FinAgent: Benchmarking Financial Analysis with Stock Agent Framework

Anonymous

Abstract:

Constructing stock agents that facilitate investment analysis is an important research direction in finance. The key technology is the agent's capability to automatically identify user queries and integrate multimodal data for analysis by large language models (LLMs).
Currently, LLMs have made some progress, primarily in retrieving text and time-series data from knowledge bases based on user queries and providing these data in a basic combination to LLMs.
However, they have not efficiently integrated these data to enhance the performance of the LLMs.
Also, they do not fully exploit image information and depend on extensive knowledge bases that require real-time updates.
To overcome these limitations, we propose the FinAgent dataset, which encompasses research datasets, financial Q\&A, stock charts, and handwritten chain-of-thought (CoT) data.
Moreover, we innovate a Stock-Agent that efficiently discerns user intent and retrieves necessary information via APIs and knowledge bases to tackle financial tasks.
Additionally, we propose an efficient multimodal information fusion method that enhances data sorting and organization, thereby improving LLMs' analytical quality. We conduct extensive experiments to demonstrate the effectiveness of our framework in financial analysis.

Paper Type: long

Research Area: NLP Applications

Contribution Types: Data analysis

Languages Studied: Englisg

Preprint Status: We are considering releasing a non-anonymous preprint in the next two months (i.e., during the reviewing process).

A1: yes

A1 Elaboration For Yes Or No: section 7

A2: no

A3: yes

A3 Elaboration For Yes Or No: section 1

B: no

B1: no

B2: no

B3: no

B4: no

B5: no

B6: yes

B6 Elaboration For Yes Or No: section 2 and 4

C: yes

C1: no

C2: no

C3: yes

C3 Elaboration For Yes Or No: section 4

C4: yes

C4 Elaboration For Yes Or No: section 4

D: no

D1: no

D2: no

D3: no

D4: no

D5: no

E: no

E1: no