BERYL ELITES KNOWLEDGE CENTER
PANEL – Machine Learning and Artificial Intelligence Investing, Hype versus Reality: How much value do advanced analytics add to tactical and strategic asset allocation, portfolio construction, optimization, and risk management in multi-asset portfolios.
Panelists: ❖ Vidak Radonjic, The Beryl Consulting Group ❖ Jessica Stauth, Fidelity Investments
❖ Poul Kristensen, New York Life Investments ❖ Clark Cheng, Merrimac ❖ Peng Cheng, J.P. Morgan
* Don’t forget to subscribe to Beryl Elites YouTube channel for updates *
The application of Artificial Intelligence (AI) and Machine Learning (ML) in investing could be best described as the process of enabling machines to figure out pattern and create models from data. Both hypes and realities exist in such application. On the one hand, quantitative researchers and portfolio managers can utilize AI and ML to screen large amounts of data and discern signals from noise in the short-term, conduct portfolio optimization, execute trades, and screen through documents to reduce human intervention. That is because AI and ML perform best in developing short-term algorithms and figuring out patterns, thanks to the rapid growth of alternative data and computation power of machines in recent years. On the other hand, it is also noteworthy that some challenges could lead to hypes in the application of AI and ML in the investment management.
The first hype is signal generating and prediction in long-term. Almost all of the panelists mention that AI and ML do not work well in long-term decision making, or time series prediction. The second hype could happen due to the lack of adequate data in macro investing. The challenge behind is known as ‘curse of dimensionality’ – when applying ML methods, the amount of data needed would increase exponentially as the number of parameters or predictors grows. While the advantage of ML largely lies on its capability to handle large number of predictors, enormous amounts of input data is required. However, most of the data only have been created over the last few years. The third hype relates to availability of adequate historical data. To apply ML or AI in investment, quantitative researchers or portfolio managers need data that could go back multiple cycles to avoid skews. However, as mentioned above, most of the data available on the market do not have that long history. That could lead to overfitted or underfitted models. The fourth hype comes from data source. Most of the raw market data are in the hands of few big companies. While some of the companies sell data on the market, such data are only processed data, which might not fit for specific ML or AI model training purpose and could not be reprocessed. Without qualified input data, the models trained could be misleading. The fifth hype lies in the mechanism behinds ML and AI – the black-box natures of ML and AI. Most of the people who deploy ML or AI algorithms do not really know how the algorithms work behind the screen. Therefore, they might not know risks that lie behind the models. Without understanding the mechanisms and risks behind the models, people do not know when the algorithms might not work, and thus lose money. Such lack of understanding poses the challenge of persuading others to trust the model. The sixth hype lies in the scalability. Even though an AI or ML based model could succeed in managing a ten million dollar fund, that same model might not be able to manage a billion dollar fund. If a model could not be scaled up to a business level, it might not be very valuable.
Besides the challenges that lead to potential hypes mentioned above, panelists also point out that when AI or ML models are too complex for companies to embed into existing systems, or when quantitative researchers or portfolio managers could not figure out actual causes and effects from models, they might reach statistically plausible but actually wrong conclusions. In addition, panelists also discuss open-source application and sentiment based quantification. Panelists believe that open-source could make models more robust, if structured correctly. However, they differ in opinions on whether the more robust models are needed by financial service companies. In terms of the application of emotions and AI/ML for prediction, panelists indicate that AI and ML would be better fit for short-term predictions, while emotions would be better for long-term predictions. That being said, some panelists still hold pretty optimistic view of AI and ML application in investment world in the future. One panelist believes that the number of people applying AI or ML would grow rapidly in the next two to three years. Two of the panelist forecast that AI and ML would take more human elements in. However, it might take longer time for AI and ML to actually work in boosting long-term investment decision makings.
Copyright 2020 The Beryl Consulting Group. All Rights Reserved.
source