PaperlayerView exact revision on arXiv

Source-linked AI summary

FinGPT: Democratizing Internet-scale Data for Financial Large Language Models

Xiao-Yang Liu, Guoxuan Wang, Hongyang Yang, Daochen Zha

arXiv:2307.10485v2cs.CLcs.LGq-fin.GN

TL;DR

Financial LLM development faces limited public financial text data, costly training, and the challenge of keeping models current as markets change. FinGPT combines real-time data collection and curation with lightweight model adaptation; the authors report 34 data pipelines and lower adaptation costs than BloombergGPT, while noting room for further improvement.

  • Problem

    Financial LLM development is constrained by limited financial text datasets, BloombergGPT’s unavailable data and model, costly training, and the need to adapt as markets evolve.

    Source: §1 · p. 2

  • Method

    FinGPT is an open, data-centric framework that automates real-time financial-data collection and curation and supports lightweight adaptation of general-purpose LLMs.

    Source: §1 · p. 2

  • Results

    FinGPT introduces 34 data pipelines and parameter-efficient adaptation that the authors say reduces adaptation costs and computational requirements compared with BloombergGPT.

    Source: §7 · p. 10

  • Takeaways & Limitations

    The pipeline provides accessible financial-data infrastructure for researchers and practitioners developing FinLLMs.

    Source: §1 · p. 2

    The authors state that substantial room for improvement remains in democratizing financial data.

    Source: §7 · p. 11

Abstract

from arXiv · show

Large language models (LLMs) have demonstrated remarkable proficiency in understanding and generating human-like texts, which may potentially revolutionize the finance industry. However, existing LLMs often fall short in the financial field, which is mainly attributed to the disparities between general text data and financial text data. Unfortunately, there is only a limited number of financial text datasets available, and BloombergGPT, the first financial LLM (FinLLM), is close-sourced (only the training logs were released). In light of this, we aim to democratize Internet-scale financial data for LLMs, which is an open challenge due to diverse data sources, low signal-to-noise ratio, and high time-validity. To address the challenges, we introduce an open-sourced and data-centric framework, Financial Generative Pre-trained Transformer (FinGPT), that automates the collection and curation of real-time financial data from 34 diverse sources on the Internet, providing researchers and practitioners with accessible and transparent resources to develop their FinLLMs. Additionally, we propose a simple yet effective strategy for fine-tuning FinLLM using the inherent feedback from the market, dubbed Reinforcement Learning with Stock Prices (RLSP). We also adopt the Low-rank Adaptation (LoRA, QLoRA) method that enables users to customize their own FinLLMs from general-purpose LLMs at a low cost. Finally, we showcase several FinGPT applications, including robo-advisor, sentiment analysis for algorithmic trading, and low-code development. FinGPT aims to democratize FinLLMs, stimulate innovation, and unlock new opportunities in open finance. The codes have been open-sourced.

1 Introduction

FinLLM development is constrained by scarce, inaccessible financial datasets, costly training, and the market’s time sensitivity. FinGPT responds with an open, data-centric framework for real-time financial-data collection and curation, alongside application demonstrations.

Motivation

  • Financial datasets are limited, BloombergGPT is not publicly available, and its training cost about $2.67 million as market conditions make frequent model updates necessary.The paper also identifies diverse sources, low signal-to-noise ratio, and time sensitivity as data-access and curation challenges.Source: §1 · p. 2

FinGPT

  • FinGPT is an open-sourced, data-centric framework that automates real-time financial-data collection and curation while enabling lightweight adaptation of general-purpose LLMs.The framework prioritizes data sources and data quality.Source: §1 · p. 2

Data Curation Pipeline

  • FinGPT automates real-time curation across more than 34 varied sources, making Internet-scale financial data accessible through APIs.The sources span news, social media, filings, and scholarly datasets.Source: §1 · p. 2

Application Effectiveness

  • The authors report empirical validation of curated data for financial applications, including robo-advisors, sentiment analysis for algorithmic trading, and low-code development.They state that the applications’ performance and accuracy improve in real-world financial settings.Source: §1 · p. 2

2 Related Work

Prior work uses financial text for prediction, trading, and model training, but the authors identify a lack of open collection and curation pipelines suited to time-sensitive markets. FinGPT emphasizes an accessible data pipeline and efficient fine-tuning with real-time market feedback.

