FinRobot: An Open-Source AI Agent Platform

for Financial Applications using Large Language Models

Hongyang (Bruce) Yang1 , Boyu Zhang1 , Neng Wang1 , Cheng Guo1 ,
Xiaoli Zhang1 , Likun Lin1,2 , Junlin (Jason) Wang 1,4 ,Tianyu Zhou1 , Mao Guan1 ,
Runjia (Luna) Zhang1,4 , Christina Dan Wang1,3∗ ;
1
AI4Finance Foundation; 2 Columbia University;
3
Shanghai Frontiers Science Center of Artificial Intelligence and Deep Learning,
NYU Shanghai; Business Division, NYU Shanghai;
arXiv:2405.14767v2 [q-fin.ST] 27 May 2024

4
Shanghai AI-Finance School ECNU;
[email protected];

Abstract vestment strategies [Greenwald et al., 2004; Penman, 2010;

Berman and Knight, 2013]. Financial analysis precedes fi-
As financial institutions and professionals increas- nancial decisions. Suppose we were to distinguish analysis
ingly incorporate Large Language Models (LLMs) performed in the equity space into two categories. Then, there
into their workflows, substantial barriers, including is fundamental analysis [Abarbanell and Bushee, 1997] that
proprietary data and specialized knowledge, per- studies the companies and gives valuations, while technical
sist between the finance sector and the AI com- analysis studies market action to forecast future price trends
munity. These challenges impede the AI com- [Murphy, 1999].
munity’s ability to enhance financial tasks effec-
tively. Acknowledging financial analysis’s critical The analysis is grounded in data. As the digital revolu-
role, we aim to devise financial-specialized LLM- tion progresses, more data becomes accessible for analysis.
based toolchains and democratize access to them Given the ever-growing volume and complexity of the data
through open-source initiatives, promoting wider involved in the process, financial professionals increasingly
AI adoption in financial decision-making. resort to algorithmic and machine intelligence for data pro-
cessing [Awotunde et al., 2021]. Artificial intelligence (AI)
In this paper, we introduce FinRobot, a novel [Cao, 2022] has revolutionized financial analysis by automat-
open-source AI agent platform supporting multi- ing tasks typically performed by human analysts, such as sen-
ple financially specialized AI agents, each pow- timent analysis [Medhat et al., 2014; Sohangir et al., 2018;
ered by LLM. Specifically, the platform consists Huang et al., 2023; Zhang et al., 2023b] and market predic-
of four major layers: 1) the Financial AI Agents tion [Henrique et al., 2019; Nabipour et al., 2020; Kumar et
layer that formulates Financial Chain-of-Thought al., 2022; Jiang, 2021]. Traditional AI models were initially
(CoT) by breaking sophisticated financial prob- limited to straightforward, single-task operations. However,
lems down into logical sequences; 2) the Finan- advancements in LLMs [Brown et al., 2020; Wu et al., 2023b;
cial LLM Algorithms layer dynamically configures Yang et al., 2023] and computing power have significantly
appropriate model application strategies for spe- enhanced AI’s capabilities. LLMs, with their extensive pa-
cific tasks; 3) the LLMOps and DataOps layer pro- rameter space and diverse training datasets, are adept at as-
duces accurate models by applying training/fine- similating global knowledge, understanding contexts, and ex-
tuning techniques and using task-relevant data; 4) ecuting complex logical reasoning. This expansion in capa-
the Multi-source LLM Foundation Models layer bilities not only allows managing multiple tasks simultane-
that integrates various LLMs and enables the above ously but also extends utility in the financial sector for more
layers to access them directly. Finally, Fin- complex analysis and decision-making.
Robot provides hands-on for both professional-
grade analysts and laypersons to utilize power- Building on the foundation of LLMs, AI agents have
ful AI techniques for advanced financial analy- evolved to employ these models for an array of sophisti-
sis. We open-source FinRobot at https://github. cated functionalities, including planning, memory manage-
com/AI4Finance-Foundation/FinRobot. ment, and tool usage, which facilitate intelligent and au-
tonomous actions in real-world scenarios [Xi et al., 2023;
Gao et al., 2023; Wang et al., 2023c; Durante et al., 2024].
1 Introduction These capabilities allow AI agents to execute complex finan-
Financial analysis typically involves interpreting market cial analyses and operations that traditionally required exten-
trends, predicting economic outcomes, and providing in- sive human intervention and expertise. Recent applications
of AI agents, such as FinAgent [Zhang et al., 2024] for trad-
∗
Corresponding author ing and FinMem [Yu et al., 2023c] for financial decision-
making, underscore the growing reliance on these technolo- In Section 5, we present the performance evaluation for two
gies to perform sophisticated financial operations. Despite demo applications. Section 6 concludes this work and points
their advances, several critical challenges remain unaddressed out directions for future work.
by these solutions:
• Transparency Improvement: In what ways can AI Agent- 2 Related Work
driven financial analysis platforms improve transparency in 2.1 LLM-Based AI Agents
their decision-making processes to build trust among users?
AI Agents, powered by large language models (LLMs), are
• Global Markets Adaptation: How can AI agents effec- intelligent entities capable of perceiving environments, mak-
tively adapt to global stock markets’ multilingual and mul- ing decisions, and executing actions independently. Recent
ticultural nuances to ensure comprehensive and accurate advancements in AI have led to the proliferation of LLM-
market analysis? based AI agents across multiple domains, significantly im-
• Model Diversity: What are the limitations of relying solely pacting areas ranging from creative industries to technical
on a single LLM architecture, such as GPT-4, for complex fields [Sumers et al., 2023; Park et al., 2023; Zeng et al.,
financial analysis, and how can diversifying model archi- 2023]. For example, MusicAgent [Yu et al., 2023a] en-
tectures enhance performance? hances music composition, MedAgents [Tang et al., 2023;
Thirunavukarasu et al., 2023] improves medical diagnostics,
• Real-Time Data Processing: How can AI agent platforms
Data Analytics Agent [Qiao et al., 2023] provides insights
efficiently manage and route vast amounts of financial data
from big data, Programmer Agent [Rasheed et al., 2024] au-
to ensure timely and precise financial analysis?
tomates coding tasks, and ResearchAgent generates research
Recognizing these challenges, this paper introduces Fin- ideas [Baek et al., 2024]. These applications illustrate the
Robot, an innovative open-source AI agent platform that har- versatility of LLMs and set the stage for their pivotal role in
nesses multi-source LLMs for diverse financial tasks. Fin- advancing financial technology.
Robot refines financial workflows from data processing to
strategy implementation, broadening access to advanced an- 2.2 Financial AI Agents
alytical tools and enhancing scalability and transparency. By In the financial sector, AI agent-driven systems like FinAgent
utilizing diverse LLM architectures, it tailors solutions to [Zhang et al., 2024] and FinMem [Yu et al., 2023c] leverage
global market needs, ensuring precise adaptation and perfor- LLMs to transform trading strategies by utilizing real-time
mance optimization. The platform’s multi-source integration market data for informed decision-making. However, the fo-
architecture facilitates the selection of accurate LLMs for spe- cus on metrics such as backtesting and individual stock re-
cific tasks, and its real-time data processing pipeline assures turns often overshadows the need for robust, process-oriented
timely financial analysis. approaches. Effective trading agents should not only perform
Our contributions are significant and address previous well statistically but also enhance operational workflows, au-
shortcomings in the field: tomating tasks from strategy execution to order placement
• Comprehensive Financial AI Agent Framework: Fin- with minimal human intervention. This shift from perfor-
Robot provides a holistic framework for developing finan- mance to process orientation can lead to more sustainable and
cial AI agents that can perform a wide range of tasks from adaptable financial technologies.
market forecasting to financial document analysis. Notably,
FinRobot is the first AI agent platform dedicated to finan- 2.3 Open-Source AI Agent Frameworks
cial analysis, marking a significant advancement in the ap- The development of open-source AI agent frameworks has
plication of AI in finance. been instrumental in democratizing access to advanced AI
• Multi-Source LLM Integration: FinRobot introduces a technologies. Prominent platforms and frameworks such as
novel Smart Scheduler mechanism that allows for the seam- AutoGPT [Gravitas, 2023], AutoGen [Wu et al., 2023a],
less integration of multi-source LLMs, enabling the plat- MetaGPT [Hong et al., 2023], HuggingGPT [Shen et al.,
form to leverage the strengths of various state-of-the-art 2024], ChatDev [Qian et al., 2023], Dify [LangGenius, 2024]
LLMs and select the most suitable ones for specific finan- and Voyager [Wang et al., 2023b] facilitate collaborative en-
cial tasks. This adaptability is crucial for handling the com- hancements and innovation by allowing a global community
plexities of global financial markets and multilingual data. of developers to contribute and refine AI models. Given the
growing need for sophisticated financial tools, the emergence
• Open-Source Platform for Financial AI: By adopting an of an open-source framework specialized for financial tasks
open-source approach, FinRobot fosters broader collabora- is imminent.
tion and accelerates innovation within the financial AI com-
munity. This openness not only promotes transparency but 2.4 AI4Finance Foundation and Open-Source
also allows for continuous enhancements and adaptations, Culture
which are essential in the rapidly evolving financial sector.
The AI4Finance Foundation 1 plays a crucial role in shaping
The remainder of this paper is organized as follows. Sec- the open-source culture within financial technology, driven
tion 2 briefly reviews related work. In Section 3, we de- by its mission to promote standardized practices and develop
scribe the four layers in FinRobot. In Section 4, we present
1
the financial Chain-of-Thought (CoT) Prompting technique. https://www.ai4finance.org

