into 169 question archetypes, finding that 75% are automatable via text extraction or database queries, with LLMs successfully automating approximately 80% of statements when used in an ensemble.
This paper introduces FinRpt, a benchmark and multi-agent framework for Equity Research Report generation, demonstrating that fine-tuned agents outperform single LLMs in producing professional, accurate financial reports.
The paper introduces FinAR-Bench, a benchmark evaluating LLMs on financial statement analysis through information extraction, indicator computation, and logical reasoning, revealing that while large models excel at extraction, they struggle with precise numerical calculation despite improved reasoning capabilities.
The paper argues that standard accuracy benchmarks for financial LLM agents are insufficient and proposes a risk-centric auditing framework, SAEA, which reveals that high-performing agents still harbor critical safety vulnerabilities like hallucination and adversarial susceptibility.
TAT-LLM is a specialized 7B-70B LLM fine-tuned on a Step-wise Pipeline (Extractor, Reasoner, Executor) that outperforms GPT-4 and prior SOTA models on financial tabular-textual QA benchmarks like FinQA and TAT-QA.
Fin-R1 is a 7B-parameter financial reasoning LLM trained via supervised fine-tuning and Group Relative Policy Optimization on a distilled 60k-sample chain-of-thought dataset, achieving competitive benchmark performance and practical utility in compliance and robo-advisory tasks.
FinNLI is a novel benchmark dataset for Financial Natural Language Inference comprising 21,304 premise-hypothesis pairs from SEC filings, annual reports, and earnings calls, revealing that current LLMs struggle with financial reasoning and that instruction-tuned financial models often underperform general-domain baselines.
GPT-InvestAR demonstrates that LLM-generated insights from 10-K filings, when used as features in a non-negative linear regression model, can outperform the S&P 500 in annual stock selection with minimal transaction costs.
This paper demonstrates that GPT-3.5 summaries of corporate disclosures amplify sentiment and better explain stock market reactions than original texts, enabling a novel 'bloat' metric that predicts lower price efficiency and higher information asymmetry.
The paper introduces FinLongDocQA, a benchmark for cross-table numerical reasoning in long financial reports, and proposes FinLongDocAgent, a multi-agent RAG system that significantly improves LLM accuracy by iteratively retrieving evidence and verifying calculations.
This study evaluates five major LLMs on 10-K Business sections using human, automated, and behavioral metrics, finding no single model dominates and highlighting significant inter-rater disagreement in qualitative assessments.
The paper proposes a scalable framework using Cohere's Command-R+ LLM to extract, clean, and quantitatively rate SEC 10-K filings across confidence, environment, innovation, and people dimensions, enabling longitudinal and comparative corporate analysis.
EDINET-Bench evaluates LLMs on complex Japanese financial tasks like fraud detection and earnings forecasting, revealing that state-of-the-art models perform only marginally better than logistic regression, highlighting the need for richer reasoning scaffolding.
This study proposes a lightweight, tiered text comparison framework using TF-IDF, BM25, and historical consistency analysis to detect semantic mismatches in XBRL tag mappings for SEC 10-K filings, revealing persistent industry-specific tagging heterogeneity and offering practical validation checkpoints for disclosure workflows.
HiFi-KPI introduces a large-scale dataset of 1.65M iXBRL-tagged paragraphs with unified taxonomies and contextual metadata, demonstrating that encoder-based models excel at classification while LLMs struggle with structured extraction, particularly regarding date normalization.
This study provides a comprehensive assessment of LLM consistency across five finance tasks, finding that while simple tasks are highly reproducible, complex tasks vary, yet aggregation strategies and downstream statistical robustness mitigate these inconsistencies.
FinAuditing introduces a taxonomy-structured benchmark using real XBRL filings to evaluate LLMs on semantic matching, relationship extraction, and mathematical reasoning, revealing significant gaps in cross-document consistency and numerical accuracy.
The paper introduces SEC-former, a bidirectional transformer fine-tuned to reconstruct standardized SEC 10-K item structures from unstructured filings, demonstrating improved document similarity and topic detection for downstream financial analysis.
The paper introduces the Financial Numeric Extreme Labelling (FNXL) dataset with 2,794 XBRL tags and benchmarks sequence labeling versus extreme classification pipelines, finding that while sequence labeling excels on frequent labels, the extreme classification pipeline better handles the long-tail distribution of rare accounting metrics.
This paper introduces a benchmark and five-stage evaluation framework to assess LLMs' capabilities in automating financial statement audits, revealing that while models can detect errors, they struggle with accounting standard citations and table revisions.
Finch introduces FinWorkBench, a benchmark of 172 real-world enterprise finance and accounting workflows sourced from Enron and other institutions, revealing that even top AI agents like GPT-5.1 Pro pass only 38.4% of complex, long-horizon spreadsheet tasks.
This paper benchmarks five LLM variants within the LARD-SC framework for supply chain risk identification, demonstrating that GPT-4o variants achieve superior accuracy and lower false positive rates compared to earlier models when analyzing news about Apple's Tier 1 suppliers.
The paper introduces LUEI, a lightweight unsupervised framework using LLMs and semantic similarity to identify supply chain risk contributing events in news without training data, achieving significantly higher accuracy than keyword-based or commercial alternatives.
