Guojin Securities' big model enables investment research scenario exploration

In the current era of rapid development of digitalization and intelligence, the financial sector is undergoing profound changes. As an important part of the financial system, the securities industry is facing unprecedented opportunities and challenges. With the continuous advancement of artificial intelligence technology, especially the continuous breakthroughs in large model technology, the intelligent development of the securities industry has ushered in new opportunities. So far, at least 30 securities companies have announced the access to large model products such as DeepSeek or promoted their localized deployment, which marks an important step forward for the securities industry on the road to intelligent development.

The application of big models in the securities industry is mainly concentrated in the fields of investment research, investment consulting, and operations. Through document question and answer, indicator prediction, text generation, etc., big models provide strong technical support for the securities industry. But at the same time, its application in the securities industry also faces three major problems and challenges.

How to choose the most appropriate big model application scenario to ensure its rapid adaptation and maximize its effectiveness is a core issue that securities companies urgently need to solve. Securities companies need to choose typical scenarios to apply big models based on the technical advantages of big models and their own business needs and technical capabilities to ensure that they can give full play to their technical advantages and see results as soon as possible.

Although the big model has been deployed locally in some securities companies and preliminarily verified in multiple scenarios, the specific effect and accuracy data have not yet been fully disclosed. Issues such as accuracy evaluation and performance optimization of the big model adaptation scenario need to be further resolved to ensure its actual application effect in the securities industry.

The application of big models and their Chain of Thought (CoT) technology in the securities industry faces the problem of how to fully tap their potential and apply deep thinking capabilities to innovative applications. Specifically, the Chain of Thought technology can demonstrate the complete thinking and reasoning process and improve the explainability and transparency of reasoning, but there is no clear answer to how to apply the Chain of Thought technology to the securities industry and provide incremental innovation support for investment research and analysis. In practical applications, how to optimize the information extraction process, improve the clarity of reasoning logic, and refine investment strategies through the Chain of Thought are issues that need to be explored in depth in specific scenarios.

Currently, big model technology is in the "long slope and thick snow" stage, and its versatility and spillover effects are triggering a profound productivity change. As a data-intensive industry, the financial industry has rich data resources, a sound digital foundation and diversified business scenarios, and is expected to become a pioneer in the implementation of big model technology.

To adapt to this trend, Guojin Securities Co., Ltd. (hereinafter referred to as "Guojin Securities") proposed and practiced the concept of "AI-friendly organization", aiming to achieve the deep integration of artificial intelligence and business scenarios by optimizing business and data processes, improve efficiency, optimize decision-making, stimulate creativity, and provide strong impetus for the intelligent transformation of the securities industry.

In this context, Guotai Junan Securities conducted research on the adaptability of big models in the securities industry. Based on document question-and-answer tests, it evaluated the performance of big models in financial text comprehension tasks. It also combined Guotai Junan Securities' characteristic scenarios, namely big model industry chain graph mining and research report factor mining, to analyze how thinking chain technology can optimize the investment research process and improve the transparency and rigor of data analysis and reasoning.

Guojin Securities connected the DeepSeek big model to the company's characteristic big model industrial chain map intelligent mining scene (its technical architecture is shown in Figure 1), and originally proposed the Guojin Securities industrial chain evaluation system (as shown in Figure 2). Specifically, it uses the tree structure of the industrial chain to perform relevant edit distance calculations, and compares the industrial chain map automatically mined by the DeepSeek big model with the existing industrial chain benchmark library data to calculate the average edit distance between the two.

Figure 1 Guojin Securities' large-scale model industry chain graph intelligent mining technology architecture

Figure 2 Guojin Securities Industry Chain Evaluation System

After evaluation, the average edit distance between the industrial chain map automatically mined by DeepSeek and the benchmark library is similar, which shows that DeepSeek can accurately mine and construct the industrial chain map, and its results have high reliability and effectiveness. This further proves the application value of thinking chain technology in industrial chain analysis. In terms of industrial chain diversity and richness, Guojin Securities also evaluated the number of nodes in the industrial chain map. The number of nodes in the industrial chain map mined by DeepSeek is relatively large, which means that the richness of the industrial chain is relatively high. DeepSeek can disassemble the industrial structure more finely and deeply explore the industry context, thus forming a more comprehensive and detailed industrial chain analysis framework.

