Large Language Models for Financial Applications: A Comprehensive Survey of Progress, Prospects, and Challenges

Why Large Language Models Are Reshaping Finance

The financial services industry generates an extraordinary volume of unstructured data every single day. From SEC filings and earnings call transcripts to analyst reports, social media commentary, and central bank policy communications, the sheer breadth of textual information that influences market movements has long exceeded human processing capacity. Large language models (LLMs) have emerged as the most promising technology for bridging this gap, offering the ability to analyze, synthesize, and reason over financial text at unprecedented scale and speed.

A landmark survey published by researchers from multiple institutions provides the most comprehensive mapping to date of how LLMs are being applied across the financial landscape. The paper catalogues more than 50 domain-specific models, hundreds of datasets, and dozens of benchmark tasks—painting a picture of a field that has evolved from experimental curiosity to production-grade infrastructure in barely three years. For anyone working at the intersection of artificial intelligence and finance, this survey represents essential reading.

What makes LLMs particularly compelling for financial applications is their combination of contextual language understanding, emergent reasoning capabilities through chain-of-thought prompting, and the ability to perform useful work with minimal task-specific training through zero-shot and few-shot learning. These properties make them uniquely suited to a domain where language is nuanced, context-dependent, and laden with implicit meaning that traditional NLP systems struggled to capture. If you’re interested in how AI is transforming document analysis more broadly, explore our interactive library of AI research for deeper insights.

The Evolution of Financial NLP: From Rules to Deep Learning

The journey from early financial text processing to today’s LLM-powered systems spans several decades of incremental progress. In the early 2000s, financial NLP relied on hand-crafted dictionaries and rule-based systems. The Loughran-McDonald sentiment dictionary, developed specifically for financial text, became a standard tool for quantifying the tone of 10-K filings and earnings releases. These approaches were interpretable and reliable within narrow domains but fundamentally limited in their ability to handle ambiguity, context, and the evolving nature of financial language.

The arrival of word embeddings through Word2Vec and GloVe in the mid-2010s introduced the notion that semantic relationships could be captured in dense vector representations. Financial applications quickly adopted these techniques for tasks like company similarity scoring and sector classification. However, these models still treated words in isolation, lacking the contextual awareness needed to distinguish between “Apple reported strong growth” (referring to the company) and “apple growth was strong this season” (referring to agriculture).

The transformer architecture, introduced in the seminal “Attention Is All You Need” paper by Vaswani et al., fundamentally changed the landscape. BERT and its variants demonstrated that pre-training on large corpora followed by fine-tuning on specific tasks could achieve state-of-the-art performance across virtually all NLP benchmarks. For finance, this led to the development of domain-adapted models like FinBERT, which was pre-trained on financial communications and quickly became the go-to model for financial sentiment analysis. The evolution continued with larger decoder-only models—GPT-3, GPT-4, LLaMA, and BLOOM—which shifted the paradigm toward generative capabilities, enabling not just classification but open-ended analysis, summarization, and reasoning over complex financial documents.

Domain-Specific Financial LLMs: BloombergGPT, FinBERT, and Beyond

One of the most significant developments in the field has been the creation of LLMs specifically designed for financial applications. Rather than relying solely on general-purpose models that may lack deep financial knowledge, researchers and institutions have invested heavily in domain adaptation—producing models whose pre-training data, instruction tuning, and evaluation are explicitly financial in nature.

BloombergGPT stands as perhaps the most ambitious example. With 50 billion parameters and training on Bloomberg’s proprietary financial data corpus alongside general web data, BloombergGPT was designed to excel at financial tasks while retaining general language capabilities. The model demonstrated strong performance on financial NLP benchmarks including sentiment analysis, named entity recognition, and financial question answering, validating the hypothesis that domain-specific pre-training delivers measurable advantages over prompting general models.

The FinBERT family represents a different approach—smaller, BERT-based models fine-tuned on financial text that can be deployed efficiently for specific classification and extraction tasks. Multiple FinBERT variants exist (developed between 2019 and 2021), each targeting slightly different aspects of financial language processing. These models remain widely used in production systems where inference speed and cost are critical constraints.

The survey catalogs an impressive ecosystem of newer finance LLMs built on open-source foundations: FinMA and Fin-Llama (LLaMA-based), FinGPT (open-source financial model), FinTral (built on Mistral 7B), DISC-FinLLM (Baichuan-13B backbone), and XuanYuan 2.0 (a Chinese financial chat model). Each represents a different trade-off between model size, training cost, specialization depth, and deployment flexibility. The trend is clearly toward smaller, more efficient models that can be fine-tuned with techniques like Low-Rank Adaptation (LoRA) and quantization, making domain-specific AI accessible to organizations without massive compute budgets.

Sentiment Analysis: How LLMs Decode Market Signals

Sentiment analysis has historically been the primary entry point for NLP in finance, and LLMs have dramatically expanded what’s possible. Traditional approaches relied on dictionary-based scoring or supervised classification with limited context windows. Modern LLM-based sentiment analysis can process entire earnings call transcripts, cross-reference multiple news sources, and detect subtle shifts in tone that carry meaningful market signals.