SHIELD integrates LLM-driven schema induction with GCN-based prediction to accurately forecast EV battery supply chain disruptions, outperforming baseline models and direct LLM prompting.
Stock prediction, trading, alpha, portfolio construction, allocation and market analysis.
GPT-4 accurately predicts stock market reactions to news headlines with ~90% hit rates and significant drift predictability, particularly for small caps and negative news, though strategy returns decline as LLM adoption increases.
This study demonstrates that GPT-4-32k, prompted with financial fundamentals and sentiment scores, outperforms traditional Wall Street analysts in predicting S&P 500 stock ratings based on forward returns, while revealing that raw news text introduces bias compared to structured sentiment inputs.
The FINSABER framework reveals that LLM-based investing strategies fail to outperform simple benchmarks like Buy-and-Hold over long horizons and broad universes due to survivorship bias in prior evaluations and poor regime-aware risk control.
FinTradeBench is a benchmark evaluating LLM financial reasoning over company fundamentals and trading signals, revealing that while retrieval-augmented generation (RAG) significantly improves fundamental analysis, it degrades performance on trading signal reasoning due to LLMs' inability to parse raw numerical time-series data.
StockGPT is a lightweight decoder-only transformer trained on 70 million daily US stock returns that automatically learns predictive patterns, yielding long-short portfolios with significant alphas against standard asset pricing factors.
The paper proposes a Decision-Informed Neural Network (DINN) that integrates Large Language Model embeddings with differentiable portfolio optimization to align return predictions directly with investment decision quality, outperforming traditional prediction-focused models on S&P100 and DOW30 datasets.
The paper introduces an LLM-based agentic framework for top-down sector allocation that integrates macroeconomic indicators and aspect-based sentiment analysis, achieving a Sharpe ratio of 2.51 and 8.79% return in backtests compared to a baseline cross-momentum strategy.
This paper evaluates ChatGPT's zero-shot capability in multimodal stock movement prediction, finding it underperforms traditional baselines and specialized deep learning models despite the utility of tweet data, while highlighting its limitations in reasoning stability and explainability.
This paper introduces a three-stage experimental framework to quantify latent investment biases in LLMs, revealing that models exhibit strong, persistent confirmation biases toward technology stocks, large-cap firms, and contrarian strategies, which cause them to ignore significant counter-evidence.
This study demonstrates that LLM-based portfolio management strategies, utilizing persona-based ensembles and economic indicators, outperform buy-and-hold in Sharpe ratio during rising CPI periods but lag during market downturns.
This paper presents a two-stage task-adaptive LLM framework that combines granular sentiment extraction with a six-dimensional cognitive reasoning model to generate persuasive investment reports from earnings call transcripts, achieving top rankings in the FinNLP 2025 Earnings2Insights shared task.
InvestLM is a 65B parameter financial LLM instruction-tuned on a curated dataset of ~1,300 high-quality examples, demonstrating expert-rated performance comparable to GPT-3.5/4 and strong generalization across financial NLP benchmarks.
This paper proposes a structured six-decision framework to guide financial institutions in the responsible integration of Large Language Models, balancing innovation with regulatory compliance, data governance, and ethical oversight.
The paper introduces the AI Financial Intelligence Benchmark (AFIB) to evaluate five AI systems across five dimensions, finding that SuperInvesting achieves the highest aggregate performance in factual accuracy and analytical completeness for Indian equity research.
FinLlama fine-tunes Llama 2 7B using LoRA for financial sentiment analysis, achieving superior cumulative returns and Sharpe ratios in long-short portfolio construction compared to FinBERT and lexicon-based methods.
The paper introduces a decoupled training mixture model that adaptively combines quantitative factor predictions with LLM-generated newsflow representations, achieving superior stock return prediction and portfolio performance compared to standard fusion learning, particularly in long-short strategies.
FinVis-GPT is a multimodal large language model fine-tuned on a proprietary dataset of Chinese A-share stock charts that outperforms general-purpose vision-language models in generating descriptions, answering financial questions, and predicting market trends.
This paper empirically demonstrates that off-the-shelf LLMs suffer from severe hallucinations in financial tasks, but accuracy can be significantly restored using Retrieval Augmented Generation (RAG) and prompt-based tool learning for external API calls.
MarketSenseAI leverages GPT-4 with Chain of Thought and In-Context Learning to analyze multi-source financial data for stock selection, achieving up to 72% cumulative returns and 10-30% excess alpha on S&P 100 stocks over 15 months.
The paper demonstrates that anonymizing company names in news headlines mitigates the 'distraction effect' in GPT-3.5 sentiment analysis, leading to superior in-sample trading performance compared to original headlines where look-ahead bias is outweighed by negative interference from the model's pre-existing knowledge.
This paper demonstrates that LLMs, particularly GPT-4 with Chain-of-Thought prompting, can outperform traditional statistical and machine learning baselines in forecasting NASDAQ-100 stock returns while providing human-readable explanations derived from multi-modal financial data.
The paper proposes a hybrid LSTM-GRU model integrated with FinBERT-derived social media sentiment to predict Ethereum and Solana prices, demonstrating superior accuracy over baseline models.