2
 FinRobot: An Open-Source AI Agent Platform
for Financial Applications using Large Language Models
Financial AI Agents
Market Forecasting Agents Document Analysis & Generation Agents
Earnings Report Risk Assessment
Trading Strategist Agent Multimodal Agent
Analysis Agent Report Analysis Agent
Financial Report Equity Research Report
Global Stock Market Agent
Generation Agent Generation Agent

Financial LLMs
Financial Large Langauge Model (FinGPT)

Financial Multimodal LLMs Financial Reinforcement Learning (FinRL) Financial Machine Learning (FinML)

LLMOps model DataOps

interaction FinRL-Meta Financial Data & FinNLP
Prompt LoRA/
SFT Tabular Data Gym Env Market News Social Media
Engineering QLoRA
Model Model Unstructured Third Party Market Financial
CI/CD
Deployment Evaluation data data API Sentiment Ratios
interaction
Instruction Tuning RLHF Smart Scheduler Vector Database RAG Knowledge Graph

Multi-Source LLM Foundation Models

Llama3 Falcon Gemma

ChatGLM3 ChatGPT Baichuan2

Qwen InternLM Mistral

Figure 1: Overall Framework of FinRobot.

open-source resources. This commitment is aimed at bene- Llama series [Meta, 2023] for the U.S. market and Chat-
fiting both researchers and industry professionals, fostering GLM [Zeng et al., 2022] for the Chinese market, each
a collaborative environment where industry-specific knowl- optimized for regional specificities. The Falcon [Al-
edge meets and innovation thrives. By leading initiatives in mazrouei et al., 2023] model excels in financial rela-
applying AI technologies for financial services, AI4Finance tionship analysis. Additionally, multimodal models in-
not only accelerates technological advancements but also en- tegrate text with candlestick charts, while FinRL [Yang
sures these developments are accessible, transparent, and et al., 2020; Liu et al., 2020] optimizes tasks like portfo-
beneficial across the financial sector. lio allocation, and traditional machine learning methods
[Yang et al., 2018] refine stock selection. This approach
3 Overview of FinRobot: An Open-Source ensures high precision in sensitive operations such as
Platform for Financial Tasks market forecasting and financial document analysis.
As delineated in Fig. 1, the overall framework of FinRobot is • LLMOps and DataOps Layers: The LLMOps layer
organized into four distinct layers, each designed to address implements a multi-source integration strategy that se-
specific aspects of financial AI processing and application: lects the most suitable LLMs for specific financial tasks,
• Financial AI Agents Layer: The Financial AI Agents utilizing a range of state-of-the-art models. Initially,
Layer now includes Financial Chain-of-Thought (CoT) general LLMs are deployed; if performance is subopti-
prompting, enhancing complex analysis and decision- mal, the system dynamically switches to fine-tune LLMs
making capacity. Market Forecasting Agents, Document to enhance effectiveness. This adaptive approach en-
Analysis Agents, and Trading Strategies Agents utilize sures tailored solutions for diverse financial scenarios,
CoT to dissect financial challenges into logical steps, enhancing the platform’s overall performance. Concur-
aligning their advanced algorithms and domain exper- rently, the DataOps layer manages real-time data pro-
tise with the evolving dynamics of financial markets for cessing [Liu et al., 2022; Liu et al., 2024; Liu et al.,
precise, actionable insights. 2023], which is crucial for rapid market responsiveness.
This dual-layer configuration bolsters FinRobot’s ability
• Financial LLMs Algorithms Layer: The Financial to deliver timely and accurate financial insights under
LLMs Algorithms Layer configures and utilizes spe- dynamic market conditions.
cially tuned models tailored to specific domains and
global market analysis. It employs FinGPT [Wang • Multi-source LLM Foundation Models Layer: This
et al., 2023d] alongside multi-source LLMs such as foundational layer supports the plug-and-play function-

3
 workflow. This role entails prioritizing financial tasks, allo-
OpenBB FinnHub cating resources, and orchestrating team efforts to optimize
FinNLP YFinance analysis timelines and outcomes.
Text2
Params Data Source API
Planning Memory Decision Assistant The Assistant in financial analysis focuses on ini-
Text2 Function/Tool Calling
Structured Instructions
Code
MplFinance BackTrader
tial data management tasks such as collecting, processing,
Inner
and conducting preliminary analyses of financial data and fa-
Scikit-learn Pyfolio
Reflection cilitates communication between agents. This role supports
Brain Perception
more specialized analysts by preparing datasets, maintaining
Action
Market News / Financials / SEC Filings …
databases, and executing basic financial computations and vi-
Trading Records / Backtest Report …
sualizations. This foundational work enables other analysts to
LLM Analyst
Textual focus on more complex and interpretive aspects of financial
Perception
analysis.
Visual
Query
Assistant LMM Analyst Perception
Response LLM Analyst The LLM Analyst employs advanced com-
UserProxy FinRobot: Agent
··· ··· Workflows
putational techniques to analyze financial texts, such as re-
Director Analyst A Analyst B Analyst C
ports, filings, and news articles, using LLMs. This role is
essential for extracting detailed insights, conducting senti-
Figure 2: Financial AI Agents Workflow Diagram ment analysis, and forecasting market trends from qualita-
tive data. By integrating these insights with quantitative data,
ality of various general and specialized LLMs. It forms the LLM Analyst provides a comprehensive assessment of fi-
the platform’s backbone, ensuring that all models are nancial health and market conditions, enhancing the decision-
up-to-date, optimized, and consistently aligned with the making process.
latest advancements in financial technologies and data
Financial Analysts Operating under the LLM Analyst’s
standards.
guidance, Financial Analysts engage in detailed quantitative
data analysis across various sectors like portfolio manage-
3.1 Financial AI Agents Layer ment, risk assessment, and market analysis. Utilizing sta-
The Financial AI Agents Layer in FinRobot, as depicted in tistical tools and financial models, they interpret data, eval-
Fig. 2, consists of domain-specific AI agents tailored to en- uate investment opportunities, and formulate risk mitigation
hance financial analysis through advanced data perception, strategies. Their expert analysis contributes significantly to
cognitive processing, and dynamic action execution: the development of sound financial strategies and informed
recommendations.
Perception This module captures and interprets multi-
The collaboration among these roles within the multi-agent
modal financial data from market feeds, news, and economic
workflow enables a holistic approach to financial analysis,
indicators, using sophisticated techniques to structure the data
addressing macroeconomic conditions and specific financial
for thorough analysis.
metrics. This structured yet flexible approach ensures thor-
Brain Acting as the core processing unit, this module per- ough exploration and interpretation of financial data, leading
ceives data from the Perception module with LLMs and uti- to informed strategic decision-making.
lizes Financial Chain-of-Thought (CoT) processes, as ex-
plained in Section 4, to generate structured instructions. Utilizing LLMs for Tool Usage
Action This module executes instructions from the Brain API Interaction through Text2Params This method
module, applying tools to translate analytical insights into ac- translates natural language queries into API requests [Wang
tionable outcomes. Actions include trading, portfolio adjust- et al., 2024], effectively combining the capabilities of gen-
ments, generating reports, or sending alerts, thereby actively erating function calls and forming API calls. In this ap-
influencing the financial environment. proach, the LLM first parses the text to identify and extract
key parameters. These parameters are then utilized to dynam-
Multi-Agent Workflow ically generate function calls or compile directly into API re-
Multi-agent workflow systems allow multiple agents, each quests that interact with financial software or databases. This
with unique roles and responsibilities, to tackle intricate tasks method is especially useful for executing predefined opera-
in a cooperative manner. The deployment of a multi-agent tions and fetching or manipulating financial data from exter-
workflow system is beneficial when navigating complex fi- nal services.
nancial datasets and ensuring a high standard of analytical
accuracy and depth. This system incorporates multiple spe- Code Compilation through Text2Code For more com-
cialized roles that work in concert to process, analyze, and plex financial tasks that require dynamic solution generation,
derive actionable insights from diverse financial data sources. LLMs can employ text2code techniques to write and compile
code on the fly. This capability is essential for developing
Director As the strategic leader of the project, the Direc- custom algorithms based on user queries’ unique market con-
tor in a financial analysis context oversees all aspects of the ditions or financial scenarios.

