Alpha Go has such a paragraph: when Alpha Go evolves to Alpha master, it is already at the level of 13-dan Go. Facing professional 9-dan players can get 60 consecutive victories, no one can beat him. So Alpha Go came to the Chinese A-share market, and finally fled after more than half a year of challenges. This paragraph illustrates one point: Any smart financial project that wants to make investment transactions through artificial intelligence, and then lie down to make money is doomed to fail. So far, we have heard of a lot of Niu×’s smart financial products, none is Kensho, which was acquired by S&P Global for US$550 million. The main service provided by kensho to investors is to find the correlation between events and assets. And the impact of events on prices, especially long-term price trends. For example, what is the impact of the Brazilian earthquake on iron ore prices? What is the correlation between them? This is the main service provided by kensho, and its core is based on correlation analysis of historical big data, not asset pricing! Therefore, several years have passed. So far, we have not seen kensho evolve into a stock trading tool, or even popularized like search engines. Its current positioning is still intelligent investment research, providing special data services for researchers and fund managers.
The author once also did an intelligent financial project that is close to trading: position warning. The problem to be solved is to predict the impact of information related to your holdings on stock price fluctuations. The so-called relevant information refers to the information that has a 0 degree, 1 degree, and 2 degree relationship with the stock described by a knowledge graph. For example, if you hold a position in company A, then company A’s own information has a 0 degree relationship; for company A’s competitor, company B’s information is a 1 degree relationship for company A. The usual business research logic is that when competitor B has huge bad news, company A will show good news. According to this logic, we prepared sample data, used NLP for information analysis (event, sentiment analysis, popularity, influence, importance, etc.), and knowledge graph for relationship definition (competitor relationship, guarantee relationship, supply relationship, upstream and downstream relationship, etc.) ), machine learning to make model predictions. At first, we set the Y value as the influence of the information on the stock price increase after a period of time. We found that the influence is basically negligible. Then we revised the Y value to the effect on the stock price the next day. It is still not ideal, and then we continue to revise the impact of the next day's rise and fall, but it still does not work. Finally, the Y value is defined as the absolute value of stock price fluctuations. Finally found some relevance. From the perspective of this artificial intelligence project, the initial goal of the project is very close to the transaction, and we want to make stock price rise and fall forecasts. At the end of the project, when we move away from the goal of "asset pricing", we adjust the forecast results to relevant information for the next day. By analyzing the correlation of stock price fluctuations, we barely produced some results. In hindsight, is there any problem with the original research logic? From the perspective of a researcher familiar with the industry, there is no problem, but what is the competitive relationship for machine AI? Competitive relations can be divided into strong and weak ones. Weak competition relations have little effect, and strong competition relations may not necessarily have reverse influence. Take the famous longevity biological fake vaccine incident as an example. The bad incident of a leading company actually affects the whole The industry is bad, and all companies in the industry are affected to varying degrees. There is no good news.