The study demonstrates that the OPT large language model significantly outperforms BERT, FinBERT, and the Loughran-McDonald dictionary in analyzing U.S. financial news sentiment, achieving a Sharpe ratio of 3.05 for a long-short trading strategy based on its predictions.
This paper demonstrates that fine-tuning encoder-only and decoder-only LLMs on financial news to directly predict stock returns yields superior portfolio performance compared to conventional sentiment-based methods, with decoder models like Mistral showing robustness across diverse investment universes.
The paper proposes SCRL-LG, a framework using Self-Correlated Reinforcement Learning to align LLM-generated news embeddings with quantitative stock features, significantly improving stock return prediction accuracy and portfolio performance in the China A-share market.
The paper demonstrates that instruction-fine-tuning Llama-3-8B with 4-bit QLoRA on a dataset combining earnings transcripts, financial metrics, and external market factors significantly outperforms GPT-4 in predicting next-day stock direction following earnings announcements.
Ploutos is a financial LLM framework that combines multimodal expert analysis with a novel training strategy to achieve state-of-the-art stock movement prediction accuracy and high-quality interpretable rationales.
The paper proposes a framework using ChatGPT to infer dynamic stock networks from financial news, which are then processed by a Graph Neural Network and LSTM to predict stock movements, outperforming baselines in both classification accuracy and portfolio performance.
This study demonstrates that incorporating FinBERT-derived sentiment scores from New York Times news summaries significantly improves the accuracy of LSTM-based S&P 500 index forecasts compared to models using only historical price data.
BondBERT, a transformer fine-tuned on bond-specific news, outperforms general financial sentiment models by correctly capturing the inverse relationship between economic optimism and bond prices, yielding higher directional accuracy and lower forecasting error for UK sovereign bonds.
The paper demonstrates that small LLMs (under 1.5B parameters) can outperform larger generalist and specialized financial models in financial sentiment analysis and related tasks through a two-stage adaptation process involving further pre-training on financial documents and instruction fine-tuning with synthetic data augmentation.
Alpha-GPT introduces a human-AI interactive framework for quantitative alpha mining, leveraging LLMs to translate trading ideas into formulaic factors and achieving top-10 global rankings in the WorldQuant International Quant Championship 2024.
This study demonstrates that ChatGPT-generated sentiment scores from corporate annual reports regarding environmental policy effectively predict firms' risk management adjustments and subsequent stock return performance.
LLMFactor introduces Sequential Knowledge-Guided Prompting to extract explainable factors from news and historical prices, achieving superior stock movement prediction accuracy over keyphrase, sentiment, and time-series baselines across US and Chinese markets.
This study compares FinBERT, GPT-4, and Logistic Regression for predicting the NGX All-Share Index using financial news, finding that optimized Logistic Regression outperforms both advanced LLMs in accuracy and efficiency.
The ECC Analyzer framework leverages LLMs to extract fine-grained textual and audio features from earnings conference calls, achieving a 27.7% reduction in MSE for short-term stock volatility prediction compared to state-of-the-art baselines.
The ECC Analyzer framework leverages LLMs to extract fine-grained textual and audio features from earnings conference calls, achieving a 27.7% reduction in Mean Squared Error for stock volatility prediction compared to state-of-the-art baselines.
Trading-R1 is a financial reasoning LLM trained via supervised fine-tuning and reinforcement learning with a three-stage curriculum to generate structured, evidence-based investment theses and volatility-adjusted trading decisions.
This study evaluates GPT-4's ability to predict earnings direction and stock returns in China's A-share market using anonymized financial statements, revealing high confidence but low and inconsistent accuracy with no reliable correlation to actual returns.
FinReasoning introduces a hierarchical benchmark decomposing financial research into semantic consistency, data alignment, and deep insight, revealing that closed-source models outperform open-source and financial-domain models in multi-agent role suitability.
This conceptual paper proposes a research agenda for deploying large language models in financial advice, identifying domain expertise, trustworthiness, and regulatory compliance as critical challenges that must be addressed before LLMs can safely replace or augment human financial advisors.
CN-Buzz2Portfolio introduces a rolling-horizon benchmark mapping Chinese daily trending news to macro and sector ETF allocation, revealing that LLMs generate significant alpha in sector rotation but struggle in sideways markets due to regime-dependent efficacy and narrative hallucination.
FinBERT2, a specialized Chinese financial bidirectional encoder pretrained on 32B tokens, outperforms leading LLMs in classification and retrieval tasks while enabling superior topic modeling, demonstrating that fine-tuned encoders remain superior to LLMs for discriminative and feature-extraction tasks in finance.
The paper presents a reproducible workflow for fine-tuning FinBERT on a new 1,500-headline gold-standard dataset, demonstrating that while the resulting sentiment signal is reactive rather than predictive of returns, it supports profitable sector-specific long-short strategies with high Sharpe ratios.
The paper demonstrates that combining FinBERT-extracted sentiment features with Bayesian-optimized feature selection significantly improves SPY ETF return sign prediction and cumulative trading profits compared to existing benchmarks.
FinRipple aligns LLMs with financial markets via time-varying knowledge graphs and CAPM-guided reinforcement learning to predict event ripple effects, significantly outperforming baselines in asset pricing and portfolio management.
FinCARE integrates SEC 10-K derived knowledge graphs and LLM reasoning into causal discovery algorithms (PC, GES, NOTEARS), achieving significant F1 score improvements and reliable counterfactual scenario analysis for portfolio management.