Prior Financial Text Research

  • Earlier studies use financial text for prediction, trading, task-specific models, or training from scratch, while FinGPT uses it for efficient fine-tuning with real-time market feedback.The authors note that previous work lacked an open-sourced data collection and curation pipeline for practical use in time-sensitive markets.Source: §2 · p. 3

Related Efforts

  • FinGPT distinguishes itself from a contemporary financial-text effort by offering a streamlined data pipeline and from the FinGPT vision paper by adding dataset-focused evaluations.The current paper centers on datasets that users can use to train their own FinLLMs.Source: §2 · p. 3

3 Data-centric FinGPT Framework for FinLLMs

The paper frames open financial-data access as central to developing FinLLMs, which must contend with diverse, noisy, time-sensitive data. FinGPT combines accessible data, curation, lightweight adaptation, and application layers; BloombergGPT is presented as a capable but closed and costly comparator.

3.1 Challenges of Training FinLLMs

  • Financial data comes from diverse sources, has low signal-to-noise ratio, and loses representativeness as markets change, motivating specialized pipelines, curation, and real-time collection.The section sets the objective of obtaining an open-source FinLLM with strong performance on financial tasks.Source: §3.1 · pp. 3–4

3.2 Overview of FinGPT Framework

  • FinGPT provides unified access to Internet-scale financial data, data-curation pipelines, lightweight adaptation with reinforcement learning and LoRA or QLoRA, and a four-layer architecture.The four layers are data source, data curation, LLM, and application; the overview reports adaptation costs as low as $262.Source: §3.2 · p. 4

3.3 Proprietary Model BloombergGPT

  • BloombergGPT demonstrated strong results across financial tasks and can generate valid Bloomberg Query Language, but its closed access, $2.67 million training cost, and costly retraining constrain public use and updating.The paper attributes the need for retraining to the financial market’s dynamic nature.Source: §3.3 · p. 4

4 Demoncratizing Internet-scale Financial Data

FinGPT’s financial-data infrastructure collects and prepares data for model training, with source-specific access interfaces and automated real-time curation. Its described preparation includes text cleaning, document-quality filtering, and use of a pretrained tokenizer.

4 Demoncratizing Internet-scale Financial Data

  • The paper presents collecting, preparing, and processing high-quality training data as central to its data-centric strategy.It directs readers to code and usage examples for the data strategies.Source: §4 · p. 5

4.1 Financial Data Sources

  • FinGPT draws on financial news for market and company information, social media for public sentiment, company filings and market data, and curated research datasets for financial analysis.Examples include Yahoo and Seeking Alpha, Twitter and Reddit, SEC filings, and Stocknet, CHRNN, TTE, Astock [42], FiQA SA, and FPB [17].Source: §4.1 · p. 5

4.2 Data Interface

  • The unified data interface offers date-range retrieval and streaming access to the latest content, though not every source supports both.Streaming access uses a pages parameter; interface availability depends on source limitations.Source: §4.2 · pp. 5–6

4.3 Automated Real-Time Data Curation Pipeline

  • The automated real-time curation pipeline supplies data for LLM training or fine-tuning in markets where news and sentiment can quickly affect prices.The authors describe real-time processing as important because delays may mean missed opportunities or increased risk.Source: §4.3 · p. 6

4.4 Data Cleaning

  • Data cleaning standardizes whitespace and removes URLs, uncommon characters, and excessively long words to reduce non-natural-language content.The authors state that these steps enhance data usability and reliability.Source: §4.4 · p. 6

4.5 Document Filtering

  • Document filtering screens for length, special-character abundance, repetition, low perplexity or language-identification scores, and duplicate or near-duplicate documents.The authors use n-gram frequencies to analyze repetition and retain one representative instance of duplicate content.Source: §4.5 · pp. 6–7

4.6 Tokenization

  • FinGPT tokenizes text with a pretrained tokenizer provided through HuggingFace.The cited tokenizer is the pre-trained tokenizer used in the framework.Source: §4.6 · p. 7

5 Lightweight Adaptation of General-Purpoose LLMs to FinLLMs

FinGPT addresses costly full-model fine-tuning and the difficulty of obtaining real-time labels by combining parameter-efficient adaptation with market-derived stock-price labels. Users can train their own LoRA weights, which the framework does not present as direct financial advice.

