Investing with AI

AI (Artificial Intelligence), a term from advanced computer science, has become a word that almost everyone has heard of. If you look at the newspaper, you will find news of products and services using AI on a daily basis. Asset management is no exception. AI investment trusts have also begun to appear where AI selects stocks instead of a human fund manager. So, in this blog, I would like to review AI from the viewpoint of asset management.

However, there is no generally agreed definition of AI, even among experts, and the interpretation varies depending on the person. That is because anything that runs on a computer can, in a sense, be called AI. This time, I would like to focus on the techniques of “text mining” and “machine learning” and think of them as AI and this kind of interpretation of AI is beginning to be considered as standard.

Is this time different?

When the older AI researchers get together, the conversation often turns to "Is this time different?". The current boom is said to be the third time that AI has been widely picked up by the mass media (see chart above) I remember the last boom, at that time I was doing AI research at university. In 1982, a national project called the "5th Generation Computer Project" was launched in Japan, and a huge national budget of 57 billion yen was invested. However, expectations for AI were very high and, when it failed to meet those expectations, the boom ended with disappointment, so my generation is the “second AI boom frustration generation”. The question of whether this time is different from the last boom has become a big topic among the older generation of AI researchers. The current situation of AI, whatever you think, reminds them of the last boom. Furthermore, the machine learning technology called “deep learning”, which is most often cited as the state-of-the-art AI technology, is in fact just a refinement of earlier technology called neural network (an algorithm that mimics the neural structure of the brain) that was actually developed in the first AI boom era. In other words, AI itself cannot be said to have made “revolutionary” technological progress since it is just a refinement of earlier work. So, will it end with a mere boom and excessive expectations this time around too?

AI + Big Data

I think this time is a little different from the past two booms. The reason is the existence of big data. It is the rapid spread of the Internet in the 21st century that has made the environment surrounding A different from the past, since it is now possible to digitize and store vast amounts of information (that is, big Data). When I was studying AI at university, the data used for analysis was hardly digitized. I was exhausted just by entering the data published in the newspaper or statistical surveys into my computer. In the end, we were not able to gather enough data aligned. Furthermore, the vast amounts of text information were barely digitized and could not be used.

Now, not only numerical data but also all information such as text and images are digitized, making it much easy to analyze. In particular, text and image information has much more information content and more immediacy than numerical data, so it can be a very powerful weapon if it can be used digitally. It is only with this big data that AI can be applied in various fields. In other words, the protagonist this time around is “AI + big data” rather than just AI.

The application of AI + big data to asset management is expected in various areas. For example, in investor profiling to build a portfolio that is optimal for each investor, or in investment model development in order to make the model more sophisticated. In the following sections I discuss two basic methods that are commonly used in applications of AI.

Converting text data into numerical data

The data used until now in most asset management models is mainly stock data, corporate financial data, GDP statistics, etc. However, these numerical data have less information content than text information, and it takes time to be released (apart from market data such as stock prices) and so lack immediacy. Even preliminary GDP data is delayed by two months. If it is possible to convert enormous amounts of text information available online on a daily basis into numerical data, it is possible to use more information immediately and improve the asset management model in various ways.

For example, in my laboratory, I am conducting research on applying text information from the press conference of the Bank of Japan President Haruhiko Kuroda to numerical data using an AI technology called text mining and applying it to an investment model. Text mining is a technology that mechanically reads sentences and breaks them down into meaningful words and phrases. The decomposed words are further converted into numerical data and used. By incorporating this new data into the asset management model, it is possible to refine the model and increase its immediacy. This is possible because the text of the press conference is digitized and available on the Internet (anyone can access for free from the Bank of Japan website.)

Automatic updating of models by machine learning

The investment model used for daily portfolio management is usually reviewed by researchers at regular intervals. That is because the economy and market structure change over time. Under normal circumstances, it is best to review the model when the structure changes. However, it is very difficult to detect structural changes, and it is common to carry out reviews at predetermined intervals, such as once a year or half a year. However, if you use machine learning, which is another AI technique, it is possible for the investment model to make a decision on its own regarding structural changes and thus improve by itself. A machine learning system is a program that can change the structure of the model based on newly acquired data. The program also decides for itself whether the structure has changed. This technology, like text mining, is by no means new. But with the explosion of information available through big data, machine learning capabilities have also greatly improved. My laboratory is also researching the development of a model that automatically learns and updates market sentiment data created by aggregating big data such as information published by companies and data on SNS.

In conclusion, is this time different?

So far, we have examined the possibility of applying AI + Big Data to asset management. There are many other research results in other fields that I have not had time to introduce this time. While the “Singularity” in which AI completely replaces human fund managers is unlikely to occur for the time being, given the current situation where all information has become digitized, success in asset management may depend largely on successful use of AI + Big Data. So, it seems that this time may be a little different after all.

Article By Prof. Yasuyuki Kato

Professor at Kyoto University Graduate School. Masters degree from Tokyo Institute of Technology, PhD from Kyoto University. Previously Executive Director at Nomura Securities Co., Ltd. Director of the Financial Engineering Research Center at Kyoto University. His specialty is investment theory and financial engineering. Academic adviser to Money Design Co., Limited.