The paper proposes an LLM-based framework that integrates monetary policy transmission path reasoning and a dynamic uncertainty decoding module to achieve state-of-the-art accuracy and reliability in classifying Federal Reserve policy stances.
PolyBench evaluates seven LLMs on live Polymarket data, revealing that only MiMo-V2-Flash and Gemini-3-Flash achieve positive financial returns, highlighting the gap between language fluency and profitable probabilistic reasoning under market uncertainty.
The paper proposes a dynamic hedging framework for derivatives that integrates LLM-driven sentiment analysis from news and social media to adjust positions in real-time, achieving superior risk-adjusted returns compared to static hedging baselines.
FinSheet-Bench evaluates LLMs on synthetic private equity spreadsheet QA, revealing that even top models like Gemini 3.1 Pro achieve only 82.4% accuracy, with performance collapsing on complex aggregations, indicating a need for hybrid architectures separating parsing from computation.
FinanceQA is a benchmark evaluating LLMs on realistic financial analysis tasks, revealing that state-of-the-art models fail approximately 60% of professional-grade questions due to poor handling of accounting conventions and incomplete information, though fine-tuning significantly improves performance.
MIRAGE-VC introduces an information-gain-driven graph path retriever and multi-agent RAG framework to predict venture capital startup success, achieving state-of-the-art performance by distilling complex investment networks into interpretable reasoning chains.
The provided paper text describes KD-MSLRT, a lightweight sign language recognition model using knowledge distillation, which is unrelated to the cataloged paper 'LLM-Guided Evolutionary Strategy Generation for Quantitative Trading' (LLM-GA) that uses LLMs and genetic algorithms for quantitative trading strategy generation.
MM-DREX is a multimodal LLM framework that uses a vision-language router to dynamically allocate weights among four specialized trading experts, achieving superior risk-adjusted returns across stocks, futures, and crypto compared to static baselines.
AlphaForgeBench introduces a deterministic benchmark for LLMs that evaluates their ability to generate executable alpha factors and strategy code rather than direct trading actions, revealing that this approach eliminates the severe behavioral instability and non-reproducibility inherent in direct-trading agent evaluations.
The paper introduces Option Query Language (OQL), a neuro-symbolic pipeline that translates natural language trading intents into executable option strategies by using LLMs as semantic parsers for a domain-specific intermediate representation, significantly improving execution accuracy and risk management over direct code generation baselines.
This paper presents a due diligence dashboard for venture capital that integrates GPT-4 for document analysis and XGBoost for startup success prediction, achieving a 4.50/5.00 user satisfaction score in a case study.
QuantCode-Bench evaluates LLMs on generating executable Backtrader trading strategies, revealing that while syntax is mastered, semantic alignment and logic operationalization remain significant challenges even for frontier models.
The paper proposes an intelligent trade order recognition pipeline using five LLMs to convert natural language trading instructions into standard JSON formats, revealing high syntactic validity but low end-to-end accuracy and excessive unnecessary clarifications.
The paper proposes an LLM-powered Monte Carlo Tree Search framework for formulaic alpha mining that integrates multi-dimensional backtesting feedback and frequent subtree avoidance to discover interpretable, high-performing alpha factors.
The paper develops a fused large language model combining BERT-generated embeddings of textual self-descriptions with fundamental variables from Crunchbase profiles to predict startup success, achieving a statistically significant improvement in balanced accuracy and a substantial increase in estimated investment ROI compared to baselines using only structured data or traditional text representations.
LOB-Bench introduces a comprehensive benchmark for evaluating generative models of limit order book data, demonstrating that autoregressive state-space models outperform traditional parametric and GAN-based approaches in reproducing realistic market microstructure dynamics.
Financial LLM agents, trading agents, multi-agent markets, agent benchmarks, autonomous workflows.
FinMem introduces an LLM-based trading agent with a layered memory system and dynamic character design that significantly outperforms deep reinforcement learning baselines in stock trading by effectively prioritizing time-sensitive financial data.
TradingAgents introduces a multi-agent LLM framework that simulates professional trading firm dynamics through specialized analyst, researcher, trader, and risk management roles, achieving superior risk-adjusted returns compared to rule-based baselines on major tech stocks.
InvestorBench introduces a comprehensive benchmark and agent framework for evaluating LLM-based financial decision-making across stocks, cryptocurrencies, and ETFs, revealing that proprietary models significantly outperform open-source alternatives in complex, volatile markets.
STOCKBENCH evaluates LLM agents in a contamination-free, multi-month stock trading environment, revealing that while top models like Kimi-K2 can outperform buy-and-hold baselines, most struggle with profitability and risk management compared to simple passive strategies.
FinRobot is an open-source multi-agent LLM framework that automates equity research by integrating quantitative data processing, conceptual reasoning, and narrative synthesis to generate professional-grade investment reports.
FinRobot is an open-source AI agent platform that integrates multi-source LLMs, financial Chain-of-Thought prompting, and a Smart Scheduler to automate complex financial analysis tasks such as equity research and market forecasting.
FinSphere is a real-time stock analysis agent that integrates proprietary quantitative tools with an instruction-tuned LLM, achieving superior expert-rated analytical quality compared to general and domain-specific baselines.