4
3.2 Financial LLMs Algorithms Layer: FinRobot: Smart Scheduler
This layer encompasses advanced AI algorithms specifically Agent Task
Agent 1 Agent 2
Provide Financial
designed to address various needs within the financial sector, Sector Analysis
Adaptor Adaptor
Agent
enhancing the platform’s capabilities across a range of finan- Agent 3 Agent 4 Registration
cial applications. Adaptor Adaptor
Director Agent
Agent 5 Agent 6 Agent
Financial Large Language Models (FinGPT) Adaptor Adaptor Reflections Adaptor
FinGPT [Yang et al., 2023], a domain-specific LLM, is metic- Task
ulously engineered to elevate natural language understand- Select the Best Agent Scores

ing specifically within the financial context. These models based on Task Score
Target
adeptly analyze and interpret financial narratives, extracting Task Manager Agent
critical data from complex documents such as annual reports Agent 1 Agent 2 Agent 3 Agent 4 Agent 5 Agent 6
Golden
Adaptor Adaptor Adaptor Adaptor Adaptor Adaptor
and real-time financial news, thereby supporting enhanced (Llama3) (GPT3.5) (Falcon) (Local) (FinGPT) (GPT-4) Dataset
decision-making processes. 100 80 90 92 98 95 Role
The FinGPT models are mainly developed by supervised Prompt
Store
finetuning with the financial in-domain paired “instruction-
Consumer Future
response” data on the open-source large language models. Financials IT
Discretionary
Energy Utilities Industrials Reflection
Sector Sector Sector Sector Sector
This process is achieved by minimizing the following neg- Sector

ative log-likelihood,
Figure 3: Smart Scheduler Structural Layout in LLMOps Layer
T
X
LCausalLM = − log P (wt |w1 , w2 , . . . , wt−1 ; θ), (1)
t=1  
ST +f,i
where T is the length of the input sequence, wi represents rT +f,i = log , i = 1, . . . , nT , (3)
ST,i
the i-th token in the sequence, θ denotes the model param-
where r denotes the log-return, S is the stock price, nT repre-
eters, and P (wi |w1 , w2 , . . . , wi−1 ; θ) represents the condi-
sents the number of companies monitored at time T , and f is
tional probability of predicting the target token xi given the
the forecast horizon. This measure is essential for evaluating
preceding tokens x1 , . . . , xi−1 . By optimizing this objective
investment performance and shaping financial strategies.
function, the model learns to maximize the probability of gen-
erating the expected response given the finance task instruc- Financial Multimodal LLMs
tion. Financial documents frequently incorporate diverse data
types beyond textual content, such as graphs [Bhatia et al.,
Financial Reinforcement Learning (FinRL)
2024] and tables [Wang et al., 2023a]. These additional
FinRL [Yang et al., 2020] optimizes trading strategies using modalities offer rich, complementary insights, significantly
ensemble deep RL algorithms to analyze historical and real- enhancing the depth of analysis possible. To effectively in-
time market data. This dynamic adaptation helps maximize tegrate and leverage these varied data types, we have devel-
financial returns while minimizing risks, making stock port- oped Financial Multimodal LLMs, which is specifically en-
folio allocation an ideal application for this methodology. gineered to process and synthesize information from multiple
In FinRL, stock portfolio allocation is modeled as a modalities, thereby providing a comprehensive and nuanced
Markov Decision Process (MDP). At any time t, an agent understanding of financial documents.
in state st ∈ S selects an action at ∈ A based on policy The mathematical representation of the integration process
πθ (st ). This action leads to a new state st+1 and a reward in our Financial Multimodal LLM is as follows:
r(st , at , st+1 ). The goal is to optimize this policy:
F (xt , xg , xh ) = L(T (xt ), G(xv ), H(xl )) (4)
πθ∗
= argmax J(πθ ), where F (xt , xg , xh ) denotes the output of the model.
θ Here, xt , xg and xh represent the inputs of textual, graphi-
(2)
" T #
X
t
cal, and tabular data, respectively. The functions T , G and H
where J(πθ ) = E γ r(st , at , st+1 ) , transform these inputs into a unified embedding space. The
t=0 LLM L then synthesizes these embeddings to produce a co-
where γ ∈ (0, 1] is the discount factor. herent and reliable output, enhancing both the accuracy and
reliability of the financial analyses.
Financial Machine Learning (FinML)
FinML [Yang et al., 2018] leverages diverse machine learning 3.3 LLMOps Layer
techniques, from regression to advanced neural networks, to The LLMOps Layer is designed for high modularity and
boost predictive analytics in finance. These algorithms play a pluggability, accommodating rapid integration and dynamic
crucial role in forecasting market trends, consumer behavior, swapping of LLMs in response to evolving technological ad-
credit risks, and other pivotal financial indicators, thus facil- vancements and financial market demands. This layer facil-
itating informed decision-making. A key measure used by itates seamless model integration and includes mechanisms
FinML is the log-return, calculated as follows: for rigorous evaluation and selection of the most suitable

5
models for specific financial tasks. These capabilities are crit- Scoring Metrics
ical for maintaining operational efficiency and ensuring the The scoring process within the Smart Scheduler involves sev-
adaptability of solutions to diverse financial scenarios. eral critical steps to evaluate the performance of different
LLMs for task-specific applications:
Smart Scheduler
The Smart Scheduler, depicted in Figure 3, is central to en- 1. Data Collection: Collect performance data of various
suring model diversity and optimizing the integration and se- LLMs across multiple evaluation tasks.
lection of the most appropriate LLM for each task. 2. Normalization: Normalize the results for each evalua-
tion task to scale between 0 and 1.
Smart Scheduler Architecture
3. Weight Assignment: Assign weights to different evalu-
The Smart Scheduler optimizes task distribution among ation dimensions based on industry standards or expert
agents and comprises: opinions.
• Director Agent: This component orchestrates the task 4. Calculation of Composite Score: Multiply the normal-
assignment process, ensuring that tasks are allocated to ized scores by their respective weights and sum them to
agents based on their performance metrics and suitabil- derive the task score.
ity for specific tasks.
5. Results Analysis: Analyze the scores to rank and as-
• Agent Registration: Manages the registration and sess the LLMs, providing performance comparisons and
tracks the availability of agents within the system, fa- recommendations for selection.
cilitating an efficient task allocation process.
This enhanced structure not only boosts the operational effi-
• Agent Adaptor: Tailors agent functionalities to spe- ciency of financial AI agents but also supports scalable, dy-
cific tasks, enhancing their performance and integration namic management of diverse financial tasks, significantly
within the overall system. improving the system’s capability to manage complex, multi-
agent scenarios in real-time financial environments.
• Task Manager: Manages and stores different general
and fine-tuned LLMs-based agents tailored for various 3.4 DataOps Layer
financial tasks, updated periodically to ensure relevance
and efficacy. The DataOps Layer manages the extensive and varied
datasets necessary for financial analysis, ranging from pub-
Smart Scheduler Initialization Process lic tabular data to proprietary market sentiments. This layer
The initialization process is streamlined to establish a robust is crucial in ensuring that all data fed into the AI process-
foundation for agent operations effectively: ing pipelines is of high quality and representative of the cur-
rent market conditions. Effective data management is foun-
• Golden Dataset Creation: Populates with industry- dational for the accurate and reliable performance of AI mod-
specific data els, and as such, the DataOps Layer employs advanced data
handling and processing techniques to prepare and deliver
• Prompt Store Population: Fills with custom prompts these inputs. By optimizing data accessibility and quality,
for various agents. the DataOps Layer supports the overall efficacy of the Fin-
• Task Score Filling: Invokes adaptors to evaluate re- Robot platform, enabling sophisticated and precise financial
sponses against best answers, storing scores in the Task decision-making based on solid data-driven insights.
Scores database.
Retrieval Augmented Generation
Smart Scheduler In Action Process Retrieval-augmented generation (RAG) is a pivotal technique
incorporated within FinGPT [Zhang et al., 2023b; Zhang et
The operational phase follows this sequence:
al., 2023a], as it seamlessly combines the strength of both
1. Task Initiation: A user initiates a task. context retrieval mechanisms and LLMs to optimize language
generation tasks.
2. Director Agent’s Role: Evaluates task inputs, ranking The LLMOps and DataOps layers form the backbone of
agents by performance and relevance. FinRobot’s operational capabilities, ensuring that the plat-
3. Agent Selection and Task Routing: Routes tasks to the form remains at the forefront of technological innovation
highest-ranked agent while providing reliable and effective financial AI services.