5 Lightweight Adaptation of General-Purpoose LLMs to FinLLMs

  • Real-time financial fine-tuning is constrained by the cost of training many parameters and the difficulty of obtaining human annotations.The section identifies RLHF as requiring annotations that are difficult to obtain in real time.Source: §5 · p. 7
  • LoRA and QLoRA reduce trainable parameters and training cost, while RLSP uses relative stock-price changes as labels for positive, negative, or neutral outputs.The implementation labels texts positive above a 2% price increase, negative below a 2% decrease, and neutral within -2% to 2%, without human participation.Source: §5 · p. 7
  • Users can train plug-and-play LoRA weights themselves; FinGPT supplies tools rather than direct financial advice and frames this design as protecting user privacy.The framework describes user control over the weights as supporting community engagement and safeguarding privacy.Source: §5 · p. 7

6 Demonstrative Applications of FinGPT

FinGPT is demonstrated through robo-advisory, sentiment analysis for quantitative trading, and low-code financial coding tasks. The examples include improved sentiment results after fine-tuning and applications for summarizing news and developing or finding factors.

6 Demonstrative Applications of FinGPT

  • The demonstrations cover personalized robo-advice, trading signals that can be combined with risk control, and graphical-interface-based low-code development.The project provides tutorials for these applications and guidance for lightweight model adaptation.Source: §6 · pp. 7–8

6.1 Application I: Robo-Advisor

  • In an AAPL news example, FinGPT summarizes positive catalysts and risks, then describes a positive but potentially volatile price outlook.The example analyzes news from March 3, 2023.Source: §6.1 · p. 8

6.2.1 Labeling by Market

  • FinGPT consistently outperforms LLaMA on market-labeled sentiment, with substantial gains excluding neutral and an improved average cumulative return rate in simulated trading.Table 1 defines ACC and F1 measures for all labels and without neutral, and Avg. CRR for simulated trading.Source: Table 1 · p. 9, §6.2.1 · p. 9

6.2 Application II: Sentiment Analysis for Quantitative Trading

  • The sentiment-analysis experiments use either RLSP market labels or labels from GPT-4, and both settings show significant enhancements over prevailing LLMs.The market-label experiment uses news for S&P 500 constituents and a 2% price threshold.Source: §6.2 · pp. 8–9, §6.2.1 · p. 9

6.2.2 Supervised Fine-tuning

  • Fine-tuning with FinGPT-collected data significantly enhances performance on labeled academic datasets.Table 2 reports F1 scores; it omits BloombergGPT on TFNS and NWGI because those results are unavailable.Source: §6.2.2 · p. 10, Table 2 · p. 10

6.3 Application III: Low-code Development

  • FinGPT’s code-generation capability is presented as reducing the time and effort of factor-library construction, and tailored prompts can expedite the search for new factors.The examples concern financial coding tasks and factors used in development and production environments.Source: §6.3 · p. 10

7 Conclusion, Discussions, and Future Work

FinGPT combines 34 data pipelines with parameter-efficient adaptation of existing LLMs and reports experiments on three representative financial tasks. The conclusion also says there remains ample room for improvement.

7 Conclusion, Discussions, and Future Work

  • FinGPT introduces 34 data pipelines and adapts pre-existing LLMs with parameter-efficient methods, reducing adaptation costs and computational requirements compared with BloombergGPT.The paper presents this as a more accessible, flexible, and cost-effective FinLLM solution for the open-source community.Source: §7 · p. 10
  • Experiments on three representative financial tasks demonstrate FinGPT’s efficacy, while the authors note that the work leaves ample room for improvement.The conclusion frames these experiments as showing the promise of Internet-scale financial data for training FinLLMs.Source: §7 · pp. 10–11

A Discussion of Dynamic Datasets and Fine-tuning Methods

Financial interpretations shift with market conditions, motivating continual model updates and alignment with both human judgment and market dynamics. The section introduces RLSP, which uses stock-price fluctuations as fine-tuning labels, while noting its risk of overfitting market trends.

A Discussion of Dynamic Datasets and Fine-tuning Methods

  • Market interpretation can change even when the core information remains similar, as the 2022 Meta metaverse news example illustrates.The passage attributes the shift in interpretation to a change in the project's net present value as rates escalated.Source: A Discussion of Dynamic Datasets and… · p. 16
  • Continual updates are needed to calibrate models to prevailing market conditions; the inaugural release uses LoRA for resource-efficient weight fine-tuning.The authors invite evaluation of other fine-tuning methods in finance.Source: A Discussion of Dynamic Datasets and… · p. 16
  • Financial alignment should reflect both human judgment and market dynamics, since human labels are costly and insufficient for predicting practical outcomes such as stock prices.Source: A Discussion of Dynamic Datasets and… · p. 16
  • RLSP fine-tunes FinGPT using stock-price fluctuations as labels, automating label collection and reflecting market trends.The authors say this is intended to keep the model in sync with market movements.Source: A Discussion of Dynamic Datasets and… · p. 16
  • The authors identify potential overfitting to market trends because stock prices respond to influences beyond news.They plan to examine alternative market indicators for labeling.Source: A Discussion of Dynamic Datasets and… · p. 16

B Data Sources

This section describes FinGPT’s financial and social-media data sources, their access interfaces and retrieval examples, and how their coverage and access modes vary. It also presents downloader interfaces and crawling practices for efficiency, incremental collection, and recovery from failures.