The Finance Agent Benchmark evaluates 20 LLMs on 537 expert-authored financial research tasks requiring tool use, revealing that even the best model (o3) achieves only 46.8% accuracy, highlighting significant gaps in autonomous financial analysis capabilities.
FinAgent is a multimodal LLM-based trading agent that integrates news, prices, and charts via a dual-level reflection mechanism and tool-augmented decision-making, significantly outperforming baselines in profit across stock and crypto markets.
LiveTradeBench introduces a live trading environment for evaluating LLM agents in U.S. stocks and Polymarket, revealing that high general reasoning scores do not guarantee trading competence and that models exhibit distinct, market-specific portfolio styles.
DeepFund introduces a live, multi-agent arena platform to evaluate LLM-based fund investment strategies in real-time, addressing critical flaws in existing benchmarks such as data leakage, theoretical disconnect, and excessive human intervention.
TradingGPT introduces a multi-agent LLM framework with layered memory and distinct character profiles to enhance automated stock and fund trading performance through collaborative debate and hierarchical information processing.
DeepFund introduces a live, multi-agent benchmarking framework for LLM-driven fund investment that eliminates historical data leakage, revealing that most state-of-the-art models incur net losses in real-time trading, with only Grok 3 achieving profitability through prudent risk management.
FinCon is an LLM-based multi-agent system with a hierarchical manager-analyst structure and conceptual verbal reinforcement that outperforms DRL and other LLM agents in single-stock trading and portfolio management by synthesizing multi-modal data and dynamically updating investment beliefs.
FinArena is a human-agent collaboration framework that integrates specialized LLM agents for time series, news, and financial statements with adaptive RAG and investor risk preferences to improve stock trend prediction and personalized investment decisions.
StockAgent is an LLM-driven multi-agent system that simulates stock trading to analyze how external factors like macroeconomic events and social sentiment influence agent behavior and profitability while mitigating test set leakage.
FLAG-Trader integrates a partially fine-tuned LLM with Proximal Policy Optimization to enable a small 135M-parameter model to outperform larger proprietary models in single-asset stock and crypto trading.
The paper proposes a three-stage LLM-driven multi-agent framework for automated alpha factor generation and dynamic weight optimization, achieving 53.17% cumulative return on the SSE50 index by adapting to market regimes.
This 2025 review synthesizes 84 studies on LLM applications in equity markets, categorizing them by financial use cases and technical methods to identify trends in forecasting, sentiment analysis, and multi-agent trading while highlighting gaps in real-world validation and interpretability.
The paper introduces MarS, a financial market simulation engine powered by the Large Market Model (LMM), which generates realistic, interactive, and controllable order-level market trajectories to enable forecasting, risk detection, market impact analysis, and reinforcement learning agent training.
HedgeAgents is a multi-agent system using LLMs to coordinate specialized asset experts and a fund manager through hedging conferences, achieving 400% total return and superior risk management across Bitcoin, stocks, and forex compared to baselines.
This survey systematically categorizes and reviews the applications of large language models in finance, covering linguistic tasks, sentiment analysis, time series forecasting, reasoning, and agent-based modeling, while providing a comprehensive collection of associated datasets, benchmarks, and code to facilitate research and adoption.
The paper proposes a multi-agent collaboration framework for financial investment research, demonstrating that an ensemble structure combining single agents for simple tasks and vertical multi-agents for complex risk analysis outperforms uniform architectures in accuracy and efficiency.
P1GPT introduces a layered multi-agent LLM workflow that fuses technical, fundamental, and news data through structured reasoning to generate interpretable trading signals, demonstrating superior risk-adjusted returns compared to baselines on major U.S. equities.
MASS introduces a multi-agent simulation framework with backward optimization for end-to-end portfolio construction, demonstrating a scaling effect where increasing agent count up to 512 improves excess returns on the Chinese A-share market.
The paper proposes the SEP framework, which uses a self-reflective agent and PPO to fine-tune LLMs for explainable stock prediction and portfolio construction without human annotations.
CryptoTrade is a zero-shot LLM-based multi-agent system that integrates on-chain transaction statistics and off-chain news to guide daily cryptocurrency trading, outperforming time-series baselines but matching traditional technical indicators.
The paper proposes Heterogeneous Agent Discussion (HAD), a zero-shot framework using five specialized LLM agents guided by financial sentiment error types, which improves classification accuracy by 25-35% of the fine-tuning gap compared to naive prompting.
This paper proposes an optimization framework for individual and household budgeting that integrates LLM-generated recommendations as initial feasible solutions, validated through a prospective verification mechanism and extended coevolutionary theory to enhance personalization and financial resilience.
MarketSenseAI 2.0 leverages a multi-agent LLM architecture with RAG to integrate SEC filings, earnings calls, and macroeconomic reports, achieving 125.9% cumulative returns on S&P 100 stocks (2023-2024) and a 33.8% higher Sortino ratio on S&P 500 stocks (2024) compared to benchmarks.
PyFi introduces a pyramid-structured dataset and adversarial multi-agent synthesis framework to enable vision-language models to perform hierarchical financial image understanding, significantly improving accuracy on complex decision-support tasks through chain-of-thought fine-tuning.