4. Workflow Progression and Self-Evaluation: Post-task 3.5 Multi-source LLM Foundation Models Layer
completion, the agent performs a self-assessment, the re-
sults of which are stored in the Future Reflection storage. The Multi-source LLM Foundation Models Layer equips Fin-
Robot with advanced capabilities to manage and integrate a
5. Workflow Completion and Evaluation: At the end of diverse array of Large Language Models (LLMs), essential
the workflow, the agent evaluates the results and pro- for adapting to the dynamic demands of global financial mar-
vides feedback for continuous improvement. kets. Key features of this layer include:

6
 • Plug-and-Play Functionality: Enables seamless inte- LLMs for financial statement analysis, it conducts
gration and updating of both general and specialized comparative analyses across industry competitors and
LLMs, ensuring the platform remains adaptable and up- against the company’s historical performance. This
to-date with financial technology advancements. comprehensive approach provides profound insights
• Model Diversity and Evaluation: Incorporates LLMs into a company’s business landscape and the logical un-
with parameters ranging from 7 billion to 72 billion, derpinnings of its financial metrics. Specifically, it nor-
each rigorously evaluated for effectiveness in specific malizes financial ratios to identify anomalies, enhancing
financial tasks, allowing for optimal model selection the accuracy of financial insights beyond human capa-
based on performance metrics such as accuracy and bilities.
adaptability. • Business-Specific Analysis: Leveraging Retrieval-
• Global Market Compatibility: Supports multilingual Augmented Generation (RAG), FinRobot revolutionizes
model integration, enhancing the platform’s ability to the process of gathering data on a company’s products
analyze and process diverse financial data, which is cru- and services, complementing the information available
cial for global market operations. in financial reports. By tapping into the vast resources
of the web, FinRobot enriches its understanding of a
This streamlined version succinctly outlines the layer’s company’s offerings, enabling a comprehensive analy-
functionality and strategic role in maintaining FinRobot’s sis of product lines, trends across channels and regions,
leadership in AI-driven financial analysis. cost structures, supply chain dynamics, and R&D con-
versions.
4 Financial Chain-of-Thought (CoT)
• Market Analysis: FinRobot abstracts real-world finan-
Prompting cial contexts, utilizing a combination of financial ra-
4.1 Introduction tios, market data (including stock price trends and can-
The chain-of-thought prompting technique [Wei et al., 2022; dlestick charts), market news sentiment, and alterna-
Xia et al., 2024] structures prompts to foster a step-by-step tive data sourced online. By employing LLMs, Fin-
reasoning process within AI models, akin to human problem- Robot simulates market participants’ decision-making
solving strategies. This method significantly enhances perfor- processes in response to contextual changes. This allows
mance on complex reasoning tasks such as mathematics and for a comprehensive evaluation of a company’s stock
commonsense reasoning by encouraging models to articulate price trajectory and valuation. FinRobot provides nu-
intermediate reasoning steps, leading to a final answer. This anced investment assessments, offering strategic guid-
method improves accuracy and enhances the interpretability ance on optimal investment timing and suitable financial
and transparency of the decision-making process. instruments, ensuring robust, well-informed investment
strategies considering various horizons, structures, and
4.2 Concept of Financial CoT risk thresholds.
Financial CoT Prompting adapts the CoT technique for AI- • Valuation Analysis: FinRobot integrates its compre-
driven financial analysis, integrating advanced cognitive pro- hensive analysis of financial ratios, market data (in-
cessing techniques to refine decision-making capabilities. It cluding stock price trends and candlestick charts), mar-
stimulates the thought process of financial professionals to ket news sentiment, and alternative data collected on-
tackle complex financial issues by guiding AI models through line. FinRobot employs LLMs to evaluate a company’s
a logical, sequential reasoning process. This approach breaks stock price trajectory and valuation using the insights de-
down complex financial scenarios into smaller, manageable rived from these diverse sources. This approach enables
components, analyzes each component, and synthesizes the FinRobot to provide nuanced investment assessments
findings to form conclusions or recommendations. As noted across various horizons, structures, and risk thresholds,
by [Kim et al., 2024], this method mimics human-like rea- offering strategic guidance on optimal investment tim-
soning in the analysis chain and is particularly valuable for ing and suitable financial instruments, thereby ensuring
tasks requiring deep analytical thought, such as valuation, in- well-informed and robust investment strategies.
vestment strategy formulation, market trend analysis, and risk
assessment. 4.4 Advantages of Financial CoT Prompting
4.3 Financial CoT Implementation Despite their merits, a readily available tool that can intelli-
gently perform industry- and company-specific analytics still
Financial Chain-of-Thought (CoT) Prompting marks a trans-
needs to exist in a way that parallels the work of an investment
formative shift in security analysis, enhancing traditional an-
professional. Existing data providers such as Bloomberg,
alytical tools across various domains by leveraging the capa-
FactSet, CapitalIQ, and Refinitiv Eikon rely on rudimentary,
bilities of LLMs such as GPT-4. Detailed case studies, such
generic templates for data extraction, inadvertently overlook-
as those conducted by [Kim et al., 2024], highlight this ap-
ing key operational metrics and qualitative information intri-
proach’s practical applications and benefits.
cately linked with the business essence. Manual analysis can
• Financial Analysis: FinRobot transcends basic data be time-consuming and occasionally repetitive while existing
enumeration and ratio computation. By employing automated tools can only sometimes extract raw data from

7
unstructured sources like annual reports. Although these plat- and analyzed could provide invaluable insights into market
forms abound with vast amounts of information, extracting behavior and strategy development.
essential information and its subsequent analysis still require A detailed framework for such a simulation could involve:
the expertise of investment professionals. (1) Defining the roles and behaviors of different market par-
With CoT Prompting, FinRobot establishes a unique niche ticipants (e.g., institutional investors, retail investors, mar-
among AI Agent Platforms by emulating human cognitive ket makers). (2) Implementing a variety of data inputs and
processes in financial analysis, transcending traditional re- decision-making models for each participant. (3) Simulating
liance on mere numerical computations. Our methodol- market interactions over different time horizons and under
ogy leverages LLMs to deconstruct and analyze the method- varying conditions to assess potential outcomes and strate-
ologies utilized by financial professionals when scrutinizing gies. (4) Utilizing reinforcement learning to refine the agents’
business data. This approach unveils a practical analysis in- strategies based on observed performance in the simulation.
trinsically dependent on a comprehensive understanding of As outlined by [Park et al., 2023], this concept presents
a company’s business fundamentals, as reflected through de- a promising avenue for future research and development in
tailed accounting data, tailored industry-specific metrics, and financial market analysis and strategy optimization.
qualitative information.
This approach overcomes the above limitations because: 5 Demo Applications and Hands-on Tutorial
(1) It does not rely on a rigid template due to the generative
nature of LLMs. (2) It avoids misclassification because error- 5.1 Application I: Market Forecaster
checking is embedded within layers of prompting. (3) It sim- Market Forecaster is a set of AI Agents designed to synthesize
plifies human labor by codifying investment logic within its recent market news and financial data, delivering comprehen-
reasoning. (4) It can extract relevant historical numbers and sive insights into a company’s latest achievements and poten-
qualitative information from complex and unstructured data tial concerns, along with predictions for stock price move-
formats such as earnings call transcripts. (5) Given its multi- ments. As a junior robo-advisor, Market Forecaster embodies
layered structure, it provides stronger grounds for explain- a substantial leap forward in AI-driven financial advisory.
ing the sources and derivations of each recorded and derived
value. (6) It is highly adaptable and evolving as LLMs may
be improved through supervised fine-tuning and instruction- Instruction: You are an experienced stock market analyst.
tuning. Your task is to list the company’s positive developments and
potential concerns based on the company’s relevant news and
4.5 Market Simulation: Transcending the Analyst quarterly financials in the past few weeks, and then combine
Mimicking them with your views on the overall financial economic market
judgment, providing predictions and analysis of the company’s
To expand the concept of market simulation and explore its stock price changes in the coming week. Your answer format
possibilities, one could view the GPT’s ability to mimic hu- should be as follows:
man analytical processes as a gateway to simulating broader
market dynamics. Market simulation transcends mere nu- [Positive development]:
merical analysis by incorporating human-like reasoning pro- 1. ...
cesses, embodying multiple personas and scenarios within [Potential concerns]:
a financial context. This simulation could encompass a va- 1. ...
riety of market participants whose decisions are influenced [Forecast and Analysis]:
by an evolving set of data inputs—ranging from financial ...
statements, reports, market data, and economic indicators to
Information:
global news. a. Company Introduction
The core of this approach involves constructing a detailed b. Stock Price Changes
simulation environment where these generative agents, repre- c. Recent News Information
senting different market players, interact [Park et al., 2023]. d. Recent Basic Financials
Each agent’s decision-making process would be guided by
data-driven prompts, mimicking the cognitive steps a hu- Instruction: Based on all the information before 2024-04-19,
man analyst might take. This includes interpreting complex let’s first analyze the positive developments and potential con-
market signals and reacting to new information in real time, cerns for AAPL. Come up with 2-4 most important factors re-
which advanced machine learning models and reinforcement spectively and keep them concise. Then make your prediction
of the AAPL price movement for next week (2024-04-22 to
learning techniques could facilitate.
2024-04-26). Provide a summary analysis to support your pre-
By leveraging LLMs, one could theoretically encode these diction.
decision-making frameworks into the agents, enabling them
to perform tasks that require a deep understanding of the mar-
ket’s quantitative and qualitative aspects. This would simu- Figure 4: Market Forecaster Prompt Template
late individual trading decisions and the broader market dy-
namics resulting from multiple decision-makers interactions. Data
The potential to use this technology to create a virtual mar- Market Forecaster stands at the crossroads of global mar-
ket environment where hypothetical scenarios could be tested kets, leveraging diverse data sources to gather information