B.1.1 News

  • News sources cover financial and company news across markets, but access varies: Seeking Alpha and Finnhub support streaming and date ranges, while many other sources are streaming-only.The section also reports stock-related document counts for six U.S. tech companies in Table 4, without providing the counts in these passages.Source: §B.1.1 · pp. 16, 17, 19, Table 4 · p. 19

B.1.2 Social Media

  • Social-media sources include Twitter, Reddit, Weibo, Xueqiu, Facebook, StockTwits, and Eastmoney; Twitter and Weibo support date-range access, while several others are streaming-only.The wallstreetbets subreddit can provide company-related data and broader market changes, but its access is streaming-only.Source: §B.1.2 · pp. 19–20

B.1.3 Filing

  • SEC EDGAR and Juchao provide searchable company filings and both support streaming and date-range interfaces.Juchao is the disclosure platform for Shenzhen-listed companies; SEC provides access to U.S. filings and forms.Source: §B.1.3 · pp. 20–21

B.1.4 Research Dataset

  • Stocknet covers two years of price movements for 88 stocks; CHRNN, TradeTheEvent, and Astock [42] provide datasets for stock prediction, event detection, or trading.Stocknet spans 01/01/2014 to 01/01/2016; TradeTheEvent is an open-source corporate-event detection and news-based prediction benchmark, and Astock is an open-source dataset and automated trading system.Source: §B.1.4 · p. 21
  • FPB [17] supports three-way financial-news sentence sentiment classification, while FiQA SA covers sentiment in English financial news and microblog headlines.FPB labels are neutral, positive, and negative; FiQA SA was released for the 2018 financial question-answering and opinion-mining challenge.Source: §B.1.4 · p. 21

B.2 Example Codes for Accessing Data

B.3.1 Asynchronous Crawling

  • Multiprocessing pools are used to crawl large volumes of content asynchronously and improve efficiency, with an example pool size of 40.The implementation checks which date files already exist before crawling and closes and joins the pool when work is complete.Source: §B.3.1 · p. 27

B.3.2 Incremental Saving

  • Saving each stock’s results separately by day enables incremental crawling and makes failed content easier to recrawl.The paper presents daily CSV files as an example of this storage structure.Source: §B.3.2 · p. 27

C Quantitative Analysis of Training Cost

The cost analysis estimates FinGPT's LoRA fine-tuning expense against training costs for other LLMs. Its estimates use AWS prices as a reference and may vary with the training systems used.

C Quantitative Analysis of Training Cost

D Performance Metrics

This section defines accuracy, F1 score, and cumulative return rate (CRR) as metrics used in the experiments. The CRR calculation averages per-stock returns, while the final portfolio values depend on the trading strategy.

D Performance Metrics

D.1 Accuracy

  • Accuracy is the fraction of input samples classified correctly, calculated from true positives, true negatives, false positives, and false negatives.Source: §D.1 · pp. 28–29

D.2 F1 Score

  • F1 score is the harmonic mean of precision and recall, with precision and recall defined using true positives, false positives, and false negatives.Source: §D.2 · p. 29

D.3 Cumulative Return Rate (CRR)

  • Average CRR averages each stock's cumulative return, calculated from its final and initial portfolio values; those values depend on the trading strategy.Source: §D.3 · p. 29

E.1 Data

This section specifies the financial sentiment datasets, market-derived labels, data splits, training settings, and models used in the experiments.