TiMi is a rationality-driven multi-agent system that decouples strategy development from minute-level deployment, achieving superior risk-adjusted returns and low latency across 200+ trading pairs by leveraging specialized LLMs for semantic analysis, code generation, and mathematical optimization.
The paper introduces Agent Market Arena (AMA), a live multi-market benchmark demonstrating that agent architecture, rather than LLM backbone, primarily determines trading performance and behavioral patterns in real-time financial markets.
InvestorBench introduces a comprehensive benchmark and agent framework for evaluating LLM-based financial decision-making across stocks, cryptocurrencies, and ETFs, demonstrating that proprietary and large open-source models significantly outperform smaller or domain-specific models in sequential trading tasks.
This paper introduces an open-source simulation framework where heterogeneous LLM agents act as trading participants in a realistic market with a persistent order book, demonstrating that LLMs can execute specific strategies, generate emergent market dynamics like bubbles, and serve as a tool for experimental finance.
QuantAgent introduces a multi-agent LLM framework for high-frequency trading that decomposes OHLC data analysis into specialized indicator, pattern, trend, and risk agents, achieving superior directional accuracy and risk-adjusted returns across diverse asset classes compared to traditional baselines.
AlphaQuanter introduces a single-agent reinforcement learning framework that optimizes tool-augmented decision-making for stock trading, achieving state-of-the-art performance and higher faithfulness than multi-agent or prompt-only baselines.
The paper introduces the Agent Trading Arena, a closed-loop multi-agent simulation where LLMs trade in a zero-sum environment, demonstrating that visual inputs and reflection modules significantly improve numerical reasoning and trading performance over textual data and static baselines.
ATLAS introduces a multi-agent trading framework with Adaptive-OPRO, a dynamic prompt optimization method that consistently outperforms static prompts and reflection-based feedback across diverse market regimes and LLM families.
This paper introduces a role-based multi-agent LLM framework that generates investment reports from earnings call transcripts, achieving 58.1% decision accuracy and outperforming professional analyst reports in human evaluations by prioritizing logical structure and practical utility over surface-level quality.
The paper develops an agentic framework where LLMs iteratively discover stochastic differential equations to estimate market risk, demonstrating that model-informed trading decisions significantly outperform standard sentiment-based LLM agents in both historical backtests and synthetic market simulations.
TradeTrap demonstrates that LLM-based trading agents are highly vulnerable to system-level perturbations, where small attacks on market intelligence, strategy, or state management can cause catastrophic portfolio drawdowns and unreliable behavior.
This paper proposes a multi-agent LLM framework for Chinese Public REITs that decomposes analysis into four specialized agents and compares a general-purpose LLM (DeepSeek-R1) against a fine-tuned small model (Qwen3-8B), finding both significantly outperform buy-and-hold with superior risk-adjusted returns.
This paper demonstrates that decomposing investment analysis into fine-grained, expert-defined tasks within a hierarchical multi-agent LLM system significantly improves risk-adjusted returns and interpretability compared to coarse-grained baselines in Japanese equity backtesting.
TradingGroup is a multi-agent quantitative trading system that integrates specialized agents for sentiment, fundamentals, and forecasting with self-reflection mechanisms and a dynamic risk management module, achieving superior backtesting performance over baselines and enabling effective parameter-efficient fine-tuning via an automated data-synthesis pipeline.
The deployed MarketSenseAI multi-agent LLM system generates strong-buy recommendations that significantly outperform passive equal-weight benchmarks and random selection on the S&P 500, with performance driven by an adaptive synthesis of specialist agent signals rather than a single dominant factor.
StockSim is an open-source, dual-mode simulation platform that evaluates LLM trading agents using realistic order-book microstructure and candlestick data, revealing distinct strategic behaviors between models like GPT-o3 and GPT-o4-mini.
This paper presents an automated multi-agent framework for venture capital due diligence that integrates real-time web retrieval, reverse-engineered access to the Greek Business Registry for official financial filings, and a structural fallback mechanism to prevent hallucination by explicitly flagging data gaps.
The paper introduces the Agent Trading Arena, a zero-sum multi-agent simulation framework that demonstrates LLM-based agents significantly outperform traditional baselines when using visual chart inputs and a reflection module to handle dynamic market microstructure.
PolySwarm is a multi-agent LLM framework that deploys 50 diverse personas to trade prediction markets via Bayesian aggregation and KL-divergence-based arbitrage, outperforming single-model baselines in calibration.
DeltaHedge is a multi-agent framework integrating transformer-based forecasting, sentiment analysis, and an ensemble of reinforcement learning agents to dynamically hedge equity portfolios with protective puts, significantly outperforming traditional and standalone RL baselines in risk-adjusted returns.
AlphaAgent is an autonomous multi-agent framework that uses LLMs to mine decay-resistant alpha factors by enforcing originality via AST similarity, hypothesis alignment, and complexity control, achieving superior risk-adjusted returns in CSI 500 and S&P 500 markets.
JAX-LOB is the first GPU-accelerated limit order book simulator that enables large-scale parallel reinforcement learning for optimal trade execution, achieving at least a 7x speedup over CPU-based implementations.
