Microsoft AI Research Methods for Quantitative Investing
Summary
This article surveys Microsoft Research’s published work on AI for quantitative investing, especially Chinese equity selection. It explains methods that use graph neural networks to represent relationships among stocks, attention mechanisms to adapt to changing market patterns, recurrent networks to process event sequences, and fund holdings or sentiment as model inputs. Other topics include learned risk factors with controls for factor overlap, reinforcement learning for order splitting, data augmentation, and graph-based multivariate forecasting.
The article reports that studies tested these methods on historical Chinese equity data and describes comparative results for several models, including improvements in selected prediction or portfolio metrics over stated baselines. These are backtest findings from particular samples and setups, not proof of future returns. The article also notes that research papers do not necessarily disclose the methods used in commercial partnerships. Its conclusions emphasize broadening AI research beyond stock selection and combining technical methods with practical domain questions. The evidence is a secondary discussion of published studies, and results depend on the datasets, periods, and evaluation designs used.
Key ideas
- Graph neural networks can encode explicit and learned relationships among stocks for prediction.
- Attention mechanisms can combine models or trading patterns as market behavior changes.
- Event sequences, sentiment, and fund holdings offer inputs beyond conventional price and factor data.
- Learned risk factors can be regularized to reduce overlap among factors.
- Reported advantages come from historical tests and may not generalize to future markets.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.