Before AI was used for financial stock analysis, I've seen two papers:
Today, I am looking at a new one: FinSphere. It is a conversational AI assistant for stock analysis, aiming to revolutionize the field of financial analysis.
FinSphere has three major innovations
Stocksis Dataset: Carefully curated by industry experts, specifically designed to enhance the stock analysis capabilities of LLMs.
AnalyScore Evaluation Framework: A systematic tool for evaluating the quality of stock analysis, making the objectivity and accuracy of the results quantifiable and comparable.
FinSphere Intelligent Agent: Capable of generating high-quality stock analysis reports based on specific user needs.
FinSphere Intelligent Agent
4.1 Strong Quantitative Tools Based on Real-time Databases
One of FinSphere's core advantages lies in its seamless integration with the company’s mature quantitative analysis tools, which have been widely deployed and validated in production environments. These tools can access the company's real-time financial database that comprehensively covers all market stocks, including structured data (stock price trends, trading volumes, financial indicators) and unstructured data (company announcements, analyst reports, market news), etc.
When FinSphere identifies specific quantitative analysis requirements, it automatically triggers the corresponding tools. The tools query the real-time database, extract the latest relevant data, and perform complex calculations to generate professional analyses such as technical indicators, fundamental valuations, or market sentiment assessments. Each tool is customized for the user's queries, fully utilizing the constantly updated database to ensure that the analysis results accurately reflect real-time market conditions. This architecture ensures that FinSphere's responses are always based on the latest market data while benefiting from proven quantitative methods.
4.2 Specialized Instruction Tuning
We fine-tuned the Qwen2-72B model using the expert-curated Stocksis dataset, significantly enhancing the model's financial analysis capabilities. The Stocksis dataset contains 5,000 high-quality training pairs, each consisting of structured prompts, comprehensive outputs from quantitative tools, and corresponding expert analyses. The fine-tuning process employed LoRA (Low-Rank Adaptation) technology, which efficiently updates parameters while maintaining the model's general capabilities. Through this method, the model learned to understand various outputs from quantitative tools, integrate multidimensional analytical perspectives, and generate well-structured reports following professional analysis patterns.
4.3 Overall Workflow of FinSphere
FinSphere completes comprehensive financial analysis through a systematic, multi-stage process. After the user submits a query request, FinSphere first uses the chain-of-thought (CoT) method to break down the analysis request into structured subtasks and determine the required quantitative tools for each task.
After completing the task decomposition, the selected quantitative tools independently access the real-time financial database, retrieve the latest market data, and conduct professional analyses. These analyses cover multiple dimensions, from technical indicators to fundamental indicators, ensuring the comprehensiveness and real-time nature of the content.
The final stage is completed by the model fine-tuned with the Stocksis dataset. The model integrates and deeply analyzes the professional analyses generated by each quantitative tool, ultimately outputting a high-quality comprehensive analysis report. Through instruction fine-tuning, the model demonstrates excellent capabilities in understanding quantitative outputs and generating professional-level financial analyses. This integrated workflow ensures that FinSphere's analysis results combine the precision of quantitative analysis with the deep insights of expert financial reasoning, while always maintaining real-time market relevance.
Comparison
Example