E.1 Data

  • The experiments use 582,734 FMP news items with Yahoo Finance prices, split chronologically at 2021-11-01 into 465,645 training/validation and 117,089 test items.The training/validation portion is randomly split 80/20 with seed 42.Source: §E.1 · p. 29

E.2 Market-Related Financial Labels

  • News receives Positive, Negative, or Neutral labels from stock-price changes: above 2%, below -2%, or between those thresholds, respectively.The resulting label shares are about 42% positive, 37% negative, and 21% neutral.Source: §E.2 · p. 29

E.3 Training Details

  • LoRA training uses rank 8, alpha 16, dropout 0.05, batch size 128, and learning rate 3e-3.Other implementation details are referred to the project website.Source: §E.3 · p. 30

F.1 Datasets and Splits

  • The benchmark dataset splits follow the setting described in BloombergGPT [1].The provided passage does not specify further split details for this subsection.Source: §F.1 · p. 30

F.1.1 Financial Phrasebank (FPB)

  • FPB contains about 5,000 investor-perspective news sentences, with sentiment annotated by 5–8 finance-informed annotators; experiments use the at-least-50%-agreement configuration.The evaluation uses a random seed-42 split aligned with BloombergGPT [1] and five-shot weighted F1.Source: §F.1.1 · p. 30

F.1.2 FiQA SA

  • FiQA SA’s continuous sentiment scores are discretized into negative, neutral, and positive classes, then evaluated in a five-shot setup using weighted F1.The experiment combines microblogs and news in a random split; the training and test sets contain 970 and 243 sentences, respectively.Source: §F.1.2 · p. 30

F.1.3 TFNS; F.1.4 NWGI

  • TFNS provides 11,932 labeled finance-related tweets, while NWGI supplies 16.2k training and 4.05k test samples with seven ChatGPT-generated labels and rationales.TFNS labels are Bearish, Bullish, and Neutral; NWGI labels range from strong negative to strong positive.Source: §F.1.3 · p. 30, §F.1.4 · p. 31

F.2 Models

  • Comparisons cover pretrained BloombergGPT [1], ChatGLM2 [55], Llama2, ChatGPT, and GPT-4 [6], plus FinGPT-data fine-tuned ChatGPT via OpenAI’s API and Llama2/ChatGLM2 using LoRA [37], with ChatGLM2 int-8 [58] and QLoRA [36] variants.Source: §F.2 · p. 31

F.3 Training Details

This section specifies the LoRA fine-tuning configuration and illustrates building a factor library from fundamental and price momentum factors, including their combination. It also clarifies that FinGPT offers an automated real-time data-curation pipeline and access interface—not a new dataset—with open-source resources and ongoing maintenance.

F.3 Training Details

  • LoRA fine-tuning used rank 8, alpha 16, dropout 0.05, batch size 16, and learning rate 1e-4 on an A100 GPU.The authors direct readers to their website for further details.Source: §F.3 · p. 31

G Example of Building a Factor Library

  • The factor-library example implements price-to-earnings ratio, moving average, and volatility functions, then adds five fundamental factors: ROE, debt-to-equity, EPS, current ratio, and gross margin.The example also notes that factor functions can be customized or expanded, with inputs supplied from available financial data.Source: G Example of Building a Factor Library · pp. 32–33

H Examples of Building Factors

  • The stock-factor workflow loads CSV data into a pandas DataFrame and displays its first rows before factor construction.The example names open, high, low, close, volume, and amount as available columns.Source: H Examples of Building Factors · p. 34
  • For a fundamental momentum factor, the example calculates a chosen metric’s percentage change over a period, ranks companies, and standardizes ranks to mean 0 and standard deviation 1.EPS and a 12-month period are defaults; the dataset must include the selected fundamental metric.Source: H Examples of Building Factors · pp. 34–35
  • The price momentum example applies the same percentage-change, ranking, and standardization steps to a technical indicator, using closing price and 12 months as defaults.The example requires price data and allows other technical indicators to be substituted.Source: H Examples of Building Factors · pp. 35–36
  • The combined momentum factor averages or weights the fundamental and price momentum factors, reflecting their complementary views of momentum.The example assigns a default weight of 0.5 to FMF when weights are unspecified.Source: H Examples of Building Factors · p. 36
  • The Combined Momentum Factor combines FMF and PMF using assigned weights, defaulting to 0.5 for each when weights are omitted.Source: H Examples of Building Factors · pp. 36–37
  • The example Non-Linear Momentum Factor multiplies log-transformed FMF by exponentially transformed PMF to represent nonlinear combinations.Source: H Examples of Building Factors · p. 37

I Accessibility and Maintenance

J.1 Data-centric AI

This discussion presents FinGPT as a data-centric financial LLM framework, connecting curated data and lightweight adaptation with market feedback, openness, and privacy-conscious use.

