The proper selection and monitoring of fund managers is one of the most important tasks of Japan’s Government Pension Investment Fund (GPIF). Its approach depends on the track records and qualitative explanation of candidates and commissioned fund managers.

This can be improved with the use of artificial intelligence (AI).

In its 2016 fiscal year annual report, GPIF said that excess earnings from active funds had disappointed over the previous 10 years, while it had paid ¥70.9 billion ($623 million) to asset management companies to invest into domestic and foreign bonds and stocks over the previous three years.

This led to GPIF deciding to link the amount of management fees it will pay active managers with the level of active return they are able to supply (a concept chief investment officer Hiromichi Mizuno discussed with AsianInvestor in Tokyo in March 2018).

However, this does not solve all the problems within the manager structure practices.  

Other concerns were raised during a series of meetings with the Board of Governors of the GPIF, including a desire for the pension fund to select funds from the point of view of the whole universe, and the fact that evaluating funds on a qualitative basis could be somewhat arbitrary and lack objectivity. Additionally, GPIF’s evaluation and selection of active managers is implemented under very strong constraints by a small number of internal experts.

The pension fund recognised the need to fully modernise the process, possibly using data science and AI to rectify the problems within the current process. We sought to research how to create an AI system that enables GPIF to select and monitor fund managers based on data and stringent analysis of trading behaviour data.

CREATING VIRTUAL INVESTORS

A joint team formed between GPIF and Sony CSL went through GPIF’s manager structure development and maintenance processes in depth and agreed to develop a proof-of-concept prototype system that uses deep learning to detect the investment style of managers from trading behaviour data collected daily by GPIF. 

To evaluate investment styles in more quantitative detail, we applied deep-learning technology to devise an investment style analyser, called a ‘Style Detector Array’ system. To test the operating principle, we restricted the universe to 100 mainly large-cap stocks.

The Style Detector Array was then trained by simulating the trading behaviours of virtual managers that each adopt one of eight trading strategies: high dividend; minimum volatility; momentum; value; growth; quality; fixed weight; and technical. 

After the Style Detector Array was trained, we used data from 10 real domestic equity funds that offered sufficient available data to evaluate it. From this, distinct styles were evident – including temporal changes of styles even within a single fund. 

CONCLUSIONS 

Introducing our prototype AI system should enable GPIF to obtain detailed analysis of investment styles, based on data from fund managers submitted in advance. We believe that enables dialogue in presentations to be more precise and backed up by data. 

Furthermore, the ability to analyse the investment styles of fund managers already under contract makes it possible to analyse the management conditions of each fund manager, which should make it possible to increase efficiency. Clarity of communication in dialogue is promptly responsive to changes in style.

The style and drift of individual fund managers will also be detected, and that may encourage a certain discipline among the fund managers. GPIF will also be capable of analysing the behaviour of multiple fund managers from a comprehensive perspective – a bird’s-eye view of fund manager investment styles. 

When asset managers recognise that GPIF can independently analyse their investment styles and intends to develop even more advanced technology, they will recognise that they cannot justify their results with only qualitative explanations. As a result, they will look to improve the efficiency of their investment process by introducing more sophisticated technologies, including AI, to explain their behaviour and be accountable for their investment practices. 

This will eliminate the dependence on individual persons from management strategies, and promote optimisation. This sequence of developments will further promote the science and technology of asset management.

This Chalk Talk was taken from the December 2018/January 2019 edition of AsianInvestor. It is an edited summary of a report entitled ‘A Study on the Use of Artificial Intelligence within Government Pension Investment Fund’s Investment Management Practices’. It can be found by clicking here.