In terms of industrial chain mining, the current stage of the securities industry is generally carried out manually. After a new industrial chain is generated, relevant personnel need to spend a certain amount of time to organize and mine, and cannot be updated in real time. The use of large-model industrial chain intelligent mining technology for industrial chain mining can not only achieve rapid processing, but also its accuracy is comparable to manual mining. The application of this technology is expected to change the ecology of the securities industry, improve industry efficiency, and enhance industrial chain analysis and decision-making capabilities.

In the context of digital transformation and intelligent development, artificial intelligence technology is profoundly changing the way various industries operate. In order to more objectively evaluate the accuracy of DeepSeek in factor generation, Guojin Securities uses ChatGPT-4o to score each dimension of DeepSeek output, striving to more objectively and accurately evaluate the performance of the model. The evaluation results are shown in Table 1.

Table 1 Evaluation results of ChatGPT4o on DeepSeek

In terms of language fluency and expression, the content sentences generated by DeepSeek are concise and clear with good fluency; although some details are simplified, this expression style is very suitable for quickly understanding the report. There is no complex sentence structure and the tone is relatively uniform. The score is 4.5 points.

In terms of structural clarity, the DeepSeek report has a clear structure and distinct levels. Each part has a clear title, and the content progresses in a logical order. From the introduction to the conclusion, each part can be connected to each other, making it easy to track and quickly extract key information. The score is 5 points.

In terms of information richness, DeepSeek provides detailed content, especially in the "Incremental Information" section, which clearly lists the specific content and shows the performance of factors at different frequencies. Although the information is simplified, it does not sacrifice the necessary details, and the score is 4.5 points.

In terms of the correctness and rationality of factor logic, DeepSeek factor logic can reflect the core role of expected inertia factors in the generation process, which is consistent with the actual market response pattern. In terms of factor performance, the relationship between frequency adjustment and factor priority is reasonably displayed. The overall logic is very consistent with the practical needs of the investment field, with a score of 4.5.

In terms of code readability and usability, the content generated by DeepSeek is very concise, suitable for quick report generation and application in actual operations, and easy to understand and implement; in the code support part, the concise logic and clear expression make the code more usable in actual applications, and easier to maintain and modify, with a score of 4.5.

In comprehensive comparison, DeepSeek has significant advantages in report generation and application. It performs well in report generation speed, readability, simplicity and content depth, and is suitable for a fast and efficient working environment. Its screening and simplification of information makes the content more suitable for quick decision makers. Moreover, its code usability is good, the factor logic is reasonable, and its overall performance is excellent.

In the featured scenario, Guojin Securities conducted an in-depth evaluation of DeepSeek (as shown in Figure 3). The results showed that DeepSeek significantly improved business efficiency and decision-making quality.

Figure 3 Evaluation results of DeepSeek in Guojin Securities’ featured scenarios

Note: The maximum score for questions on positioning and questions on analysis and summary is 100 points.

In specific applications, DeepSeek shows a complete thinking process, making its relevance to user questions clearly visible. Through deep thinking and open thinking processes, the intuitiveness of reasoning is significantly improved, which can help users obtain key information more quickly and improve the accuracy and efficiency of document question and answer. Especially in the general document question and answer scenario, DeepSeek provides guidance on the original text positioning ideas of document question and answer, which is used to generate a logical reasoning report of document analysis summary, helping users to better understand and locate key information in the document. In the intelligent mining scenario of industrial chain graph, DeepSeek generates an industrial chain deep thinking report based on its thinking chain technology, and conducts in-depth analysis and mining of key factors and relationships in the industrial chain. In addition, DeepSeek also generates a factor deep thinking report and a code deep thinking report, which respectively analyzes the interpretation of large model research reports and factor mining process, large model factor mining and code generation process in detail, further improving users' understanding and application capabilities in these fields. The advantages of DeepSeek thinking chain technology are shown in Figure 4.

Figure 4 DeepSeek Thinking Chain Technical Advantages

After in-depth evaluation, Guotai Junan Securities focused on the innovative application of DeepSeek big models in the securities industry, combining its technological advantages with its own business needs, achieved a number of breakthroughs, and provided strong support for the intelligent transformation of the securities industry.

In the scenario of intelligent mining of large-scale industrial chain graphs, Guojin Securities introduced DeepSeek's thinking chain technology, which not only demonstrates the complete process of thinking reasoning, but also greatly improves the systematic nature of industrial chain research. With the reasoning ability of thinking chain, Guojin Securities can carry out more detailed hierarchical disassembly of complex industrial structures, deeply sort out the industry context, form a logically clear and insightful industrial chain analysis framework, and finally process and generate in-depth thinking reports.