8
 Forecaster for NVDA on 2024-01-29 Forecaster for Kweichow Moutai on 2024-01-28
[Positive Developments]: [积极发展]:
1. NVDA’s stock price has been steadily increasing over the 1. 贵州茅台的购买率持续上涨，表明市场对茅台的需求
past weeks, suggesting a strong investor sentiment. (Stock 持续增长。(News)
Price) 2. 公司的新闻发布，如“美茅台美世界”2024全球品牌
2. The company’s CEO, Jensen Huang, has expressed 文化活动，表明公司继续追求国际化发展，并且将继
optimism about the future of AI and Nvidia’s role in it, which 续以“美”的国际表达，坚持追求美，实现看好未来的发
could boost investor confidence. (News) 展。(News)
3. The company has received attention from Zacks.com users, 3. 茅台的营业总收入同比增长率达到了19.80%，表明公
indicating a strong performance in the market. (News) 司的经营状况良好，具有稳定的盈利能力。(Basic Finan-
4. The company has a strong balance sheet with a healthy cials)
gross margin and a high EV. (Basic Financials) 4. 茅台的流动比率和速动比率都相对较高，表明公司的
资金流动性较好，可以更好地应对市场的压力。(Basic Fi-
[Potential Concerns]: nancials)
1. The company’s long-term debt-to-total asset and debt-to- [潜在担忧]:
total capital ratios are relatively high, indicating a potential 1. 茅台的资产负债率达到了17.98%，显示公司的负债
risk of financial instability. (Basic Financials) 压力较大，可能会对公司的盈利能力产生影响。(Basic
2. The company’s high PE ratio and high net debt-to-total Financials)
equity ratio suggest high valuation and potential financial risk. 2. 公司未来可能面临的市场竞争压力可能会增加，如果
(Basic Financials) 公司无法继续提供高质量的产品和服务，可能会受到市场
3. The company’s quick ratio is relatively low, indicating a 反应。(News)
potential liquidity risk. (Basic Financials) 3. 茅台的购买率上涨，可能会导致市场价格的偏高，可
能会影响到公司的股价。(Stock Price)
[Prediction & Analysis]: [预测和分析]:
Prediction: Up by 0-1% 预测涨跌幅: 上涨0-1%
Analysis: Despite some potential concerns, the positive devel- 总结分析：虽然贵州茅台的资产负债率较高，但公司的经
opments for NVDA are significant. The company’s strong bal- 营状况良好，营业总收入同比增长率达到了19.80%，显
ance sheet and increasing stock price suggest a strong investor 示公司在市场的需求和竞争中具有稳定的盈利能力。考虑
sentiment. The optimistic outlook from the CEO and the at- 到茅台的购买率持续上涨，表明市场对茅台的需求持续增
tention from Zacks.com users also suggest a positive future 长，可能会对公司的股价产生积极影响。同时，公司的新
outlook. However, the high long-term debt-to-total asset and 闻发布，如“美茅台美世界”2024全球品牌文化活动，表明
debt-to-total capital ratios are a cause for concern. These high 公司继续追求国际化发展，可能会带来未来的发展机会。
ratios could potentially increase the company’s financial risk 因此，预计茅台的股价在下一周将上涨0-1%。
and affect its liquidity.
Figure 6: Market Forecaster Result of Kweichow Moutai
Figure 5: Market Forecaster Result of Nvidia
gineering to format the instructive prompt with the struc-
and make forecasting decisions. In detail, Market Fore- ture of ‘Task Instruction & Company Information (Company
caster gathers data regarding multifaceted company informa- Overview + Recent Stock Prices + Recent News + Latest Ba-
tion such as recent news, latest basic financials, and stock sic Financials)’. Figure 4 shows an example of the Market
prices for the target markets, including the US stock market, Forecaster prompt template.
Chinese stock market, Crypto market, and other possible ex-
tensions. This section will give a solid overview of the Market Sample Forecastings
Forecaster targeted at the US stock market and Chinese stock The sample answers provided for two prominent stocks in
market with data sources Finnhub and EastMoney. the US and Chinese markets, namely Nvidia and Kweichow
Moutai, as depicted in Figures 5 and 6, respectively, illus-
Model trate the Market Forecaster’s adeptness in synthesizing and
This AI Agent is powered by fine-tuned FinGPT-Forecasters offering valuable insights gleaned from diverse information
using Llama-2-7b-chat-hf with LoRA, leveraging data from sources. Additionally, the Market Forecaster furnishes rec-
the latest year’s DOW 30 for the US Market and SSE 50 for ommendations regarding the future trajectory of the stocks,
the Chinese Market to ensure precise forecasts for these ma- underscoring its capability to provide actionable guidance
jor stocks and demonstrating robust generalization abilities based on the analyzed data.
across various stock symbols. Moreover, this paper follows
the Financial Multi-task Instruction Tuning paradigm[Wang 5.2 Application II: Document Analysis &
et al., 2023d; Yu et al., 2023b] to align the power of the base Generation
model with the specific functions of Market Forecaster.
The Document Analysis & Generation application represents
Prompts a groundbreaking use of AI Agents and Large Language
Aligning with the multi-task instruction tuning framework, Models (LLMs) in the realm of financial document man-
the Market Forecaster follows a sophisticated prompt for- agement and report creation. This application harnesses the
mat. Specifically, the Market Forecaster gathers data of mul- power of AI to perform deep analysis of financial documents
tifaceted company information, then it performs prompt en- and generate detailed, insightful reports automatically.

