State Super, the $37 billion pension fund for the NSW state government, has deployed two machine learning (ML) tools to help inform asset allocation tilts and summarise economic data, as asset owners develop practical AI applications to power their existing investment processes.
As a case study in the CFA Institute Research and Policy Center’s latest report, Pensions in the Age of Artificial Intelligence, State Super said it started experimenting with AI tools when it realised its investment team receives a significant amount of “unstructured” information from asset managers (such as manager commentary reports). The challenge was to identify ways to use the information without human biases.
One of the two AI tools deployed was a reinforcement learning (RL) tool. The model is trained on structured market data every month and stored in a cloud service.
“Training takes approximately 24 to 48 hours, after which new data are fed into the model daily, and the same parameters are used to produce outputs before being retrained the next month,” the report said.
The outputs facilitate equity and currency tilts and inform conversations with managers about investment strategy.
The second tool is an AI agent, named “Vision” after the Marvel superhero, which scans publicly available data from organisations such as IMF and OECD and market commentary in news outlets. State Super chief investment officer Charles Wu said the goal was to be “resource efficient” while gaining economic insights.
“It compares and contrasts the information (akin to assessing intrinsic value against market price) and summarises key insights into a table so that the information can be used internally to support investment decisions,” the report said of the agent.
“Because the quality of the data used by the GPT agent is high (i.e., from official government databases), some of the risks associated with LLMs (large language models) can be mitigated.
State Super established an academic oversight body between 2019 and 2020 to support the development and governance of the AI tools.
Huge potential
The CFA Institute report also outlined the potential for AI enhancements up and down the pension value chain. It highlights to asset owners how AI can be used to modernise their operations as well as problems that may arise.
The report highlighted AI and ML as useful tools to increase the accuracy of manager selection and review for asset owners, using a case study from Japan’s Government Pension Investment Fund (GPIF).
GPIF, the world’s largest asset owner, uses a variety of external managers whose performance has been uneven over the past decade. It relied on a small number of internal human experts to select and evaluate active managers, but in 2017, GPIF trialled a multi-stage program to better identify and assess manager styles using techniques like deep learning.
The CFA Institute report said an AI system would allow GPIF to “more thoroughly, accurately, and efficiently evaluate fund manager investment style, providing quantitative metrics for what was previously available only as qualitative fund management descriptions”.
“These technologies demonstrate the potential for GPIF to access the benefits of a wider array of asset managers and funds by relying on internal data-driven analyses to judge fund manager performance, rather than relying on qualitative descriptions of performance or policies,” it said.
“This can serve to eliminate bias against fund managers with a history of working with GPIF and larger firms who are better placed to market their products.”
Information gathering and analysis is one of the most common ways asset owners are using AI, even from unconventional data sources when using programs like natural language processing (NLP) and “increase the range of information available”, the report said.
For example, NLP can screen earnings call transcripts and identify changes in company positioning, while sentiment analysis can be used to predict market and investor reactions to company events.
Similarly, AI can be used to scan and formulate ESG insights by collecting and analysing large amounts of structured and unstructured data from portfolio companies. The availability and consistency of company ESG data is by far the biggest challenge for pension funds that want to exercise stewardship, and a human-driven approach with AI assistance can greatly improve the efficiency of company analysis.
Connecting the dots
Ultimately, the most valuable advantage for risk and liability aware fiduciary investors is to know what is on the horizon, and AI has proven use cases in both market and credit risk management.
The report explained that since “market behaviour is nonlinear and emergent from dynamic system-wide interactions”, AI is well-placed to identify correlations and offer predictions of market crashes. ML technologies such as random forests and artificial neural networks can effectively identify signs of recession.
AI and ML could also help improve the accuracy of credit risk assessment, the report said.
But the use of AI is not without risks. Just like humans, CFA Institute highlighted that AI models have biases and variances. “Model bias refers to discrepancies between model predictions and actual values, and variance refers to the model’s generalisability and sensitivity to variations in the training data,” it said.
“Ideally, models would have low bias and low variance, but this scenario is not always possible because there is generally a tradeoff between bias and variance.”
Strong bias suggests “underfitting”, meaning the model is not complex enough for the training data, while strong variance suggests “overfitting”, meaning the model is too complex.
“Because market data are often irregular over time… and the statistical properties of such data also change over time, it can be challenging to generate models that appropriately fit these data variations and produce accurate predictions,” CFA Institute said in the report.
“As with other applications of AI and ML, it is important to subject AI-generated outputs to appropriate testing and supervision.”
