Early attempts to apply machine learning to capital markets were met with bemusement. Perhaps trend followers could benefit from pattern recognition, but price discovery—a complex combination of market mechanics and psychology—was too difficult to model.
Today, technology and artificial intelligence (AI), in particular, threaten financial service jobs. Late to adopt, asset management will pay the price. Industry consultant Opimas predicts 90,000 fund manager, analyst and administrative support jobs will be lost globally over the next seven years, and 99% of all investment management will employ some form of machine learning over the next two decades, according to Man Group Plc. What is AI’s potential for replacing you? Here, we’ll survey and define AI capabilities.
AI, a branch of computer science, is the theory behind how machines are made intelligent and independent. Machine learning and symbolic learning are the two ways the theory is executed. Symbolic learning uses image processing and was an early input for robotics and computer vision. Machine learning is based on recognizing patterns and can be broken into statistical (for speech recognition and natural language processing) and deep learning (neural networks).
Below are some of the popular milestones demonstrating AI’s increasing relevance.
AlphaGo’s victory was impressive because of Go’s complexity, with 200 possible moves compared with chess’s 20. AlphaGo Zero’s chess win was striking as it needed only four hours after being loaded with the rules to learn the game and crush the most powerful open-source player, Stockfish.
Personal financial complexity is why Nobel laureate Bill Sharpe declared post-retirement investing to be the “nastiest, hardest problem in finance.” Perhaps with the help of AI, these challenges can be addressed.
AI falls into the following categories.
- Supervised or narrow AI: All data is labelled and the program learns to predict the output from the input data. Object recognition is an example. Scanning thousands of images with descriptions trains a machine to identify like images in new photos. Most of what we call AI falls into this category, including autonomous cars. Supervised AI can further be classified into two categories: regression and classification (see “AI in practice” for examples).
- Unsupervised learning (general AI): Data is unlabelled. The machine must learn by clustering similar properties. Unsupervised AI can be classified into clustering (like grouping clients by purchasing behaviour) and association (like associating that people who buy hot dogs also buy buns).
- Reinforcement learning: The machine is rewarded for performing the correct operation or operations in the optimal sequence. My colleagues and I use this approach for goals-based and retirement strategies, which we’ll talk about next time.
Technology and asset management
So-called smart beta strategies use a combination of regression and classification to identify approaches that might add value over an index. AI could help determine which strategies might outperform, and when, because timing remains a shortcoming in the investment world.
Today, AI has been used most frequently for following trends using pattern recognition (technical analysis) and latency arbitrage, a nice term for front-running via high-frequency trading. Can AI predict the future for capital markets? Stay tuned.
AI in practice
- AI and other advanced technologies are already being applied to test and enhance asset management strategies. Winton Capital, a U.K.-based hedge fund, used regression analysis to test Warren Buffett’s assertion that an acquiring company’s market value always fell following a takeover announcement. Data from almost 9,000 mergers proved Buffett’s hypothesis wrong.
- Professors Andrew Clare, Nick Motson and Steve Thomas (Cass Business School, City University of London, 2013) simulated portfolios constructed randomly by 10 million monkeys and classified the resulting portfolios by style (supervised learning). They found that monkeys consistently beat professionals.