Don't rush to open an account. First, have a chat with the "Investment Version of ChatGPT".

As an expat fund manager and the founder of an AI team, these two seemingly unrelated identities have coalesced in Zheng Qisen.

Leading a 20-person team, former Goldman Sachs analyst Zheng Qisen has launched an "Investment Version of ChatGPT" - ArborChat, an AI-driven investment expert robot assistant that can deeply analyze and answer various investment questions.

Quantitative finance is not a new thing. However, before the explosion of large-scale model technology, big data analysis was more about establishing statistical models to analyze the dynamic changes in the financial market and assist various investors in trading and risk management.

With the rapid advancement of large language model technology, artificial intelligence has a strong text interpretation and creation ability, and Finance + AI has been given greater expectations.

Beyond structured data such as daily trading data and financial data, can unstructured data such as corporate strategy, industry prospects, and corporate personnel changes also be summarized and concluded by AI to achieve value investment with less "machine flavor" and more "human flavor"?

ArborChat provides a direction. AI can imitate the human reasoning process. As long as you are good at giving the appropriate prompt, you can even "train" the large model into Buffett's investment style. "Buffett has always said that investing is simple but not easy. AI does not have the emotional baggage of humans and may be a better value investor!"

Finance + AI Requires a Bit of Idealism

"We hope that before making decisions, investors will first ask ArborChat what to consider and have sufficient information for reference instead of placing orders on impulse," Zheng Qisen introduced to 36Kr.

Zheng Qisen, who is 36 years old this year, graduated from the Hong Kong University of Science and Technology and once studied at Harvard University and Peking University. He is a "financial talent" from Goldman Sachs. He has both the "pragmatic spirit" of Hong Kong people and an "idealism" that is not very common in the financial circle. Even his entrepreneurial experience and product style are marked by this.

In 2014, when he was still working at Goldman Sachs, Zheng Qisen had a drink and chat with Liao Zhanpeng, a fund manager and old classmate who is engaged in value investment.

"Why is it so easy for those doing quantitative analysis next door?" Zheng Qisen was somewhat envious. AI can handle a large amount of data analysis in quantitative investment and help with many daily tasks, but value investment requires an understanding of perceptual knowledge such as business prospects, and the AI at that time was not capable of undertaking such work.

If the story only stops at envy, it definitely does not conform to the thinking path of an idealist.

What an idealist should see is that the intelligent investment advisors at that time were not "smart" but only "reliable". The portfolio strategy equation behind them has been used in the financial industry for decades and cannot perform complex and in-depth investment target analysis.

Zheng Qisen usually focuses on the TMT industry. His experience tells him that "investment is almost related to everything in the world. If AI can have general capabilities, there will definitely be a big breakthrough in intelligent investment advisors."

In fact, the seed in his heart was planted several years ago - in 2012, Nobel Prize winner Geoffrey Hinton and former OpenAI chief scientist Ilya Sutskever made a splash with AlexNet winning the ImageNet competition, which made Zheng Qisen see the possibility of AI gradually surpassing humans in different fields.

"If AI can surpass the human level in image processing, then it is imaginable that AI can also surpass the human level in language processing. To make value investment into an AI processing ability, it must exceed the human level to be solved."

In 2017, Google's Transformer model was released, and in 2018, the BERT model surpassed human scores in several natural language processing tasks for the first time, making Zheng Qisen see the possibility of the maturity of natural language processing technology. This year, he and Liao Zhanpeng also started their own entrepreneurial path.

Over the past six years, the financial AI products developed by Zheng Qisen and his team have served many banks and funds at home and abroad. In addition, they also use their own AI capabilities to optimize the account opening and fund subscription processes for high-net-worth clients and create a private equity investment platform with an asset management scale of 1.5 billion US dollars.

"For example, for one of the four major banks, we extract and analyze the financial statements of credit business customers. By extracting a large number of accounting figures and putting them into the credit model for analysis," he introduced, which compresses the usual seven-hour workflow of accountants to within half an hour.

Although the business is good, these do not really excite Zheng Qisen. Until the explosion of large models made the imagination of that year possible, and recently ArborChat was launched.

ArborChat uses a large language model system optimized for the financial field to provide retail and institutional investors with real-time investment insights, portfolio monitoring, and in-depth financial analysis.

Zheng Qisen finally迎来了 his "ChatGPT" moment. "The so-called AI investment in the past few decades has been a quantitative investment strategy based on statistical models. Based on the breakthrough large model technology, now the general artificial intelligence-like model can start to conduct very in-depth reasoning and obtain financial analysis of the fundamental investment method."

