Machine Heart report
Author: Egg sauce
Since the launch of ChatGPT last year, this AI conversational bot has become the hottest technology topic in the world.
"Ask ChatGPT" has become a highly trusted decision-making aid for many people when facing choices, including seeking advice on financial investment.
Recently, two scholars at the University of Florida released a new study that integrates ChatGPT into an investment model to predict the trend of the stock market, and the return on investment can reach a staggering 500%. According to the introduction of the content of the paper, the model will first use ChatGPT to deeply analyze the news headlines and content released by listed companies to determine whether it is good news or bad news, and then rate these contents, and then make a "ChatGPT index" through complex calculation formulas, and compare them with the company's real-time stock price to verify ChatGPT's analysis capabilities.
Of course, in the face of unpredictable financial trends, it is not so easy to create an "AI Buffett". For now, new methods like the ChatGPT Index can only be used as a reference in the decision-making process.
However, in the field of financial investment, the use of AI technology to assist decision-making is indeed a relatively popular direction, and a special term has been derived: quantitative trading. Taking fund trend simulation and forecasting as an example, AI can conduct in-depth analysis of a large number of historical performance and other data, extract useful features from it, and build a model with high generalization ability and market dynamic adaptability, so as to maintain good forecast performance under volatile market conditions.
Quantitative models need to have a deep understanding of financial scenarios, investor sentiment analysis and market sentiment prediction, which is obviously what ChatGPT is good at. In the past six months, inspired and driven by ChatGPT products, the spark between quantitative trading and AI technology has further burst out.
At the same time, we are more concerned about how this ChatGPT storm will promote the landing of AI technology in the financial field from a global perspective?
Financial intelligence, where has it come?
In recent years, the application of AI technology in the financial field is deepening, injecting digital blood into the risk control, marketing, investment advisory, management and other businesses of the financial industry, and providing engine momentum for financial institutions such as banks, insurance, funds, and securities firms to achieve digital and intelligent transformation.
From industry research and judgment, investment advisory services, insurance claims, to credit risk control, these well-known financial business scenarios are supported by many technologies such as machine learning, computer vision, and natural language processing. Each application presents different challenges and involves different technical directions. Over the years, AI technology has also continued to evolve, small sample learning, trusted AI, explainable AI, AIGC and other new tools have emerged, especially at the end of last year, large model technology has completely changed the frontier picture of artificial intelligence technology, large models can speak the language ability, compress a large number of common sense knowledge ability in a large number of fields, the reasoning scheduling ability of digital professional tools undoubtedly has a broad application space in the financial industry, bringing us challenges at the same time, but also opening up the possibility of innovation.
So, what innovative applications will these new technologies and methods of artificial intelligence have in the financial field? Here, we can choose three directions: "financial data verification", "financial data understanding" and "financial scene understanding" to talk about recent changes.
First, let's talk about "financial data verification". The essence of digital financial business is based on the value exchange of data and information flow, and the authenticity and reliability of these data and information therefore constitute the key to the smooth progress of digital financial business. For example, in the digital payment scenario, the authenticity and verifiability of the user's payment credentials are directly related to the security and efficiency of payment. In the digital lending scenario, the authenticity and verifiability of the personal loan information provided by the borrower are the basis for judging the repayment ability and risk level.
Therefore, in all kinds of digital financial transactions, the verification means of various certificates and documents are indispensable to ensure their authenticity and reliability. In particular, the exploration of non-standard document tampering detection has always been an important research direction in industry and academia.
Text in financial documents contains a lot of important and sensitive information, and any small change in a sentence can distort the overall semantics it carries. With the development of NLP text processing technology, the black ash industry uses computers to tamper with false information in fraud, marketing, or other illegal activities, so it is critical to prevent the text in documents from being tampered with.