The survey highlights research across five distinct data sources for financial sentiment: social media (Twitter, StockTwits, Reddit), news articles, corporate disclosures (10-K, 10-Q filings), research reports, and policy documents (FOMC minutes, ECB communications). Each source presents unique challenges—social media is noisy and informal, while regulatory filings are dense and formulaic—yet LLMs have shown remarkable adaptability across all categories.

Several studies cited in the survey report that GPT-3.5 and GPT-4 outperform FinBERT on headline and forex-news sentiment classification tasks. In one particularly striking example, GPT-4 was used for microblog sentiment analysis to generate trading signals for Apple and Tesla stocks, beating a buy-and-hold strategy on same-day movements. Another study found that LLM-generated sentiment scores from earnings calls correlated strongly with subsequent stock price movements, suggesting these models capture information that traditional quantitative approaches miss.

However, the research also reveals important limitations. ChatGPT sometimes underperforms domain-crafted models for multimodal and zero-shot stock movement prediction tasks, indicating that larger and more general does not always mean better. The most effective approaches often combine LLM-generated sentiment signals with traditional quantitative features, creating hybrid systems that leverage the best of both paradigms. For more on how AI is reshaping analytical workflows, check out our collection of AI and technology research summaries.

Financial Time Series Forecasting with Language Models

Perhaps the most counterintuitive application of LLMs in finance is their use for time series forecasting—a domain traditionally dominated by statistical and deep learning models operating on numerical data. The key insight driving this research is that financial time series don’t exist in isolation; they are influenced by narratives, events, policy decisions, and sentiment shifts that are inherently textual in nature.

The survey identifies three primary use cases for LLMs in financial time series: forecasting (predicting future price movements or volatility), anomaly detection (identifying unusual patterns that may indicate fraud or market manipulation), and auxiliary tasks (data imputation, augmentation, and synthetic data generation). In the forecasting category, the most promising approaches use LLMs not as direct predictors but as feature extractors—processing news, social media, and filings to generate sentiment and event features that are then fed into quantitative forecasting models.

Multimodal approaches represent the cutting edge of this research. Models like Ploutos combine GPT-4-generated insights with expert pools for interpretable stock movement prediction, processing both textual and numerical inputs through specialized architectures. The PloutosGen component generates financial analyses from text, while PloutosGPT integrates these with market data for prediction. Similarly, FinVIS-GPT extends the paradigm to chart and visual analysis, enabling LLMs to “read” financial charts and extract trading signals from visual patterns.

The challenge of signal decay is particularly acute in time series applications. Financial signals are inherently temporal—yesterday’s news is already priced in. Models need continual updating to remain relevant, and the survey identifies continual learning and timely model updates as critical open challenges. Static models trained on historical data inevitably degrade as market dynamics evolve, making this one of the most active areas of ongoing research.

LLM-Powered Financial Reasoning and Decision Support

Beyond pattern recognition and classification, the most transformative potential of LLMs in finance lies in their reasoning capabilities. The survey categorizes financial reasoning applications into four areas: planning, recommendation and advisory, decision support (auditing, compliance, fraud detection), and real-time reasoning (chatbots, question answering systems).

One of the most cited findings in the survey is a study where a GPT-4 Turbo-based workflow produced earnings prediction performance surpassing human financial analysts, with higher Sharpe ratios and alphas in backtested trading strategies. This result generated significant attention because it suggested that LLMs could not only match but exceed human expert judgment in specific, well-defined prediction tasks. However, the researchers caution that these results apply to narrow, structured prediction tasks and should not be extrapolated to the full scope of financial analysis.

Financial question answering represents another area of rapid advancement. Benchmarks like FinQA, ConvFinQA, and TATQA test models’ ability to answer complex numerical questions about financial statements—tasks that require reading tables, performing multi-step calculations, and synthesizing information across multiple data points. The survey notes that combining LLMs with code generation (program-of-thought approaches) significantly improves accuracy on these benchmarks, as models can generate Python code to perform calculations rather than attempting mental arithmetic, which remains a weakness of transformer architectures.

For compliance and regulatory applications, LLMs offer the ability to process vast volumes of regulatory text and flag potential violations or risks. The SEC’s EDGAR database alone contains millions of filings, and manually reviewing these for compliance issues is both costly and error-prone. LLM-based systems can automate initial screening, generate risk summaries, and alert compliance officers to potential issues requiring human review.

Agent-Based Financial Modeling: Autonomous AI Traders

Perhaps the most forward-looking section of the survey examines LLM-based autonomous agents in financial markets. These systems go beyond simple prediction or analysis—they represent AI entities capable of making independent trading decisions, managing portfolios, and interacting with simulated or real market environments. The implications for market structure, regulation, and investment management are profound.