Zheng Qisen introduced that this is an unprecedented technological breakthrough that officially inaugurates the new investment style of value investment AI.

AI Value Investment Becomes Possible

How does an investment robot like ArborChat achieve value investment decisions?

"Some questions that we think are ordinary are actually very difficult for AI - for example, 'How to compare the stock price prospects of Alibaba and Pinduoduo?' To answer this question, AI needs to think in multiple steps, querying the financial data, analyst ratings, stock price technical analysis, business growth, corporate governance, and investor sentiment of Alibaba and Pinduoduo respectively, and finally summarize the comparison of the two companies." It can be said that having the ability of multi-step deep thinking after surpassing single-layer thinking is the basis for ArborChat to embark on an AI investment path different from traditional quantitative analysis. And ArborChat uses a unique "Thought Tree" technology to successfully overcome the threshold of deep reasoning in financial analysis.

"We noticed as early as last year that the DeepMind team was studying how to use the Monte Carlo Tree Algorithm to try to improve the reasoning ability of the large model, so we also started to study how to use the Thought Tree (that is, tree-shaped decision reasoning, rather than the linear 'Thought Chain') to create a model reasoning deep enough for good financial analysis." Through Zheng Qisen's explanation, ArborChat is not only a custom-made artificial intelligence tool for finance, but the team also has industry-leading research and development results in system research and development, becoming an industry-leading intelligent agent.

Over the past few years of entrepreneurship, Zheng Qisen has made himself half an expert in financial AI and has formed a professional team. "The AI director of ArborChat was previously a data science expert who worked in the banking industry for many years and is a cross-disciplinary generalist."

Understanding finance, understanding AI, some idealism, and a high degree of execution have created the core competitiveness of ArborChat. The ArborChat team spans two complex disciplines of finance and data, and years of industry experience have enabled it to have a deeper understanding of customer needs and industry pain points.

Zheng Qisen believes that "judging from the situation abroad, ChatGPT does not monopolize everything, and there will definitely be its own leading products in different vertical fields."

Taking the financial industry as an example, the general large model is not a specialized financial tool and will not deliberately connect with the financial database. "When their general framework processes financial data, it often refers to a lot of outdated information."

The ArborChat system also uses the RAG technology to extract relevant information in real time. However, the conventional RAG system is flawed, and when used to extract financial-related information, there will be a lot of noise.

In order to make the results more in line with the characteristics of finance, ArborChat has adopted an innovative data annotation method. The exclusive RAG technology not only significantly improves the extraction accuracy of financial and business answers but also maintains real-time performance. This is something that many AI for stock trading at home and abroad have not achieved.

Who Will Become the Next Generation of Investment Giants?

Compared with the situation of the cold investment in the secondary market in recent years, the recent account opening boom and the influx of post-00s investors cannot help but make people sigh that ArborChat has encountered a good opportunity. "We have experienced so many ups and downs in the past six years to wait until now. I hope this good opportunity can last a little longer, haha." Zheng Qisen said with a bitter smile.

Zheng Qisen also showed rational thinking about this situation. Young people may have a higher acceptance of the financial large model, but their C-end customer targets are still on more experienced and mature investors with a higher willingness to pay.

The unique value of ArborChat, including the solutions provided to retail and institutional investors, and B-end customers are also regarded as the commercialization plate of ArborChat.

"We have an API form to cooperate with different financial institutions and become part of their investment decisions," Zheng Qisen gave an example, saying that ArborChat can help financial institutions track changes in the investment sector. Due to the recent AI investment fever, many giants are investing in small modular nuclear reactors (SMR). If you want to rank the top 100 technology companies to see if they have invested in this area, it may take a fund company analyst a week to get the answer, while "the API provided by ArborChat on the to B side can answer within 1 minute."

By helping financial institutions build a dedicated investment analysis tool, from front-end investment sector analysis to portfolio tracking, including customer risk control, ArborChat can participate in it and improve efficiency.

The emergence of ArborChat is not an isolated case. With the further implementation of AI technology, it is an inevitable result that AI will participate in part of the investment research or risk control of all fundamental or value investment funds in the future, and even a situation where AI leads investment decisions may occur. "When the first version of ArborChat came out, I felt that it was not only the first step towards the realization of the vision that has been more than ten years, but also the beginning of a new industry paradigm. The emergence of quantitative investment has given birth to multiple investment giants with a scale of hundreds of billions. The AI of value investment will open up an extremely huge blue ocean and incubate the next generation of industry leaders."