The paper proposes FAMA, a neural-symbolic agent that uses LLMs with Cross-Sample Selection and Chain-of-Experience to mine interpretable alpha factors, achieving superior predictive performance and portfolio returns on S&P 500 data compared to symbolic and neural baselines.
ABIDES is an open-source, high-fidelity agent-based discrete event simulation environment for equity markets that enables controlled experiments on market impact, co-location, and learning agents by replicating NASDAQ protocols with nanosecond precision.
SimVC-CAS predicts startup financing success by simulating collective venture capital decision-making through LLM-based investor agents interacting over a co-investment network, achieving a 25% relative improvement in AP@10 over baselines.
QuantaAlpha is an evolutionary multi-agent framework for alpha mining that improves factor discovery through trajectory-level mutation and crossover, achieving superior predictive power and cross-market robustness compared to existing LLM-based agents.
FactorEngine introduces a program-level factor mining framework that combines LLM-guided macro-evolution with Bayesian micro-optimization to discover executable, interpretable alpha factors from OHLCV data and financial reports, achieving state-of-the-art predictive and portfolio performance on Chinese A-share markets.
TradeFM is a 524M-parameter generative Transformer that learns universal market microstructure dynamics from billions of trade events across >9K equities, achieving 2-3x lower distributional error than Hawkes baselines and zero-shot generalization to APAC markets.
This conceptual paper introduces Perceived Cognitive Assistance (PCA) to explain how LLMs alter retail investors' strategy selection and risk posture, proposing a moderated regression model and a Behavioral Shift Index (BSI) for future empirical validation.
The CogAlpha framework leverages a seven-level agent hierarchy and LLM-driven evolutionary search to discover interpretable, robust alpha factors that outperform traditional machine learning and existing LLM-based baselines across multiple global stock markets.
AlphaCrafter is a full-stack multi-agent framework that unifies LLM-driven factor discovery, regime-adaptive selection, and risk-constrained execution, demonstrating robust live trading performance on CSI 300 and S&P 500 indices by outperforming static and decoupled baselines.
CFA and professional exams, financial advice, regulatory interpretation, compliance, advisory settings.
This paper benchmarks 14 LLMs on CFA exams, finding that proprietary models pass Levels I and II but fail Level III essays, while open-source models like LLaMA 3 70B can pass Levels I and II with RAG assistance.
This paper benchmarks 23 LLMs on CFA Level III exams, showing that frontier reasoning models like o4-mini pass the mock exam with 79.1% accuracy, though significant cost-latency trade-offs and human-LLM grading discrepancies persist.
This study benchmarks GPT-4o, GPT-o1, and o3-mini on 1,560 CFA mock exam questions, demonstrating that reasoning-specialized models like GPT-o1 excel in financial analysis and that a novel domain-specific RAG pipeline significantly improves accuracy, particularly for complex, knowledge-intensive tasks.
This study demonstrates that GPT-4, when paired with image-based document loading and structured prompt engineering, can accurately interpret Basel III regulations and calculate minimum capital requirements, significantly outperforming other LLMs in both information extraction and complex mathematical computation.
This study evaluates ChatGPT and GPT-4 on mock CFA Level I and II exams, finding that GPT-4 with few-shot prompting can likely pass Level I and potentially Level II, while revealing significant limitations in calculation and complex reasoning.
This paper implements and evaluates agentic crews using the CrewAI framework to automate financial modeling and model risk management (MRM) workflows, demonstrating robust performance in credit card fraud detection, approval, and portfolio risk tasks with human-in-the-loop oversight.
This study evaluates ChatGPT and Bard on 13 personal finance queries involving credit cards, bank accounts, and certificates of deposit, revealing that while outputs are fluent, both models suffer from significant mathematical errors, perceptual mistakes, and a lack of personalized, accurate financial advice.
A user study of 64 participants reveals that while LLM-advisors can match human experts in eliciting investment preferences, they often fail to resolve conflicting needs, and users exhibit a dangerous insensitivity to advice quality, preferring extroverted personas that provide worse financial guidance.
This study demonstrates that LLMs (ChatGPT, Gemini, Copilot) systematically increase portfolio risks for private investors across five dimensions—geographical, sector, trend chasing, active allocation, and expense risks—due to training data biases, with debiasing interventions only partially mitigating these effects.
This qualitative study re-evaluates ChatGPT-4o and ChatGPT-5 against 21 personal finance scenarios, finding that while prompt engineering and model upgrades improve tone and numerical accuracy, the models still struggle with legal nuances, tax implications, and genuine empathy, positioning them as supportive rather than replacement tools for human advisors.
DeepFinLLM is an intelligent financial advisor that integrates the DeepSeek Chat V3 model with real-time data from the Financial Modeling Prep API to provide accurate, low-latency financial insights, achieving 94.2% accuracy and outperforming baselines like FinGPT and BloombergGPT on standard financial benchmarks.
Retrieval-augmented generation, search, knowledge grounding, knowledge graphs, document retrieval.
FinSearchComp is an open-source benchmark of 635 expert-annotated questions evaluating LLM agents' ability to perform realistic financial search and reasoning across time-sensitive, historical, and complex investigative tasks.
The paper introduces FinDER, a benchmark dataset of 5,703 expert-annotated query-evidence-answer triplets derived from S&P 500 10-K filings to evaluate Retrieval-Augmented Generation (RAG) systems on ambiguous, real-world financial queries, demonstrating that neural retrievers and LLM-based reranking significantly outperform traditional methods but that factual correctness remains challenging without high-quality context.