The survey describes multi-agent systems where LLM-powered agents simulate market participants with different strategies, risk tolerances, and information sets. These simulations can model emergent market behaviors—bubbles, crashes, liquidity crises—that arise from the interactions of heterogeneous agents rather than from any single model’s predictions. For researchers studying market microstructure and systemic risk, these tools offer a new computational laboratory for testing hypotheses that would be impossible to study in live markets.

In practical trading applications, LLM agents are being designed to autonomously process news feeds, analyze technical indicators, reason about market conditions through chain-of-thought prompting, and execute trades based on synthesized intelligence. Some implementations use retrieval-augmented generation (RAG) to ground agent decisions in real-time market data, reducing the hallucination risk that plagues purely generative approaches. The combination of reasoning capabilities with tool use—where agents can call APIs, execute code, and interact with databases—represents a step toward truly autonomous financial intelligence.

However, the survey raises important cautions about agent-based approaches. Questions of liability, market manipulation risk, and the potential for correlated agent behavior to amplify market volatility remain largely unresolved. Regulatory frameworks have not yet caught up with the possibility of autonomous AI participants in financial markets, creating a gray area that will need to be addressed as these technologies mature.

Key Challenges: Hallucinations, Bias, and Regulatory Risks

For all their promise, deploying LLMs in financial applications comes with a distinctive set of challenges that the survey documents extensively. Understanding these risks is essential for any organization considering AI-driven financial tools, and the survey’s honest assessment of limitations adds credibility to its otherwise optimistic outlook.

Hallucination and factual accuracy tops the list of concerns. In a domain where a single incorrect number can trigger erroneous trades worth millions, the tendency of generative models to produce plausible but fabricated information is a serious liability. The survey documents cases where LLMs confidently cite non-existent financial data points, fabricate company names, or generate plausible-sounding but incorrect financial calculations. Mitigation strategies include RAG-based grounding, code-generation for numerical tasks, and multi-step verification pipelines, but no current approach completely eliminates the risk.

Numerical reasoning limitations remain a fundamental weakness. Despite improvements in chain-of-thought prompting and program-of-thought approaches, LLMs still struggle with multi-step financial calculations, percentage computations, and the kind of precise arithmetic that financial applications demand. The survey recommends hybrid architectures where LLMs handle language understanding while delegating computation to deterministic code execution environments.

Data contamination and lookahead bias present methodological challenges that can invalidate research results. When pre-training data includes information from the future relative to a backtest period, models appear to perform well but are actually “cheating” by accessing information that wouldn’t have been available in real time. The survey calls for stricter dataset curation practices and temporal awareness in model evaluation.

On the regulatory and ethical front, the challenges are equally formidable. Bias amplification—where models trained on historical data perpetuate existing market or demographic biases—raises fairness concerns. Privacy risks from processing confidential client data through cloud-based LLMs conflict with financial regulations. And the interpretability deficit of black-box neural networks clashes with regulatory requirements for explainable financial decisions. The survey emphasizes that solving these challenges requires collaboration between AI researchers, financial practitioners, regulators, and ethicists—no single community can address them in isolation.

Future Directions: Multimodal Models, Knowledge Graphs, and Beyond

The survey concludes with a forward-looking assessment of where financial LLM research is headed, and the opportunities are as substantial as the challenges. Several converging trends point toward a future where AI systems provide genuinely comprehensive financial intelligence.

Multimodal financial models represent the most ambitious near-term opportunity. Current LLMs primarily process text, but financial decision-making draws on tables, charts, time series data, audio (from earnings calls), and even satellite imagery. Building models that can jointly process and reason across all these modalities—understanding that a downward trend in a revenue chart contradicts the optimistic language in an earnings transcript, for example—would represent a qualitative leap in financial AI capability. Early work like FinVIS-GPT and DocLLM shows promise but remains far from the comprehensive multimodal vision.

Dynamic knowledge graphs offer a path toward more robust and up-to-date financial reasoning. Projects like FinDKG aim to construct financial knowledge graphs that evolve with market conditions, capturing relationships between companies, executives, products, regulations, and economic indicators. When combined with LLMs through graph-based retrieval, these systems can provide contextually grounded answers that reflect current market reality rather than the static knowledge captured during pre-training.

Efficient adaptation techniques will continue to democratize access to financial AI. As LoRA, quantization, distillation, and other parameter-efficient methods mature, the cost of creating high-quality domain-specific models will decline. This democratization could enable smaller financial institutions—regional banks, boutique advisory firms, fintech startups—to deploy competitive AI tools that were previously accessible only to firms with massive compute budgets.

Finally, the survey highlights the urgent need for standardized financial benchmarks that capture realistic decision-making scenarios. Current benchmarks often test isolated capabilities (sentiment classification, question answering) rather than end-to-end financial workflows. Developing benchmarks that evaluate models on realistic trading scenarios, compliance reviews, and advisory interactions—with proper temporal controls to prevent data leakage—would accelerate progress by giving the research community a shared yardstick for measuring advancement. For those interested in exploring more cutting-edge research across AI and technology domains, our interactive library provides an engaging way to dive deeper into these topics.