9
Document Analysis Penedo. Falcon-40B: an open large language model with
AI Agents, integrated with advanced LLMs, are employed to state-of-the-art performance. arXiv, 2023.
sift through extensive financial documents such as annual re- [Awotunde et al., 2021] Joseph Bamidele Awotunde, Em-
ports, SEC filings, and earnings call transcripts. These agents manuel Abidemi Adeniyi, Roseline Oluwaseun Ogun-
are capable of extracting critical information, identifying key dokun, and Femi Emmanuel Ayo. Application of big data
financial indicators, and highlighting trends and discrepan- with fintech in financial services. In Fintech with artifi-
cies that may require closer inspection. The ability of these cial intelligence, big data, and blockchain, pages 107–132.
AI Agents to understand and process complex financial ter- Springer, 2021.
minology and context turns vast amounts of unstructured data
[Baek et al., 2024] Jinheon Baek, Sujay Kumar Jauhar, Sil-
into structured, actionable insights.
viu Cucerzan, and Sung Ju Hwang. Researchagent: Itera-
Report Generation tive research idea generation over scientific literature with
Following the analysis phase, the same agent is utilized to large language models. arXiv preprint arXiv:2404.07738,
generate comprehensive financial reports. Leveraging the ca- 2024.
pabilities of LLMs, the system produces coherent, articulate, [Berman and Knight, 2013] Karen Berman and Joe Knight.
and detailed documents that cover various aspects of financial Financial intelligence, revised edition: A manager’s guide
analysis, including performance evaluations, market compar- to knowing what the numbers really mean. Harvard Busi-
isons, and forward-looking financial forecasts. Each report is ness Review Press, 2013.
crafted to maintain a professional tone and format, mirroring
the quality and depth expected of top-tier financial analysts. [Bhatia et al., 2024] Gagan Bhatia, El Moatez Billah
For practical insights, equity research reports generated by Nagoudi, Hasan Cavusoglu, and Muhammad Abdul-
FinRobot are included in the appendix, showcasing the appli- Mageed. Fintral: A family of gpt-4 level multi-
cation and effectiveness of our platform. modal financial large language models. arXiv preprint
arXiv:2402.10986, 2024.
6 Conclusion [Brown et al., 2020] Tom Brown, Benjamin Mann, Nick Ry-
der, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari-
FinRobot revolutionizes financial analysis by integrating
wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry,
multi-source Large Language Models (LLMs) in an open-
Amanda Askell, et al. Language models are few-shot
source platform that enhances accessibility, efficiency, and
learners. Advances in Neural Information Processing Sys-
transparency in financial operations. This innovative platform
tems, 33:1877–1901, 2020.
addresses the complexities of global markets with a multi-
layered architecture that supports real-time data processing [Cao, 2022] Longbing Cao. Ai in finance: challenges,
and diverse model integration, making sophisticated financial techniques, and opportunities. ACM Computing Surveys
tools available to a wider audience. By fostering collabora- (CSUR), 55(3):1–38, 2022.
tion within the financial AI community, FinRobot not only [Durante et al., 2024] Zane Durante, Qiuyuan Huang, Naoki
accelerates innovation but also sets new standards for the ap- Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan
plication of AI in finance, promising to significantly improve Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi,
strategic decision-making across the sector. et al. Agent ai: Surveying the horizons of multimodal in-
Future Work As FinRobot continues to evolve, we plan teraction. arXiv preprint arXiv:2401.03568, 2024.
to expand its applications to include more sophisticated tasks [Gao et al., 2023] Chen Gao, Xiaochong Lan, Nian Li, Yuan
such as portfolio allocation and comprehensive risk assess-
Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li.
ment, enhancing its utility in the financial sector. Addition-
Large language models empowered agent-based modeling
ally, we aim to broaden the platform’s reach by extending
and simulation: A survey and perspectives. arXiv preprint
its capabilities to more global markets. This expansion will
arXiv:2312.11970, 2023.
not only diversify the potential use cases of FinRobot but
also improve its adaptability to different economic environ- [Gravitas, 2023] Significant Gravitas. Autogpt. https://
ments, driving innovation and accessibility in AI-driven fi- github.com/Significant-Gravitas/AutoGPT, 2023.
nancial analysis across the world. [Greenwald et al., 2004] B.C. Greenwald, J. Kahn, P.D.
Sonkin, and M. van Biema. Value Investing: From Gra-
References ham to Buffett and Beyond. Business book summary. Wi-
[Abarbanell and Bushee, 1997] Jeffrey S Abarbanell and ley, 2004.
Brian J Bushee. Fundamental analysis, future earnings, [Henrique et al., 2019] Bruno Miranda Henrique, Vini-
and stock prices. Journal of accounting research, 35(1):1– cius Amorim Sobreiro, and Herbert Kimura. Literature
24, 1997. review: Machine learning techniques applied to finan-
[Almazrouei et al., 2023] Ebtesam Almazrouei, Hamza cial market prediction. Expert Systems with Applications,
Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, 124:226–251, 2019.
Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, [Hong et al., 2023] Sirui Hong, Mingchen Zhuge, Jonathan
Daniel Heslow, Julien Launay, Quentin Malartic, Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jin-
Badreddine Noune, Baptiste Pannier, and Guilherme lin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin,

10
 Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, [Park et al., 2023] Joon Sung Park, Joseph O’Brien, Car-
and Jürgen Schmidhuber. Metagpt: Meta programming for rie Jun Cai, Meredith Ringel Morris, Percy Liang, and
a multi-agent collaborative framework, 2023. Michael S Bernstein. Generative agents: Interactive sim-
[Huang et al., 2023] Allen H Huang, Hui Wang, and ulacra of human behavior. In Proceedings of the 36th
Yi Yang. Finbert: A large language model for extracting Annual ACM Symposium on User Interface Software and
information from financial text. Contemporary Account- Technology, pages 1–22, 2023.
ing Research, 40(2):806–841, 2023. [Penman, 2010] S. Penman. Accounting for Value. Columbia
[Jiang, 2021] Weiwei Jiang. Applications of deep learning in Business School Publishing. Columbia University Press,
2010.
stock market prediction: recent progress. Expert Systems
with Applications, 184:115537, 2021. [Qian et al., 2023] Chen Qian, Xin Cong, Wei Liu, Cheng
Yang, Weize Chen, Yusheng Su, Yufan Dang, Jiahao Li,
[Kim et al., 2024] Alex Kim, Maximilian Muhn, and Va-
Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun.
leri V Nikolaev. Financial statement analysis with large Communicative agents for software development, 2023.
language models. Chicago Booth Research Paper Forth-
coming, Fama-Miller Working Paper, 2024. [Qiao et al., 2023] Bo Qiao, Liqun Li, Xu Zhang, Shilin He,
Yu Kang, Chaoyun Zhang, Fangkai Yang, Hang Dong, Jue
[Kumar et al., 2022] Deepak Kumar, Pradeepta Kumar Zhang, Lu Wang, et al. Taskweaver: A code-first agent
Sarangi, and Rajit Verma. A systematic review of framework. arXiv preprint arXiv:2311.17541, 2023.
stock market prediction using machine learning and
[Rasheed et al., 2024] Zeeshan Rasheed, Muhammad
statistical techniques. Materials Today: Proceedings,
49:3187–3191, 2022. Waseem, Mika Saari, Kari Systä, and Pekka Abrahams-
son. Codepori: Large scale model for autonomous
[LangGenius, 2024] LangGenius. Dify. https://github.com/ software development by using multi-agents. arXiv
langgenius/dify, 2024. preprint arXiv:2402.01411, 2024.
[Liu et al., 2020] Xiao-Yang Liu, Hongyang Yang, Qian [Shen et al., 2024] Yongliang Shen, Kaitao Song, Xu Tan,
Chen, Runjia Zhang, Liuqing Yang, Bowen Xiao, and Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hug-
Christina Dan Wang. FinRL: A deep reinforcement learn- ginggpt: Solving ai tasks with chatgpt and its friends in
ing library for automated stock trading in quantitative fi- hugging face. Advances in Neural Information Processing
nance. Deep RL Workshop, NeurIPS 2020, 2020. Systems, 36, 2024.
[Liu et al., 2022] Xiao-Yang Liu, Ziyi Xia, Jingyang Rui, [Sohangir et al., 2018] Sahar Sohangir, Dingding Wang,
Jiechao Gao, Hongyang Yang, Ming Zhu, Christina Dan Anna Pomeranets, and Taghi M Khoshgoftaar. Big data:
Wang, Zhaoran Wang, and Jian Guo. FinRL-Meta: Mar- Deep learning for financial sentiment analysis. Journal of
ket environments and benchmarks for data-driven financial Big Data, 5(1):1–25, 2018.
reinforcement learning. NeurIPS, 2022. [Sumers et al., 2023] Theodore R Sumers, Shunyu Yao,
[Liu et al., 2023] Xiao-Yang Liu, Guoxuan Wang, Karthik Narasimhan, and Thomas L Griffiths. Cogni-
Hongyang Yang, and Daochen Zha. Data-centric tive architectures for language agents. arXiv preprint
fingpt: Democratizing internet-scale data for financial arXiv:2309.02427, 2023.
large language models. NeurIPS Workshop on Instruction [Tang et al., 2023] Xiangru Tang, Anni Zou, Zhuosheng
Tuning and Instruction Following, 2023. Zhang, Yilun Zhao, Xingyao Zhang, Arman Cohan, and
[Liu et al., 2024] Xiao-Yang Liu, Ziyi Xia, Hongyang Yang, Mark Gerstein. Medagents: Large language models as col-
Jiechao Gao, Daochen Zha, Ming Zhu, Christina Dan laborators for zero-shot medical reasoning. arXiv preprint
Wang, Zhaoran Wang, and Jian Guo. Dynamic datasets arXiv:2311.10537, 2023.
and market environments for financial reinforcement [Thirunavukarasu et al., 2023] Arun James
learning. Machine Learning - Nature, 2024. Thirunavukarasu, Darren Shu Jeng Ting, Kabilan
[Medhat et al., 2014] Walaa Medhat, Ahmed Hassan, and Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel
Hoda Korashy. Sentiment analysis algorithms and ap- Shu Wei Ting. Large language models in medicine.
plications: A survey. Ain Shams engineering journal, Nature medicine, 29(8):1930–1940, 2023.
5(4):1093–1113, 2014. [Wang et al., 2023a] Dongsheng Wang, Natraj Raman,
[Meta, 2023] Meta. LLaMA 2: Open foundation and fine- Mathieu Sibue, Zhiqiang Ma, Petr Babkin, Simerjot
Kaur, Yulong Pei, Armineh Nourbakhsh, and Xiaomo
tuned chat models. Preprint, 2023.
Liu. Docllm: A layout-aware generative language model
[Murphy, 1999] John J Murphy. Technical analysis of the fi- for multimodal document understanding. arXiv preprint
nancial markets: A comprehensive guide to trading meth- arXiv:2401.00908, 2023.
ods and applications. Penguin, 1999. [Wang et al., 2023b] Guanzhi Wang, Yuqi Xie, Yunfan
[Nabipour et al., 2020] Mojtaba Nabipour, Pooyan Nayyeri, Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi
Hamed Jabani, Amir Mosavi, Ely Salwana, and Shahab Fan, and Anima Anandkumar. Voyager: An open-
S. Deep learning for stock market prediction. Entropy, ended embodied agent with large language models. arXiv
22(8):840, 2020. preprint arXiv:2305.16291, 2023.

