A Comprehensive Review of Generative AI in Finance

In one sentence

Summary

This paper addresses the need for a structured overview of the rapidly evolving intersection between generative AI (GAI) and the financial sector. Unlike prior surveys that focus narrowly on large language models (LLMs) or rely on manual topic classification, this study employs BERTopic, an advanced unsupervised topic modeling technique, to systematically categorize and analyze 90 key papers published between 2018 and 2024. The research aims to uncover predominant themes, emerging areas of interest, and the broader implications of GAI technologies, including variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models, beyond just LLMs. By leveraging BERTopic's ability to capture semantic relationships and context, the authors provide an objective, data-driven taxonomy of the current landscape, addressing gaps in existing literature regarding the comprehensive integration of various GAI models in finance. The methodology involves retrieving 90 papers from Google Scholar using keywords "generative AI and finance" and "large language models and finance." The dataset includes interdisciplinary research from finance, computer science, economics, and business, sourced from major publishers and preprint servers. The authors apply a three-step BERTopic pipeline: generating document embeddings using Sentence-BERT, reducing dimensionality with UMAP, and clustering with HDBSCAN to identify coherent topics. The resulting clusters are analyzed using class-based TF-IDF to extract representative keywords. The study compares BERTopic's performance against traditional methods like LDA, highlighting its superior ability to handle semantic nuances. The experimental design focuses on thematic clustering rather than empirical performance testing of specific models, aiming to map the intellectual structure of the field. The analysis reveals three primary clusters: "LLMs for Financial Tasks" (47 papers), "The Risk and Challenge of Generative AI" (20 papers), and "Synthetic Financial Data Generation" (12 papers). The findings indicate that LLMs dominate current research, with applications ranging from sentiment analysis and robo-advisory to complex reasoning and multimodal chart interpretation. Finance-specific models like FinGPT and BloombergGPT are highlighted for their superior performance over general-purpose models in domain-specific tasks. However, the review also emphasizes significant challenges, including hallucinations, ethical concerns, regulatory compliance, and data privacy. Synthetic data generation via GANs is identified as a critical area for overcoming data scarcity and privacy issues in risk modeling and market simulation. The paper concludes that while GAI offers transformative potential, robust regulatory frameworks and ethical guidelines are urgently needed to govern its deployment in finance.

Research problem

The financial sector is experiencing rapid integration of generative AI (GAI), yet existing literature lacks a comprehensive, systematic overview that covers the full spectrum of GAI models (not just LLMs) and their diverse applications. Prior surveys often rely on manual topic selection or focus narrowly on specific models, missing emerging trends and the broader ecosystem of GAI in finance. This paper addresses the need for an objective, data-driven taxonomy to categorize and analyze the current state of research, identify predominant themes, and highlight risks and opportunities.

Core contributions

• Proposes a novel framework for categorizing GAI in finance using BERTopic, an advanced unsupervised topic modeling technique.
• Systematically analyzes 90 key papers from 2018-2024 to uncover three main themes: LLM applications, risks/challenges, and synthetic data generation.
• Demonstrates the superiority of BERTopic over traditional topic modeling methods (LDA, NMF) in capturing semantic nuances and context in financial literature.
• Provides a comprehensive overview of finance-specific LLMs (e.g., FinGPT, BloombergGPT) and their performance compared to general-purpose models.
• Highlights the critical role of synthetic data generation using GANs and VAEs for risk management, fraud detection, and market simulation.
• Identifies and discusses pressing regulatory, ethical, and technical challenges associated with deploying GAI in the financial sector.

Method

• Data Retrieval: Searched Google Scholar for "generative AI and finance" and "large language models and finance".
• Preprocessing: Lowercased text, tokenized using nltk, lemmatized using WordNetLemmatizer, and removed stopwords.
• Embedding: Generated document embeddings using Sentence-BERT (SBERT).
• Dimensionality Reduction: Applied UMAP to reduce the dimensionality of embeddings while preserving local and global features.
• Clustering: Used HDBSCAN to cluster the reduced embeddings, allowing for outlier detection.
• Topic Modeling: Employed class-based TF-IDF (c-TF-IDF) to model topic representations for each cluster.
• Comparison: Evaluated BERTopic against LDA and NMF using topic coherence and diversity metrics.
• Analysis: Interpreted clusters by examining representative keywords and their frequencies.

Results & findings

• BERTopic identified three main clusters: LLMs for financial tasks (47 papers), risks/challenges (20 papers), and synthetic data generation (12 papers).
• LLMs dominate the research landscape, with significant focus on sentiment analysis, robo-advisory, and complex reasoning tasks.
• Finance-specific LLMs like FinGPT and BloombergGPT show superior performance in domain-specific tasks compared to general-purpose models.
• Synthetic data generation using GANs is crucial for overcoming data scarcity and privacy concerns in financial modeling.
• Key challenges include hallucinations, ethical implications, regulatory compliance, and data privacy.
• Multimodal LLMs are emerging as a promising area for integrating textual, numerical, and visual financial data.
• The study highlights the need for robust regulatory frameworks to govern the use of GAI in finance.
• BERTopic outperformed LDA and NMF in capturing semantic relationships and context within the literature.

Data & experiments

• Dataset: 90 academic papers retrieved from Google Scholar using keywords "generative AI and finance" and "large language models and finance".
• Timeframe: Papers published between 2018 and 2024.
• Sources: Diverse publishers including ACM, ACL, arXiv, Elsevier, IEEE, MDPI, and preprint servers.
• Methodology: Application of BERTopic for topic modeling, involving Sentence-BERT for embeddings, UMAP for dimensionality reduction, and HDBSCAN for clustering.
• Evaluation: Comparison of BERTopic with LDA and NMF based on topic coherence and diversity metrics.
• Scope: Interdisciplinary research from finance, computer science, economics, business, and accounting.

Limitations & caveats

• The review is limited to 90 papers, which may not capture the entire breadth of the rapidly evolving field.
• Relies on Google Scholar, which may miss some niche or non-English publications.
• The BERTopic clustering, while objective, may not perfectly align with human conceptualizations of topics.
• Does not provide empirical performance comparisons of specific models, focusing instead on thematic analysis.
• The discussion of risks and regulations is based on literature review rather than new empirical findings.
• The geographical distribution of authors is skewed towards Asia and the US, potentially missing perspectives from other regions.

Curator notes

Discovery status: round1_promoted_seed_for_round2. Source links: A Survey of Large Language Models in Finance: FinLLMs; AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework; Are ChatGPT and GPT-4 General-Purpose Solvers for Financial Text Analytics? A Study on Several Typical Tasks; BloombergGPT: A Large Language Model for Finance; CFGPT: Chinese Financial Assistant with Large Language Model; DISC-FinLLM: A Chinese Financial Large Language Model based on Multiple Experts Fine-tuning; Deficiency of Large Language Models in F