Previously, the technical research objects of image tamper detection focused on natural scene images, mostly relying on relatively obvious visual tampering cues at the edges or surfaces of objects, which are almost non-existent in documents. Here's why:
First, the text tampering methods in document images are quite diverse: some are stitching, that is, copying areas from one image and pasting them on other images; Some are copying-moving, that is, changing the spatial position of objects in the image; There is also generation, which replaces areas in an image with visually reasonable but different content.
Second, tampering with text is more hidden than scene images. The area of text that has been tampered with can be very small, such as a character in a paragraph; And the contrast between the tampered area and the surrounding environment can be very low, the images of documents mostly have the same background color, and the text often has the same font and size.
Relatively speaking, the development of text tamper detection methods is not mature enough. At present, some of the industry is aimed at algorithmic work for structured documents such as ID cards and business licenses, and it is often difficult for traditional detection systems to determine and modify the content of various types of unstructured documents such as qualification certificates, contracts, and reports common in the financial field.
In view of the problem of detecting document image tampering, various methods have been proposed by the academic community. Some researchers have introduced graph neural networks (GNNs) to detect tampered areas in document images with the help of graph attention mechanisms, but this method only has good results for relatively clear and clean documents, such as scanning documents. There are also researchers who use dual-stream Faster-RCNN networks to train images end-to-end to detect a given area of tampered images, however this type of tampering clue mostly exists in generative tampering and is difficult to find in very small copy-paste tampering.
Inspired by the above approach, document tamper detection has indeed made great progress. However, the existing methods still lack sufficient robustness and cross-domain generalization capabilities when encountering complex scenarios of various documents, and need to be further explored.
The second direction worth paying attention to is "financial data understanding". In real financial business scenarios such as wealth management, credit, and insurance, entities providing financial business services must not only understand the various modal data provided by users, such as credit self-certification material data, pet pictures of pet insurance, etc., but also need to combine structured and unstructured data in the field to produce professional and controllable financial management, insurance, and industry research knowledge to answer users' questions and provide users with full-process digital financial services. The technologies involved in this field are numerous, including computer vision, natural language processing and generation, AIGC, and so on.
From the perspective of understanding, the proportion of unstructured data in financial scenarios is high, and it is diverse, diverse in form, and heterogeneous, such as user self-certification data in credit scenarios, industry research reports of industry cognitive research, company financial reports, and insurance terms, etc., and the diversified document structure and differentiated contextual semantic environment bring challenges to the knowledge structuring task of unstructured documents. At the same time, the professionalism of financial scene documents also leads to problems such as high labeling costs and insufficient sample size for a single scenario.
From the generation level, with the continuous development and application of financial technology, the professionalism of financial services is increasing, and higher requirements are also put forward for the professionalism and compliance of content production. A particular challenge is that professional knowledge and financial logic are core requirements for content production in the financial sector. However, this precisely poses a great challenge to the current popular ChatGPT type large model, and the content produced by the large model needs to be truly applied in the financial field, and it is necessary to ensure that the output content is compliant (in line with regulatory requirements), professional (in line with financial logic), and rigorous (no factual errors such as knowledge illusions). The intelligent production of financial content requires the credibility and controllability of large models, which can be exported to the outside world with compliant, professional and rigorous standards.
In addition, the technology application proposition based on "financial scene understanding" has also attracted great attention. The innovation of AI technology has also brought acceleration to the landing in this direction. For example, "quantitative trading" mentioned above, no matter what investment strategy an investor adopts, the volatility of the financial market can rely on statistical methods and programmed expectations, and professional investors often try to estimate their overall return. Previously, there have been many computer-based algorithms and models for financial market transactions, and machine learning technologies such as time series information extraction, graph learning, and model integration have shown great application value in such tasks.
In principle, "market price fluctuations are influenced by a combination of rationality and human behavior" is quoted from Alphanomics, and investors inevitably have to make their own judgments and responses to news information. For example, if an investor finds that Apple's stock price fluctuates sharply after a spike in shipments, if he wants to explore the pattern, the investor can build a model to look for this pattern in Apple's historical data on the stock market, and make decisions based on the law.