11
[Wang et al., 2023c] Lei Wang, Chen Ma, Xueyang Feng, [Yu et al., 2023a] Dingyao Yu, Kaitao Song, Peiling Lu,
Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Tianyu He, Xu Tan, Wei Ye, Shikun Zhang, and Jiang
Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large Bian. Musicagent: An ai agent for music understanding
language model based autonomous agents. arXiv preprint and generation with large language models. arXiv preprint
arXiv:2308.11432, 2023. arXiv:2310.11954, 2023.
[Wang et al., 2023d] Neng Wang, Hongyang Yang, and [Yu et al., 2023b] Xinli Yu, Zheng Chen, and Yanbin Lu.
Christina Dan Wang. Fingpt: Instruction tuning bench- Harnessing llms for temporal data-a study on explainable
mark for open-source large language models in financial financial time series forecasting. In Proceedings of the
datasets. NeurIPS Workshop on Instruction Tuning and 2023 Conference on Empirical Methods in Natural Lan-
Instruction Following, 2023. guage Processing: Industry Track, pages 739–753, 2023.
[Wang et al., 2024] Zhiruo Wang, Zhoujun Cheng, Hao Zhu, [Yu et al., 2023c] Yangyang Yu, Haohang Li, Zhi Chen,
Daniel Fried, and Graham Neubig. What are tools any- Yuechen Jiang, Yang Li, Denghui Zhang, Rong Liu, Jor-
way? a survey from the language model perspective. arXiv dan W. Suchow, and Khaldoun Khashanah. Finmem:
preprint arXiv:2403.15452, 2024. A performance-enhanced llm trading agent with layered
memory and character design, 2023.
[Wei et al., 2022] Jason Wei, Xuezhi Wang, Dale Schuur-
mans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, [Zeng et al., 2022] Aohan Zeng, Xiao Liu, Zhengxiao Du,
Denny Zhou, et al. Chain-of-thought prompting elicits Zihan Wang, Hanyu Lai, Ming Ding, Zhuoyi Yang, Yi-
reasoning in large language models. Advances in neural fan Xu, Wendi Zheng, Xiao Xia, et al. Glm-130b:
information processing systems, 35:24824–24837, 2022. An open bilingual pre-trained model. arXiv preprint
arXiv:2210.02414, 2022.
[Wu et al., 2023a] Qingyun Wu, Gagan Bansal, Jieyu Zhang,
Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun [Zeng et al., 2023] Fanlong Zeng, Wensheng Gan,
Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadal- Yongheng Wang, Ning Liu, and Philip S Yu. Large
lah, Ryen W White, Doug Burger, and Chi Wang. Auto- language models for robotics: A survey. arXiv preprint
gen: Enabling next-gen llm applications via multi-agent arXiv:2311.07226, 2023.
conversation framework. In arXiv, 2023. [Zhang et al., 2023a] Boyu Zhang, Hongyang Yang, and
[Wu et al., 2023b] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Xiao-Yang Liu. Instruct-fingpt: Financial sentiment analy-
Dabravolski, Mark Dredze, Sebastian Gehrmann, Prab- sis by instruction tuning of general-purpose large language
hanjan Kambadur, David Rosenberg, and Gideon Mann. models. FinLLM Symposium at IJCAI 2023, 2023.
BloombergGPT: A large language model for finance. [Zhang et al., 2023b] Boyu Zhang, Hongyang Yang, Tianyu
arXiv preprint arXiv:2303.17564, 2023. Zhou, Ali Babar, and Xiao-Yang Liu. Enhancing financial
[Xi et al., 2023] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei sentiment analysis via retrieval augmented large language
He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe models. ACM International Conference on AI in Finance
Wang, Senjie Jin, Enyu Zhou, et al. The rise and poten- (ICAIF), 2023.
tial of large language model based agents: A survey. arXiv [Zhang et al., 2024] Wentao Zhang, Lingxuan Zhao, Hao-
preprint arXiv:2309.07864, 2023. chong Xia, Shuo Sun, Jiaze Sun, Molei Qin, Xinyi Li,
[Xia et al., 2024] Yu Xia, Rui Wang, Xu Liu, Mingyan Li, Yuqing Zhao, Yilei Zhao, Xinyu Cai, et al. Finagent: A
Tong Yu, Xiang Chen, Julian McAuley, and Shuai Li. Be- multimodal foundation agent for financial trading: Tool-
yond chain-of-thought: A survey of chain-of-x paradigms augmented, diversified, and generalist. arXiv preprint
for llms. arXiv preprint arXiv:2404.15676, 2024. arXiv:2402.18485, 2024.
[Yang et al., 2018] Hongyang Yang, Xiao-Yang Liu, and Disclaimer: We are sharing codes for academic purposes
Qingwei Wu. A practical machine learning approach for under the MIT education license. They should not be con-
dynamic stock recommendation. 2018 17th IEEE In- strued as financial counsel or recommendations for live
ternational Conference On Trust, Security And Privacy trading. It is imperative to exercise caution and consult
In Computing And Communications/ 12th IEEE Interna- with qualified financial professionals prior to any trading
tional Conference On Big Data Science And Engineering or investment actions.
(TrustCom/BigDataSE), pages 1693–1697, 2018.
[Yang et al., 2020] Hongyang Yang, Xiao-Yang Liu, Shan
Zhong, and Anwar Walid. Deep reinforcement learning
for automated stock trading: An ensemble strategy. In Pro-
ceedings of the first ACM international conference on AI
in finance, pages 1–8, 2020.
[Yang et al., 2023] Hongyang Yang, Xiao-Yang Liu, and
Christina Dan Wang. Fingpt: Open-source financial large
language models. FinLLM Symposium at IJCAI 2023,
2023.