The paper demonstrates that structurally aware chunking of SEC financial reports using the Chipper document understanding model significantly improves Retrieval Augmented Generation (RAG) question-answering accuracy and retrieval consistency compared to fixed-size token chunking.
AlphaFin introduces a comprehensive dataset and the Stock-Chain framework, which combines LoRA-finetuned LLMs with Retrieval-Augmented Generation to achieve state-of-the-art stock trend prediction accuracy and a 30.8% annualized return while providing interpretable financial analysis.
FinBloom introduces a knowledge-grounding framework using a fine-tuned 7B parameter LLM as a Financial Agent to retrieve real-time structured and textual data, enabling accurate, low-latency responses to dynamic financial queries without relying on traditional dense-vector RAG.
The paper introduces a retrieval-augmented instruction-tuned LLM framework for financial sentiment analysis that significantly outperforms traditional models and general-purpose LLMs by combining targeted fine-tuning with external context retrieval.
RA-CFGPT introduces a retrieval-augmented Chinese financial assistant that fine-tunes a large language model on a hybrid knowledge base to improve accuracy, compliance, and risk assessment in financial question answering and document analysis.
The paper introduces FLEX, a retrieval-augmented framework that enriches financial sentences with explicit knowledge using LLMs, significantly improving financial sentiment analysis accuracy across zero-shot, few-shot, and fine-tuning settings.
OmniEval introduces an omnidirectional, automatic RAG benchmark for finance using a matrix-based topic-task evaluation framework, multi-agent data generation, and fine-tuned LLM evaluators to reveal significant performance gaps in vertical domain RAG systems.
MultiFinRAG is a multimodal RAG framework that extracts structured data from financial PDFs and uses tiered retrieval to achieve 75.3% accuracy on complex QA tasks, outperforming ChatGPT-4o by 19 percentage points.
The paper proposes HiREC, a hierarchical retrieval and evidence curation framework that improves open-domain financial question answering on standardized SEC filings by reducing near-duplicate confusion and iteratively filling information gaps, evaluated on the new LOFin benchmark.
This paper proposes a metadata-driven multi-stage RAG architecture that leverages LLM-generated metadata for pre-retrieval filtering, contextual chunk embedding, and custom reranking, achieving superior financial question-answering performance on FinanceBench compared to baseline and advanced RAG configurations.
The paper introduces a three-phase RAG pipeline for financial question answering that combines query expansion, markdown restructuring, fine-tuned hybrid retrieval, and DPO-optimized generation, achieving significant NDCG@10 improvements across seven financial benchmarks.
The paper introduces a domain-fine-tuned page scorer for financial RAG that significantly improves page and chunk retrieval accuracy by treating pages as intermediate retrieval units, closing the gap between document discovery and precise context localization.
Fin-RAG is a multimodal Retrieval-Augmented Generation system for financial documents that combines GPT-4, BERT embeddings, and CoBERT re-ranking to achieve high accuracy in natural language querying of text and image-based financial reports.
This study evaluates Retrieval-Augmented Generation (RAG) systems for answering questions about bank financial reports, finding that high-quality embedding and generation models significantly improve context relevance and answer faithfulness, particularly for qualitative inquiries.
FinVault introduces the first execution-grounded security benchmark for financial agents, revealing that state-of-the-art LLMs remain highly vulnerable to semantic and regulatory attacks in sandboxed environments with state-writable databases.
FinKario introduces an automated, event-enhanced financial knowledge graph and a two-stage graph-based RAG retrieval strategy that significantly outperforms financial LLMs and institutional strategies in stock trend prediction and backtesting accuracy.
The paper introduces ICKG, a fine-tuned LLM for generating dynamic knowledge graphs from financial news, and KGTransformer, an attention-based GNN that leverages meta-entities to outperform baselines in link prediction and thematic AI investing.
The paper introduces NatureKG, a domain-specific knowledge graph for nature finance, and demonstrates the feasibility of using fine-tuned small LLMs like Phi-3 to translate natural language into Cypher queries for structured environmental risk assessment.
FinReflectKG introduces an agentic, reflection-driven framework for constructing large-scale financial knowledge graphs from S&P 100 SEC 10-K filings, achieving superior extraction quality and schema compliance compared to single-pass and multi-pass baselines.
FinVision is a multi-agent LLM framework that integrates news, visual technical analysis, and historical reflection to generate explainable trading decisions with position sizing, outperforming traditional and RL baselines on major tech stocks.
AlphaPROBE introduces a DAG-based framework for alpha mining that uses a Bayesian Factor Retriever and a DAG-aware LLM generator to discover diverse, high-performing factors on Chinese stock markets.
Every paper is assigned to exactly one primary category; the taxonomy is a strict partition, so category counts always equal the full corpus. Trading & investment papers additionally carry a high-precision trading subtheme tag.
277 papers carry structured deep summaries — research problem, core contributions, method, results, data & experiments, limitations and tag facets — generated from full text and reviewed under a controlled schema.
The list lives on GitHub alongside the curation pipeline. Pull requests welcome — see CONTRIBUTING.md for the tagging rules.