The application of DeepSeek's thinking chain technology in Guojin Securities' characteristic industrial chain research includes the following scenarios:

(1) Intelligent dismantling of the industrial chain

Through DeepSeek's mind chain technology, Guotai Junan Securities can infer and model the relationships among enterprises, technologies, and supply chains in the upstream, midstream, and downstream of the industrial chain to form a dynamic, multi-level industrial chain map.

(2) Causal reasoning and trend prediction

With the help of the step-by-step reasoning ability of the mind chain technology, Guotai Junan Securities can not only restore the current industrial development path, but also identify key variables and make logical deductions on future trends to assist investment decisions.

(3) Open thinking process to improve user experience

The biggest feature of Thinking Chain is its explainability. During the industrial chain research process, Guotai Junan Securities opened up the reasoning process of Thinking Chain, from data input, key variable analysis, logical argumentation to the final in-depth report, so that users can intuitively see how the research team derives conclusions step by step, thereby enhancing trust and transparency.

(4) Open source thinking process to improve the level of research intelligence

Guotai Junan Securities plans to open source the thinking process of some industrial chain mining, which will not only provide users with a more intuitive experience, but also promote the collaborative optimization of investors, researchers and AI, and promote the iterative progress of intelligent research.

(5) Deep thinking to enable intelligent industrial research

Empowered by Thinking Chain, Guotai Junan Securities upgraded its industry chain research from traditional static analysis to the "intelligent reasoning + explainable analysis" model, helping research teams and investors to understand the industry landscape more efficiently and accurately, discover potential investment opportunities, and ultimately form more forward-looking in-depth research reports.

The factor mining thinking chain process of Guojin Securities' original characteristic scenario large-scale model research report (as shown in Figure 5) includes four steps: reading, tracing, defining and implementing. The dialogue tasks cover the overview of the research report, tracing the original text of the factor description, factor definition and calculation, and factor customized code. The research report factor mining task is gradually completed through multiple rounds of dialogue.

Figure 5 Guojin Securities’ original characteristic scenario large model research report factor mining thinking chain process

Guojin Securities generates factor code interpretability reports based on the deep thinking ability of the Thinking Chain technology. On the one hand, the deep thinking report of Thinking Chain can significantly improve the efficiency and accuracy of factor mining. Through the automatic generation of factor codes by the DeepSeek large model, the factor descriptions in the research report can be quickly converted into accurate code implementations, greatly shortening the time for manual code writing, while reducing the error rate caused by human negligence. On the other hand, the factor code interpretability report can promote innovation and development in the field of financial engineering. For example, the in-depth mining of the "expected inertia" factor in the report provides new construction ideas for multi-factor stock selection strategies, enriches the factor library, and helps investors find more valuable investment opportunities. In addition, the generated code follows the best practices of software engineering, has good readability and maintainability, and improves the quality of the code. At the same time, the factor code interpretability report also helps knowledge inheritance and team collaboration, making it easier for new members to quickly understand the factors and code logic, and promoting communication and collaboration among team members.

Under the AI-friendly governance framework, Guojin Securities has further expanded the application scenarios of large models, especially in the fields of intelligent mining of industrial chain maps and research report factor mining. The introduction of thinking chain technology enables complex information to be analyzed layer by layer, providing a systematic and in-depth analysis framework for securities business, thereby improving the accuracy and scientificity of investment research. The intelligent reasoning ability of large models not only expands the depth of financial analysis, but also improves the quality of investment decisions, enabling securities companies to more accurately understand industry trends and investment opportunities. At the same time, under the AI-friendly governance principles, Guojin Securities will pay more attention to data security, compliance and fairness in the application of large models, helping the securities industry to build a more robust and responsible intelligent system.

Looking ahead, Thinking Chain technology will have a far-reaching impact in multiple application scenarios in the securities industry. With the continuous innovation of technology, big models will continue to be deeply applied in investment research, intelligent question and answer, risk control management and other fields, providing stronger support for the intelligent transformation of securities companies. With transparent reasoning process and efficient information extraction capabilities, big models not only improve the systematic nature of securities research, but also enhance investors' trust in data sources and decision-making logic. With the continuous optimization of AI technology and the in-depth practice of AI-friendly concepts, big models will play a more critical role in the intelligent process of the financial industry, helping financial institutions achieve innovative development.