12
Equity Research Report: NVIDIA Corporation FinRobot
https://ai4finance.org/
Income Summarization https://github.com/AI4Finance-Foundation/FinRobot
In fiscal year 2024, NVIDIA witnessed a remarkable revenue increase to $60.922 Report date: 2024-02-21
billion, up 126% from the previous year, driven predominantly by a 217% surge in
Data Center revenue, highlighting the robust demand for its computing platforms. The
Compute & Networking segment, in particular, saw revenue skyrocket to $47.405 Key data
billion, up 215%, with significant contributions from the NVIDIA Hopper GPU Rating hold
computing platform for AI applications. Meanwhile, the Graphics segment also Target Price 850 - 1100 (md. 875.0)
experienced growth, with a 14% increase in revenue to $13.517 billion, bolstered by a
15% rise in Gaming revenue due to the launch of new GPUs and enhanced 6m avg daily val (USDmn) 46.27
AI-powered gaming experiences. The company's gross profit margin improved to Closing Price (USD) 694.49
72.7%, and net profit margin expanded to 48.9%, reflecting cost efficiency, a Market Cap (USDmn) 1663859.52
favorable product mix, and operational effectiveness. Despite challenges such as
52 Week Price Range (USD) 210.94 - 724.98
increased operational costs and supply chain complexities, NVIDIA's strategic
partnerships and continuous innovation in AI and gaming technologies position it for BVPS (USD) 17.41
sustained growth and market leadership.
Share Performance
Business Highlights
- Data Center: Revenue surged by 217% due to strong demand for NVIDIA Hopper
GPU computing platforms for AI applications. This growth was driven by the
increasing adoption of AI across various industries, including healthcare and
automotive. - Gaming: Revenue increased by 15%, driven by higher sell-in to
partners following the normalization of channel inventory levels. The growth reflects
the continued demand for gaming GPUs and the introduction of new products. -
Professional Visualization: Revenue saw a modest increase of 1%, indicating stable
demand in the professional graphics market. This stability is attributed to ongoing PE & EPS
demand for NVIDIA's visualization products across industries. - Automotive: Revenue
grew by 21%, primarily due to growth in self-driving platforms. This growth signifies
NVIDIA's expanding presence in the automotive sector, particularly in AI and
autonomous driving technologies.

Company Situation
NVIDIA Corporation, a leader in accelerated computing, thrives on innovation in GPU
technology and AI, catering to diverse sectors including AI, data analytics, and 3D
graphics. Its competitive edge is bolstered by significant R&D; investments, leading
to breakthroughs like the GPU and deep learning algorithms. Current trends like the
demand for AI, the rise of the metaverse, and the need for high-performance
computing present both opportunities and challenges, including supply chain issues
and intense competition. NVIDIA's strategic initiatives, such as launching new GPU
architectures and expanding into data center CPUs, alongside partnerships in
autonomous driving, aim to sustain its industry leadership by continuously innovating
and adapting to market needs.

Risk Assessment
The top 3 key risks for NVIDIA Corporation include: 1) Failure to adapt to rapid
changes in technology and customer requirements, which could adversely impact
financial results; 2) Intense competition that could reduce market share and financial
performance; and 3) Regulatory, legal, and political actions, including export controls
and tariffs, that may restrict business operations or increase operational costs,
particularly those affecting the semiconductor and AI industries, which could
materially impact the company's ability to manufacture, sell, and distribute products
globally.
Financial Metrics
FY (USD mn) 2020 2021 2022 2023 2024

Operating Revenue 10918.0 16675.0 26914.0 26974.0 60922.0

Adjusted Net Profit 2796.0 4332.0 9752.0 4368.0 29760.0

Adjusted EPS 1.15 1.76 3.91 1.76 12.05

EBIT Margin 0.26 0.27 0.37 0.16 0.54

ROE 0.23 0.26 0.37 0.2 0.69

PE Ratio 52.32 75.41 62.67 109.1 51.82

EV/EBITDA 40.53 58.59 54.7 66.08 44.84

PB Ratio 11.99 19.34 22.97 21.56 35.88

13
Revenue
segment FY2023Q3 FY2023Q4 FY2024Q1 FY2024Q2 FY2024Q3 FY2024Q4

AutomotiveMember 251.00 294.00 296.00 253.00 261.00 281.00

DataCenterMember 3833.00 3616.00 4284.00 10323.00 14514.00 18404.00

GamingMember 1574.00 1831.00 2240.00 2486.00 2856.00 2865.00

OEMAndOtherMember 73.00 84.00 77.00 66.00 73.00 90.00

ProfessionalVisualizationMember 200.00 226.00 295.00 379.00 416.00 463.00

total 5931.00 6051.00 7192.00 13507.00 18120.00 22103.00

Ratio
segment FY2023Q3 FY2023Q4 FY2024Q1 FY2024Q2 FY2024Q3 FY2024Q4

AutomotiveMember 0.04 0.05 0.04 0.02 0.01 0.01

DataCenterMember 0.65 0.60 0.60 0.76 0.80 0.83

GamingMember 0.27 0.30 0.31 0.18 0.16 0.13

OEMAndOtherMember 0.01 0.01 0.01 0.00 0.00 0.00

ProfessionalVisualizationMember 0.03 0.04 0.04 0.03 0.02 0.02

Yoy
segment FY2023Q3 FY2023Q4 FY2024Q1 FY2024Q2 FY2024Q3 FY2024Q4

AutomotiveMember 86.00 135.00 114.00 15.00 4.00 -4.00

DataCenterMember 31.00 11.00 14.00 171.00 279.00 409.00

GamingMember -51.00 -46.00 -38.00 22.00 81.00 56.00

OEMAndOtherMember -69.00 -56.00 -51.00 -53.00 0.00 7.00

ProfessionalVisualizationMember -65.00 -65.00 -53.00 -24.00 108.00 105.00

total -17.00 -21.00 -13.00 101.00 206.00 265.00

14
证券研究报告: 贵州茅台 FinRobot
https://ai4finance.org/
核心观点 https://github.com/AI4Finance-Foundation/FinRobot
2023年，公司实现营收1476.94亿元，同比增长19.01%，归母净利润747.34亿元，增长 Report date: 2024-04-03
19.16%。这一强劲的业绩表明，尽管市场环境充满挑战，公司依然能够通过优化销售
策略和调整产品矩阵，持续增长和盈利。特别是，公司对直销渠道的重视促进了营收
的显著增加，而在产品方面，系列酒的快速增长开始改变收入结构，为公司的持续发 基本数据
展注入新动力。这些成绩反映了公司不仅在行业中保持了领先地位，同时也拥有良好 评级 买入
的市场竞争力。对于未来，公司通过进一步优化直销渠道和扩展产品矩阵的策略，展
目标价 1953.65 - 2600.0 (md. 2246.9)
现出对增强市场竞争力和持续增长能力的明确信心。
近六月日平均成交量 (RMBmn) 0.03

业绩总览 收盘价 (RMB) 1715.11

公司发布了2023年的主要经营数据：营收达到约1476.94亿元，同比增长约19.01%；归 总市值 (RMBmn) 2125499.24

属于母公司的净利润为747.34亿元，同比增长约19.16%。该数据显示公司在面临挑战 52周价格范围 (RMB) 1555.55 - 1935.00
的市场环境中继续展现出强劲的成长动力和盈利能力，稳健的经营业绩体现了其在行
业中的领先地位和良好的市场竞争力。 股价表现

销售模式
根据提供的数据，公司的销售渠道包含直销和批发代理两大部分。从2019年至2023年
数据分析，直销营收呈现显著增长趋势，增速从65.65%逐年攀升至高峰的105.5%，尽
管在2023年增速回落至36.16%，但直销营收仍然保持强劲增长。相比之下，批发代理
营收增长相对稳定，除2022年出现-9.31%的下滑外，其他年份增速较低。此数据显示
公司销售策略逐年偏向直销渠道，直销渠道对收入的贡献日益增大，而批发代理渠道
的影响相对减弱。直销渠道的强劲增长可能是因为直接与消费者的互动提升了销售效
率和客户满意度，也反映了市场需求和消费者偏好的变化。 市盈率和每股收益

产品矩阵
根据所提供数据，公司产品矩阵围绕“茅台酒”和“系列酒”。茅台酒作为核心产品
，其收入持续占据主导地位，2019-2023年收入增速呈上升趋势，由15.75%增至17.39%
。系列酒虽然起点较低，但增速更快，尤其从2020年开始，其增速显著跃升，2023年
达到29.43%。这显示系列酒在产品矩阵中的比重正在增长，对收入的贡献越来越显著
。因此，虽然茅台酒仍是收入的主要来源，但系列酒的快速增长正在逐步改变公司的
收入结构，显示公司正通过扩展产品矩阵增加收入的多样性和稳定性。
金融指标
报告期 2019 2020 2021 2022 2023

营业收入 888.54亿 979.93亿 1094.64亿 1275.54亿 1505.60亿

营业总收入同比增长率 15.10% 10.29% 11.71% 16.53% 18.04%

净利润 412.06亿 466.97亿 524.60亿 627.17亿 747.34亿

净利润同比增长率 17.05% 13.33% 12.34% 19.55% 19.16%

销售净利率 51.47% 52.18% 52.47% 52.68% 52.49%

销售毛利率 91.30% 91.41% 91.54% 91.87% 91.96%

基本每股收益 32.8000 37.1700 41.7600 49.9300 59.4900

净资产收益率 33.09% 31.41% 29.90% 30.26% 34.19%

市盈率 42.3138 58.9284 55.8193 41.3543 34.5714

市净率 11.8922 16.357 14.9506 12.1073 11.1874

15
16