J.1 Data-centric AI

  • FinGPT emphasizes data quality because flaws in static datasets can impair practical performance and data cascades can reduce accuracy in real-world deployments.The framework uses an automated real-time curation pipeline for quality control [20] [19] [23].Source: §J.1 · p. 38

J.2 Domain-specific LLMs and Proprietary Data

  • Domain-specific LLMs can outperform general-purpose counterparts, with proprietary datasets supplying domain insights absent from public data [1] [60].FinGPT builds on open-source LLMs and fine-tunes them with specialized financial data [7] [8] [9].Source: §J.2 · pp. 38–39

J.3 From General-Purpose LLMs to FinLLMs

  • FinGPT aims to democratize financial data for adapting pretrained LLMs to evolving markets, and its sources could also support direct pretraining before task-specific fine-tuning.The paper also presents an inaugural holistic overview that brings together earlier project visions and RLSP methods [7] [8] [9].Source: §J.3 · p. 39

J.4 Parameter-Efficient Tuning of LLMs

  • LoRA updates trainable low-rank matrices while keeping the remaining model parameters fixed, enabling lightweight, plug-and-play adaptation [37].FinGPT uses LoRA but is compatible with other efficient tuning methods for adjustment to different market environments.Source: §J.4 · p. 39

J.5 Alignment of LLMs

  • RLSP uses market feedback as cost-effective annotations for refining models, supporting updates in response to changing markets.The passage contrasts this with human feedback, which is often costly.Source: §J.5 · p. 39

J.6 Openness of LLMs

  • FinGPT’s open-everything approach aims to share data, models, and tools with researchers and practitioners to support collaboration and transparency.The stated vision is an open-data, open-model, open-finance community [7] [8] [9].Source: §J.6 · p. 40, §K.1 · p. 40

K.1 Vision

  • The project frames open data and open models as foundations for a collaborative community advancing financial technology and accessible open finance [7] [8] [9].The paper acknowledges that this vision remains an expansive and challenging journey.Source: §K.1 · p. 40

K.2 Privacy, Ethics, and Openness; K.2.1 Privacy and Ethics

  • FinGPT commits to using publicly accessible, freely usable training sources and supports local inference and locally held LoRA weights to protect privacy.The paper also describes public-data use as mitigating organizational concerns about sharing private data; federated learning is discussed as an approach being explored.Source: §K.2.1 · p. 41

K.2.2 Openness

FinGPT advocates opening public-data APIs and models trained on public data while withholding private data to mitigate leakage risk. Its future work spans broader data access, fine-tuning, applications, and model capabilities.

K.2.2 Openness

  • FinGPT favors open-sourcing public-data APIs and models trained on public data, while withholding private data to mitigate leakage risk and recognizing that APIs may raise privacy concerns.The authors present this balance as facilitating research and adoption while safeguarding privacy to the greatest extent possible.Source: §K.2.2 · pp. 41–42
  • Open services provide accessible, adaptable tools and APIs, while education, research, and community initiatives invite broader use, scrutiny, and collaboration.The paper describes tutorials and best practices, encourages replication or extension of its findings, and welcomes feedback and contributions.Source: §K.2.2 · p. 42

K.3 Future Work

  • Current access focuses on US and Chinese stock markets; the authors plan to expand to Europe, Japan, and beyond.They identify broader access to high-quality data sources as future work.Source: §K.3 · p. 42
  • Experiments use LoRA and QLoRA; other parameter-efficient fine-tuning methods may achieve better results with fewer computational resources.The authors plan to investigate these alternatives.Source: §K.3 · p. 42
  • The paper fine-tunes a general LLM; the authors anticipate that fine-tuning pre-trained FinLLMs could yield higher performance, potentially with additional training resources.They report that their current approach already surpasses pre-trained FinLLMs such as BloombergGPT.Source: §K.3 · pp. 42–43
  • Other future directions include advanced prompting, longer context windows, finance-specific retrieval-augmented generation, fairness evaluation, and richer low-code support.The low-code plans include adding pseudocode; the fairness work will evaluate disparities across stocks and explore mitigation strategies.Source: §K.3 · p. 43
  • Future applications include SWAP decision support and risk analysis, as well as event and outlier detection in financial time series.The paper identifies these as potential applications of FinLLMs.Source: §K.3 · p. 43
  • The authors encourage community members to train more diverse LoRA weights using different data sources.They anticipate that collective work will advance FinLLMs.Source: §K.3 · p. 43

Related papers

Loading 2307.10485v2…