In general, the more news you get from an effective representation of an event, the more quantitative models can assist investors in making more rational decisions. In recent years, some studies have begun to apply natural language processing (NLP) techniques to learn neural network representations of news events and build event-driven trading strategies based on this.
Since last year, large-model products represented by ChatGPT have also become the object of high expectations for investors. Large models can process a large amount of heterogeneous data, such as stock transaction data, macroeconomic data, company financial reports, etc., and can also process unstructured data, such as news reports, social media information, etc., to comprehensively improve the accuracy of forecast results.
The first comprehensive financial technology list in China was released
Facing these cutting-edge technical propositions in the field of financial intelligence, both academics and the industry have a strong interest in exploration, and many competitions and lists have revolved around such propositions in the past and received widespread attention.
In order to promote the potential exploration of the field of financial intelligence, under the guidance of CCF, on June 19, Ant Fortune, Ant Insurance and MYbank under Ant Group announced the first set of domestic financial technology list (algorithm competition competition) AFAC2023 Financial Intelligence Challenge, jointly initiated and held by Zhejiang University, Shanghai Jiao Tong University, Xi'an Jiaotong University, Central University of Finance and Economics, Ant Technology Research Institute and Tianchi Platform.
Contest official website: https://tianchi.aliyun.com/specials/promotion/AntFinTechAIChallenge
It is worth noting that this is the first comprehensive financial technology list in China that starts from solving application problems, filling the gap in the market.
Wang Xiaohang, Vice President of Ant Group, CTO of Ant Group Wealth Insurance Business Group and Chairman of Ant Fintech Committee, said, "As a fintech company, Ant hopes to encourage participants to explore various innovative models and algorithms from the most cutting-edge and challenging specific problems in the industry, so as to provide a platform for young people who are concerned about fintech to discuss the innovative prospects of artificial intelligence technology in the financial field."
Three core directions, six competition questions
Combined with its own operational experience, Ant has set six sub-competition questions in three core directions around the core technical propositions in the application of domestic financial scenarios:
The list covers a number of algorithm fields such as general machine learning, computer vision, and natural language processing, and is open to the world, and personnel from colleges and universities, scientific research institutions, and Internet companies can register for the competition.
The competition invited Professor Tang Jie of the Department of Computer Science of Tsinghua University as the chairman of the competition, and dozens of professors from Peking University, Shanghai Jiao Tong University, Zhejiang University, Fudan University, Chinese Minmin University, Xi'an Jiaotong University, Sun Yat-sen University, Tianjin University, Central University of Finance and Economics and East China Normal University to form an expert judging committee. Select outstanding technical talents or teams through the automatic evaluation of online rankings and the review of university professors and industry experts.
Schedule
At present, you only need to log in to the official website of the competition platform and complete the personal information registration to register for the competition.
Interested players can form teams of no more than 3 people, each player can only join one team, and each team can choose multiple questions.
Registration: June 19 - July 31, 2023, UTC+8
Contest period: June 19 - August 4, 2023, UTC+8
Finally, the top 6 teams will be required to submit a competition code and technical report:
Report submission: 11 August 2023, UTC+8
The competition will be judged and awarded in-person at the end of August and September. The winning team will receive CNY 50,000, the runner-up prize of CNY 20,000, and the third runner-up prize of CNY 10,000, in addition to the following benefits:
1. Green channel: Outstanding players have the opportunity to get the offer green channel and join the Ant Financial Technology team.
2. Certificate of Honor: The winners will receive a certificate of honor from the competition.
3. Offline award presentation: Winners will be invited to attend the award ceremony held in Shanghai to communicate face-to-face with academic and industry leaders.
By the way, at the registration stage, successful players who recommend friends will also have a chance to get competition gifts~
Want to know more about the competition? Welcome to register for the competition~
Registration link: https://tianchi.aliyun.com/specials/promotion/AntFinTechAIChallenge